Next Limit Technologies is a computer software company headquartered in Madrid, Spain. Founded in 1998 by engineers Victor Gonzalez and Ignacio Vargas the firm develops technologies in the field of digital simulation and visualization. In December 2016, the XFlow division was acquired by Dassault ...

Next Limit Technologies - Wikipedia

Xflow | 2,569 followers on LinkedIn. Cross-border payments simplified | Xflow is a technology company that builds cross-border payments infrastructure that enables businesses to send and receive money globally in a simple, reliable, transparent and compliant way.

Xflow | LinkedIn

X-FLOW is a mobile gaming company founded in 2018.

X-FLOW

X-Flow: pioneers in membrane technology. We deliver membrane know-how to OEMs and contractors.

Pentair - X-Flow | pioneers in membrane technology

Discover XFlow, powered by Dassault Systèmes® SIMULIA, a high fidelity Computational Fluid Dynamics (CFD) software based on Lattice-Boltzmann Method (LBM).

XFlow | High Fidelity CFD - Dassault Systèmes®

Make experience management part of your organisation’s muscle memory with xFlow, our automated action engine.

xFlow - XM Workflow Management & Action Engine | Qualtrics

Discover the new features and improvements in the latest release of XFlow

Latest Release | XFlow

The Single phase solver of XFlow CFD is used by major companies in the aerospace industry. The accuracy of this solver for predicting aircraft aerodynamic behavior has been very well validated on several reference benchmarks such as the 1st and 2nd AIAA High Lift Prediction Workshops, as well ...

XFLOW CFD Software - High Fidelity Fluid Dynamics | Simuleon

XFlow represents the next generation in CFD solution technologies. Moving beyond the Navier-Stokes equation and traditional meshing, XFlow is based on the Lattice Boltzman approach, and being particle based requires no invovled meshing processes. Being free from the constraints imposed on ...

XFlow CFD - High Fidelity Computational Fluid Dynamics Software ...

SIMULIA XFlow pushes the limits of Computational Fluid Dynamics (CFD). This advanced CFD software simulates dynamic or turbulent systems, plus advanced behaviors like surface tension.

SIMULIA XFlow - Advanced CFD Simulation

XFlow

xflow

How to make the right decisions and keep your stakeholders in the loop, including the doc template I use at Typeform Your team could build a million things. At the same time, your job as a PM is to make sure you ship the most valuable product. You want to solve the most important customer problems with the smallest effort possible. It’s pretty straightforward to find new product ideas. Your customers share their and request new features. The you collect on customer behavior highlights usability issues. Your team and company executives share their new ideas with you all the time. feedback data The hard part is to find the good ideas. Where should you invest your time and effort? In which direction should the team experiment next? Which idea will be 10x better than existing solutions? In this post, you’ll discover a simple tool which helps you discern the most promising ideas from the rest, by asking 4 questions: The opportunity analysis doc. A template for evaluating product opportunities When you evaluate a promising idea, you want to ask yourself 4 essential questions: What is the customer (or business) trying to achieve? Desired outcome. How does the current solution not allow the customer or business to reach their goals? Problem with existing solutions. How will we know we have solved the problem? Measures of success. How does fixing this problem contribute to our strategic goals? Strategic fit. The idea is to write your answers to these questions in a document, no longer than 2–3 pages. When you write it down, you can share your thought process with the whole organization. It’s a lightweight way to communicate all important information to your team and stakeholders. Here’s what it looks like: You can find a link to the template at the end of this post. Let’s dig deeper into what goes in each section. 1/ Desired outcome What is your customer or business trying to achieve? Focus on the underlying interests and desired outcomes. Leave out all solution ideas. Don’t equate feature requests and customers’ ideas with desired outcomes: “If I had asked people what they wanted, they would have said faster horses.” — Henry Ford (or , it doesn’t really matter) possibly someone else I personally like to use the jobs-to-be-done framework and focus on functional and emotional jobs. In case you haven’t heard about it, you can or . read up on it here watch a great talk about it here It’s important to be specific. Make the desired outcome tangible and concrete. Here’s a great example from on what a desired outcome is : “As a marketer, I want a mobile app so I can access my data away from a computer”. Jason Evanish not It’s rather: “On their way to work on the subway, content marketers like to check how their blog traffic is doing for items they published that morning or the day before. It helps them get into work and know how they’re doing before they sit down. If a number is low, they may try promoting it extra to try to raise the number. If the number is high, they may share the win with others on the team.” Instead of asking yourself “why would someone use my product?”, ask “what is this person trying to achieve?”. Source: Kathy Sierra’s © 2013 UserOnboard.com talk on building badass users Write in a style that’s easily understandable for anyone internally at your company. A new hire should be able to get what’s going on. Customer quotes are a powerful way to illustrate a point you’re making. With a small caveat: don’t use them in isolation. Make sure they reflect a pattern and not just an individual opinion. 2/ Problem with existing solutions Define how current solutions do not allow the customer or business to reach their goals. Provide focus and frame the problem. I use these to make sure the problem statements are effective: guidelines from Facebook Be human, simple, and straightforward. Capture the hearts and minds of the people you meet. Get at the why behind something happening. Don’t try to solve “people are dropping off in the X flow” — investigate the why behind them dropping off. Then solve that problem. Stay away from solutions. Remain company wins! — agnostic. Don’t think about monetizing just yet, focus only on the user (except for company-focused initiatives). Explain how you know this is a real problem worth solving. I like to summarize customer research and include key data points like support tickets and data analysis. If you find several problems, rank them in order of priority for the customer. 3/ How we will measure success It’s essential to have a clear success metric in place. Before you launch anything, you should know exactly how you will measure success. If you don’t, confirmation bias will lead to a wrong and misleading interpretation of the data you collect. Focus on the impact you want to have. “We will have accomplished this”, not “we need to do that.” Make sure to include how a metric is calculated and what it’s meant to measure. You can also come up with a counter metric which would convince you that you’re not breaking two new things every time you fix one. Here’s an : example from Intercom Put a new button in your product and people will click it. Get enough clicks and you can call that an increase in engagement. But that’s nonsense. A counter metric is “have people stopped doing anything else?”. So if you add a metric totrack one area of your product, you must also analyse the other areas that arelikely to be impacted. 4/ Strategic fit Name the company goals you impact with your initiative. This part acts as a sanity check and makes sure all efforts are aligned. For example, if your product is focused on small and medium businesses, don’t prioritize features for enterprise customers. Avoid a common mistake: documentation for documentation’s sake Don’t mistake the opportunity analysis doc with another lifeless item on your to-do list. You can’t just ‘punch out a product document’. Following a rigid approach would mean you’re missing the whole point. The document doesn’t replace the product discovery process. The doc is just the outcome of the process, in written form. It serves as a reminder for you to ask the right questions. it neatly: Jason Evanish summarizes If you don’t know the answer to one of the sections in the Thesis [their name for opportunity analysis doc], go find out. Dive into your analytics, talk to customers, run a survey, talk to your sales/account management/support teams that interact with customers regularly. You will gain the full respect of your designers and engineers if they know you always have a customer story and/or data to back up everything they may ask you about in the Thesis. That’s it! By following the above, you will make sure you’ve asked all the important questions before investing too much effort into an idea. Writing it down will help you structure your thinking process, and make your stakeholder communication easier. In case you want to use the template for yourself, to find it on Google Docs. Just make a copy and adapt it to your own needs. Thanks for reading. click here “There is no way to write a six-page, narratively structured memo and not have clear thinking.” — Jeff Bezos, Amazon

4 essential questions to ask before you prioritize product ideas

Image Attributed to @LucianoInf Now second only to South Korea as the biggest crypto market in Asia, Hong Kong's future might be one of decentralised ledgers, Bitcoin and Crypto-Currencies. Hong Kong will always hold a special place in my heart. For just shy of two years from 2016 to 2018 I lived and worked in the city, and still today it sits close to the top of the list of my preferred jurisdictions for business incorporation. Hong Kong is, as ever, a great place to do business and with its grit and its tenacity is very much the New York City of East Asia, or at least, it was. Every so often a need arises to clear up some business or banking in the city and so I find myself landing at its world class airport and dining at its world class restaurants with old friends and contacts. People who are close to the action and keyed in to the city's inner workings. Yet, as an infrequent visitor my eyes and ears observe changes not so obvious to those stuck in the daily grind, in the same way the ageing process is invisible from one day to the next, but stark and profound as the years roll by. Hong Kong is struggling to reinvent itself. Before you boomers get all giddy with your 'muh China', hold on. This isn't an article about the Chinese Government, I rather want to lay out my thoughts around Hong Kong itself, as well as the city relative to its immediate region. I don't intend to make the case for Singapore, which, despite being arguably the greatest beneficiary of Hong Kong's headaches, is still a top contender for title of the most boring place in Asia. I intend only to outline the unique challenges and opportunities faced by one of the most magnificent - and least boring - cities in the world. Gateway to the Middle Kingdom Incorporating your business in Hong Kong still makes sense in many cases. Hong Kong is pragmatic, with systems and infrastructure that is well established to facilitate commerce and trade. Structures can be highly tax efficient, and while getting a bank account has become more difficult, once established, the banking services at your disposal are second-to-none. Hong Kong is still a respected jurisdiction globally. Yet for entrepreneurs and companies alike the case for incorporating in Hong Kong has often been as part of an overall strategy to do business in and with the mainland. Because of the prevalence of a legal system and system of governance with its foundations in the British style, foreigners could come to Hong Kong and feel just enough familiarity not to be scurrying back to wherever they came from (and often this meant quite a lot of familiarity was on offer). Recently a contact of mine who does business in the mainland linked me to a new structure in Shanghai for establishing 100% foreign owned companies. While the corporate tax rate on the mainland is higher (25% vs 16.5% in Hong Kong), the fees associated with establishing such an entity are far lower, and skipping Hong Kong to go directly to the supplier or client has the benefit of reduced friction and complexity. There is no obvious remedy to this dilemma for Hong Kong. As companies have become more confident engaging directly with the mainland, the appeal of Hong Kong in this respect has waned. Added to this the fact that Hong Kong itself is too small to be a genuine start-up hub, there exists a need for the city to take a new approach. Economic Woes At the same time as Hong Kong's role as gateway to the mainland has diminished, there has been an ever growing inflow of both capital and people from the mainland. Every year the prevalence of Mandarin on the streets of the city grows, but, in the wake of the pandemic, this growth has been offset by an exodus of both Hong Kongers (primarily to the UK, Canada and Australia) and Western Expats. This population shuffle fundamentally changes the fabric of Hong Kong society, and is a phenomenon seen elsewhere, including for example in Germany. Those leaving Hong Kong are by-and-large wealthier than the average Hong Konger, and with them goes their skills and capital worsening the city's already pessimistic demographic outlook (with a Total Fertility Rate of 0.7 according to the World Bank, Hong Kong's ageing population - an almost worldwide issue - is one of its greatest challenges in the years ahead). Arrivals from the mainland have helped to offset this exodus, but only so much, and with the collapse in the Yen, Tokyo is increasingly attractive to wealthy Chinese looking to diversify themselves both financially and by way of residency. The fact that Hong Kong remains one of the most expensive places in the world only adds fuel to the fire. As the saying goes, one man's (or city's) loss is another's gain, and as already mentioned, Singapore has been effective in attracting capital and businesses included in the exodus. At the same time, countries such as Malaysia with its sizeable Chinese diaspora and corporate structures offering tax rates as low as 3% (via the Labuan incorporation route), are also seeing an uptick in interest from those who might otherwise have stayed in the Fragrant Harbour. While the economy on the mainland has been sluggish, the PBOC has adjusted rates downwards. Hong Kong's peg however has dictated that the Hong Kong Monetary Authority has followed the Fed's increases over the past few years. This has resulted in a policy rate that sits at 5.25% vs 3.10% for the PBOC at the time of writing (5 years ago the PBOC rate stood at 4.31% while the HKMA policy rate was around 1%). While the increase in rates in the USA was implemented to reign in inflation, Hong Kong suffered no such comparable phenomenon - although it did suffer significant capital outflows - and the increase in rates has been in direct contradiction to conventional economic wisdom that may have encouraged a policy of easing rather than tightening over the last few years had it not been for the peg. Crypto In February 2022, with the invasion of Ukraine by Russia marking a dramatic escalation in hostilities, Chinese policy makers and tycoons alike were undoubtedly keenly observing the geopolitical fallout. If not already clear, this event and the close to three years since has demonstrated the closeness of these two civilisational states. However, and perhaps more importantly, it has exposed the extent to which the United States will go to both pursue sanctions and weaponise the financial system that it controls via the greenback. Many, including myself, did not think that Vladimir Vladimirovich Putin would do it, and many others believed that the Russia - which had popularly, and falsely, been described by many in the West as a 'gas station masquerading as a country' (John McCain, 2015) - would quickly crumble under the weight of economic exclusion. As is usually the case, pretty much nothing happened according to consensus. Russia which had been so successful in its quick, practically bloodless annexation of Crimea, did not scare the Ukrainians back to the negotiating table. Conversely, the Russian economy has actually experienced something of an economic boom. However, and the point I want to drive home, from the perspective of those in Asia, it was a test case for what Hong Kong may face in the wake of any escalation in the straits of Taiwan. China's approach to crypto-currencies has been mixed at best. While officially there are bans on crypto, it is no coincidence that Hong Kong, which initially pushed back so hard on this financial innovation, has now seen a dramatic uptick in crypto assets. In a recent report by Chainalysis Hong Kong saw the fastest growth in its region at 86% from July 2023 to June 2024. In light of the situation in Ukraine, it is possible that crypto could be considered a useful tool in mitigating the effects of an economic war, such as that which the West has waged on Russia. At the very least, it does provide an incentive for regulators to look the other way or turn a blind eye. Conclusion With approximately 7.5 million inhabitants, Hong Kong just doesn't have the scale to make it a stand-out start-up hub, and its role as a conduit into the mainland will in all probability be further eroded by the likes of Shanghai and Shenzhen. Nevertheless, its role as a critical financial centre may be reinvigorated by a realignment towards crypto-assets and the newly emerging financial paradigm unfolding before our very eyes. Whether Trump's reelection will inspire a rapprochement and deescalation of the now vivid geopolitical schism calving up the world is yet to be seen. Nevertheless, one thing is abundantly clear, if Hong Kong is to retain its prestige and replicate its past success, things will need to change. Image Attributed to @LucianoInf Image Attributed to @LucianoInf Now second only to South Korea as the biggest crypto market in Asia, Hong Kong's future might be one of decentralised ledgers, Bitcoin and Crypto-Currencies. Now second only to South Korea as the biggest crypto market in Asia, Hong Kong's future might be one of decentralised ledgers, Bitcoin and Crypto-Currencies. Hong Kong will always hold a special place in my heart. For just shy of two years from 2016 to 2018 I lived and worked in the city, and still today it sits close to the top of the list of my preferred jurisdictions for business incorporation. Hong Kong is, as ever, a great place to do business and with its grit and its tenacity is very much the New York City of East Asia, or at least, it was. Every so often a need arises to clear up some business or banking in the city and so I find myself landing at its world class airport and dining at its world class restaurants with old friends and contacts. People who are close to the action and keyed in to the city's inner workings. Yet, as an infrequent visitor my eyes and ears observe changes not so obvious to those stuck in the daily grind, in the same way the ageing process is invisible from one day to the next, but stark and profound as the years roll by. Hong Kong is struggling to reinvent itself. Before you boomers get all giddy with your 'muh China', hold on. This isn't an article about the Chinese Government, I rather want to lay out my thoughts around Hong Kong itself, as well as the city relative to its immediate region. I don't intend to make the case for Singapore, which, despite being arguably the greatest beneficiary of Hong Kong's headaches, is still a top contender for title of the most boring place in Asia. I intend only to outline the unique challenges and opportunities faced by one of the most magnificent - and least boring - cities in the world. Gateway to the Middle Kingdom Incorporating your business in Hong Kong still makes sense in many cases. Hong Kong is pragmatic, with systems and infrastructure that is well established to facilitate commerce and trade. Structures can be highly tax efficient, and while getting a bank account has become more difficult, once established, the banking services at your disposal are second-to-none. Hong Kong is still a respected jurisdiction globally. Yet for entrepreneurs and companies alike the case for incorporating in Hong Kong has often been as part of an overall strategy to do business in and with the mainland. Because of the prevalence of a legal system and system of governance with its foundations in the British style, foreigners could come to Hong Kong and feel just enough familiarity not to be scurrying back to wherever they came from (and often this meant quite a lot of familiarity was on offer). Recently a contact of mine who does business in the mainland linked me to a new structure in Shanghai for establishing 100% foreign owned companies. While the corporate tax rate on the mainland is higher (25% vs 16.5% in Hong Kong), the fees associated with establishing such an entity are far lower, and skipping Hong Kong to go directly to the supplier or client has the benefit of reduced friction and complexity. There is no obvious remedy to this dilemma for Hong Kong. As companies have become more confident engaging directly with the mainland, the appeal of Hong Kong in this respect has waned. Added to this the fact that Hong Kong itself is too small to be a genuine start-up hub, there exists a need for the city to take a new approach. Economic Woes At the same time as Hong Kong's role as gateway to the mainland has diminished, there has been an ever growing inflow of both capital and people from the mainland. Every year the prevalence of Mandarin on the streets of the city grows, but, in the wake of the pandemic, this growth has been offset by an exodus of both Hong Kongers (primarily to the UK, Canada and Australia) and Western Expats. This population shuffle fundamentally changes the fabric of Hong Kong society, and is a phenomenon seen elsewhere, including for example in Germany. Those leaving Hong Kong are by-and-large wealthier than the average Hong Konger, and with them goes their skills and capital worsening the city's already pessimistic demographic outlook (with a Total Fertility Rate of 0.7 according to the World Bank, Hong Kong's ageing population - an almost worldwide issue - is one of its greatest challenges in the years ahead). population shuffle population shuffle Arrivals from the mainland have helped to offset this exodus, but only so much, and with the collapse in the Yen, Tokyo is increasingly attractive to wealthy Chinese looking to diversify themselves both financially and by way of residency. The fact that Hong Kong remains one of the most expensive places in the world only adds fuel to the fire. As the saying goes, one man's (or city's) loss is another's gain , and as already mentioned, Singapore has been effective in attracting capital and businesses included in the exodus. At the same time, countries such as Malaysia with its sizeable Chinese diaspora and corporate structures offering tax rates as low as 3% (via the Labuan incorporation route), are also seeing an uptick in interest from those who might otherwise have stayed in the Fragrant Harbour. one man's (or city's) loss is another's gain While the economy on the mainland has been sluggish, the PBOC has adjusted rates downwards. Hong Kong's peg however has dictated that the Hong Kong Monetary Authority has followed the Fed's increases over the past few years. This has resulted in a policy rate that sits at 5.25% vs 3.10% for the PBOC at the time of writing (5 years ago the PBOC rate stood at 4.31% while the HKMA policy rate was around 1%). While the increase in rates in the USA was implemented to reign in inflation, Hong Kong suffered no such comparable phenomenon - although it did suffer significant capital outflows - and the increase in rates has been in direct contradiction to conventional economic wisdom that may have encouraged a policy of easing rather than tightening over the last few years had it not been for the peg. Crypto In February 2022, with the invasion of Ukraine by Russia marking a dramatic escalation in hostilities, Chinese policy makers and tycoons alike were undoubtedly keenly observing the geopolitical fallout. If not already clear, this event and the close to three years since has demonstrated the closeness of these two civilisational states. However, and perhaps more importantly, it has exposed the extent to which the United States will go to both pursue sanctions and weaponise the financial system that it controls via the greenback. Many, including myself, did not think that Vladimir Vladimirovich Putin would do it, and many others believed that the Russia - which had popularly, and falsely, been described by many in the West as a 'gas station masquerading as a country' (John McCain, 2015) - would quickly crumble under the weight of economic exclusion. As is usually the case, pretty much nothing happened according to consensus. Russia which had been so successful in its quick, practically bloodless annexation of Crimea, did not scare the Ukrainians back to the negotiating table. Conversely, the Russian economy has actually experienced something of an economic boom. However, and the point I want to drive home, from the perspective of those in Asia, it was a test case for what Hong Kong may face in the wake of any escalation in the straits of Taiwan. China's approach to crypto-currencies has been mixed at best. While officially there are bans on crypto, it is no coincidence that Hong Kong, which initially pushed back so hard on this financial innovation, has now seen a dramatic uptick in crypto assets. In a recent report by Chainalysis Hong Kong saw the fastest growth in its region at 86% from July 2023 to June 2024. In light of the situation in Ukraine, it is possible that crypto could be considered a useful tool in mitigating the effects of an economic war, such as that which the West has waged on Russia. At the very least, it does provide an incentive for regulators to look the other way or turn a blind eye. report Conclusion With approximately 7.5 million inhabitants, Hong Kong just doesn't have the scale to make it a stand-out start-up hub, and its role as a conduit into the mainland will in all probability be further eroded by the likes of Shanghai and Shenzhen. Nevertheless, its role as a critical financial centre may be reinvigorated by a realignment towards crypto-assets and the newly emerging financial paradigm unfolding before our very eyes. Whether Trump's reelection will inspire a rapprochement and deescalation of the now vivid geopolitical schism calving up the world is yet to be seen. Nevertheless, one thing is abundantly clear, if Hong Kong is to retain its prestige and replicate its past success, things will need to change.

Hong Kong and the Crypto Boom: Economic Woes, Money Flows and Sanctions

How we Achieved 11% Growth by Leveraging Personalization and Segment-based Product Flows

HackerNoon: What is your company in 2–5 words? Uros Mikic: Full-service Webflow agency, one subscription. Why is now the time for your company to exist? We’ve built Flow Ninja by leveraging Webflow as a platform to power our product and marketing website and to offer Webflow to our customers. I believe Webflow’s CMS will become the go-to choice for all CMOs. It allows marketing teams to operate independently from engineering departments, providing greater control and flexibility over their websites. Currently, there are many excellent CMS solutions available. However, once your website traffic exceeds 100,000 visits, you need a robust tool to manage your marketing efforts. At this scale, Webflow stands unrivaled. Moreover, the ongoing challenge of expanding marketing teams involves hiring roles that are not required full-time. Our company addresses this issue by offering an entire team of design, development, and growth experts for the price of a single employee, all included in one subscription. What do you love about your team, and why are you the ones to solve this problem? I’m truly grateful that the team has been with me since our company was founded and that trust is still going strong. As I was 19 when we started, there were many bumps in the road, and were it not for the team surrounding me, it would have been a much more difficult journey. We were among the first to start working in Webflow in 2015, around two years after it was launched. If you weren’t building your startup, what would you be doing? I’ve fallen in love with the game of growing businesses, and I think this will stick with me forever. It gives me a sense of purpose. It makes me seize the day. If it wasn’t for Flow Ninja, I genuinely believe I would look for a potential partnership to help another business or start a new one from scratch. I must admit it: I’ve been daydreaming about running a traditional business. I think everything we do in the startup world would be ten times more effective if applied to a traditional mom-and-pop shop type of business. Introducing all of the systems related to HR, sales, marketing etc, would allow me to grow a traditional business to new heights. One thing I would love is to spend more time in my garden. Suppose time were to stand still, relieving me of all accountability in my business. In that case, I'd likely indulge in ample moments tending to my garden, allowing my mind to unwind and rejuvenate At the moment, how do you measure success? What are your metrics? We have different metrics for different departments. We try to hold people accountable for specific metrics. COO & CPO Customer churn
Team attrition
Company billability CTO & Design Late tickets
Billability
Client initiatives
Client reviews Finance & Sales Number of closed leads
Number of daily reachouts to cold clients
Cashflow (We are a fan of profit-first)
Net profit margin
Projected vs actualized revenue In a few sentences, what do you offer to whom? We offer a full-service Webflow team, including strategy, design, development, QA, SEO, account management, and more — all within a single subscription via our custom client relationship management platform. Our clients are usually product-led growth companies with more than $20M in revenue or funding and marketing teams of more than five people. We also work with enterprises and sales-led companies. What’s most exciting about your traction to date? As I write this, we are entering the most exciting period. As a profitable business, we are currently heavily investing in our team and strategic roles. Still, one of my favorite accolades to date is the 2023 Webflow Enterprise Partner of the Year, awarded by Webflow. Where do you think your growth will be next year? We hope to increase our growth from 30% to 60% by the end of this year and expand even more next year. We recently onboarded a new sales team and a Chief Revenue Officer, so I am confident we could grow much more than this. Tell us about your first paying customer and revenue expectations over the next year. Ah, the first paying customer. No one likes to hear about that. It was a website for an education institution that was built using a simple template. Another agency outsourced this project to us, presenting our work as theirs. They wanted the template to look like they'd built everything custom. On top of that, their budget barely covered our basic expenses, so we worked for peanuts, basically. There were countless revisions, and everything took five times longer than expected. No one quit in the process, and everyone was happy in the end. Those were rookie mistakes, but you live and learn, I guess. As for the expectations — we have aggressive plans to scale to $15M+ in the next five years. We’ve set a cornerstone for that: we focus on our custom-built in-house client relations platform and concentrate on our product. Finally, we ensure that our delivery puts us among the top 1% of all agencies offering similar services. With all that working well, we can now safely invest in growth. We expect to be in the $2-3M range next year. What’s your biggest threat? There are a lot of threats, and we keep them at bay. One might think that the biggest threat to the business is that we rely solely on Webflow to develop all of our websites. But I strongly disagree. I truly believe that Webflow is the only competitive platform in the complex website builders market. I’m confident I can push all of our developments into that platform, and I sleep well because I’ve met every single member of the Webflow leadership team, and they really know what they are doing. HackerNoon : What is your company in 2–5 words? HackerNoon Uros Mikic : Full-service Webflow agency, one subscription. Uros Mikic Uros Mikic Why is now the time for your company to exist? We’ve built Flow Ninja by leveraging Webflow as a platform to power our product and marketing website and to offer Webflow to our customers. I believe Webflow’s CMS will become the go-to choice for all CMOs. It allows marketing teams to operate independently from engineering departments, providing greater control and flexibility over their websites. Flow Ninja Currently, there are many excellent CMS solutions available. However, once your website traffic exceeds 100,000 visits, you need a robust tool to manage your marketing efforts. At this scale, Webflow stands unrivaled. Moreover, the ongoing challenge of expanding marketing teams involves hiring roles that are not required full-time. Our company addresses this issue by offering an entire team of design, development, and growth experts for the price of a single employee, all included in one subscription. What do you love about your team, and why are you the ones to solve this problem? I’m truly grateful that the team has been with me since our company was founded and that trust is still going strong. As I was 19 when we started, there were many bumps in the road, and were it not for the team surrounding me, it would have been a much more difficult journey. We were among the first to start working in Webflow in 2015, around two years after it was launched. If you weren’t building your startup, what would you be doing? I’ve fallen in love with the game of growing businesses, and I think this will stick with me forever. It gives me a sense of purpose. It makes me seize the day. If it wasn’t for Flow Ninja, I genuinely believe I would look for a potential partnership to help another business or start a new one from scratch. I must admit it: I’ve been daydreaming about running a traditional business. I think everything we do in the startup world would be ten times more effective if applied to a traditional mom-and-pop shop type of business. Introducing all of the systems related to HR, sales, marketing etc, would allow me to grow a traditional business to new heights. One thing I would love is to spend more time in my garden. Suppose time were to stand still, relieving me of all accountability in my business. In that case, I'd likely indulge in ample moments tending to my garden, allowing my mind to unwind and rejuvenate At the moment, how do you measure success? What are your metrics? We have different metrics for different departments. We try to hold people accountable for specific metrics. COO & CPO COO & CPO Customer churn Team attrition Company billability Customer churn Team attrition Company billability CTO & Design CTO & Design Late tickets Billability Client initiatives Client reviews Late tickets Billability Client initiatives Client reviews Finance & Sales Finance & Sales Number of closed leads Number of daily reachouts to cold clients Cashflow (We are a fan of profit-first) Net profit margin Projected vs actualized revenue Number of closed leads Number of daily reachouts to cold clients Cashflow (We are a fan of profit-first) Net profit margin Projected vs actualized revenue In a few sentences, what do you offer to whom? We offer a full-service Webflow team, including strategy, design, development, QA, SEO, account management, and more — all within a single subscription via our custom client relationship management platform. Our clients are usually product-led growth companies with more than $20M in revenue or funding and marketing teams of more than five people. We also work with enterprises and sales-led companies. What’s most exciting about your traction to date? As I write this, we are entering the most exciting period. As a profitable business, we are currently heavily investing in our team and strategic roles. Still, one of my favorite accolades to date is the 2023 Webflow Enterprise Partner of the Year, awarded by Webflow. Where do you think your growth will be next year? We hope to increase our growth from 30% to 60% by the end of this year and expand even more next year. We recently onboarded a new sales team and a Chief Revenue Officer, so I am confident we could grow much more than this. Tell us about your first paying customer and revenue expectations over the next year. Ah, the first paying customer. No one likes to hear about that. It was a website for an education institution that was built using a simple template. Another agency outsourced this project to us, presenting our work as theirs. They wanted the template to look like they'd built everything custom. On top of that, their budget barely covered our basic expenses, so we worked for peanuts, basically. There were countless revisions, and everything took five times longer than expected. No one quit in the process, and everyone was happy in the end. Those were rookie mistakes, but you live and learn, I guess. As for the expectations — we have aggressive plans to scale to $15M+ in the next five years. We’ve set a cornerstone for that: we focus on our custom-built in-house client relations platform and concentrate on our product. Finally, we ensure that our delivery puts us among the top 1% of all agencies offering similar services. With all that working well, we can now safely invest in growth. We expect to be in the $2-3M range next year. What’s your biggest threat? There are a lot of threats, and we keep them at bay. One might think that the biggest threat to the business is that we rely solely on Webflow to develop all of our websites. But I strongly disagree. I truly believe that Webflow is the only competitive platform in the complex website builders market. I’m confident I can push all of our developments into that platform, and I sleep well because I’ve met every single member of the Webflow leadership team, and they really know what they are doing.

"Growing Businesses Gives Me a Sense of Purpose." says Flow Ninja Founder/CEO 

"Οι αναπτυσσόμενες επιχειρήσεις μου δίνει μια αίσθηση σκοπού." λέει ο Flow Ninja Founder/CEO 

"Hacer crecer mi negocio me da un sentido de propósito", afirma el fundador y director ejecutivo de Flow Ninja

«Աճող բիզնեսները ինձ նպատակի զգացում են տալիս»: ասում է Flow Ninja-ի հիմնադիր/գործադիր տնօրենը 

"גידול עסקים נותן לי תחושת מטרה." אומר מייסד/מנכ"ל Flow Ninja 

「ビジネスの成長は私に目的意識を与えてくれます」とFlow Ninjaの創設者兼CEOは語る

"ການເຕີບໃຫຍ່ຂອງທຸລະກິດເຮັດໃຫ້ຂ້ອຍມີຄວາມຕັ້ງໃຈ." Flow Ninja ຜູ້ກໍ່ຕັ້ງ / CEO 

„Rozwój firm daje mi poczucie celu” – mówi założyciel/dyrektor generalny Flow Ninja

«Развитие бизнеса дает мне чувство цели», — говорит основатель и генеральный директор Flow Ninja

"Gukura ubucuruzi bimpa kumva intego." ati Flow Ninja washinze / Umuyobozi mukuru 

"வளரும் வணிகங்கள் எனக்கு ஒரு நோக்கத்தை அளிக்கிறது." Flow Ninja நிறுவனர்/CEO கூறுகிறார் 

"ምዕባለ ንግዳዊ ትካላት ናይ ዕላማ ስምዒት ይህበኒ።" ይብል መስራቲ/ዋና ኣካያዲ ስራሕ ፍሎው ኒንጃ 

"İşletmeleri Büyütmek Bana Bir Amaç Duygusu Veriyor." diyor Flow Ninja Kurucusu/CEO'su

(1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu);
(2) Bichen Wu, Meta GenAI and Corresponding author;
(3) Jialiang Wang, Meta GenAI;
(4) Licheng Yu, Meta GenAI;
(5) Kunpeng Li, Meta GenAI;
(6) Yinan Zhao, Meta GenAI;
(7) Ishan Misra, Meta GenAI;
(8) Jia-Bin Huang, Meta GenAI;
(9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com);
(10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com);
(11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction
2. Related Work
3. Preliminary
4. FlowVid
4.1. Inflating image U-Net to accommodate video
4.2. Training with joint spatial-temporal conditions
4.3. Generation: edit the first frame then propagate


Experiments


5.1. Settings
5.2. Qualitative results
5.3. Quantitative results
5.4. Ablation study and 5.5. Limitations
Conclusion, Acknowledgments and References
A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo We highly recommend looking at our demo web by opening the https://jeff-liangf.github.io/ projects/flowvid/ to check the video results. B. Quantitative comparisons B.1. CLIP scores Inspired by previous research, we utilize CLIP [36] to evaluate the generated videos’ quality. Specifically, we measure 1) Temporal Consistency (abbreviated as Tem-Con), which is the mean cosine similarity across all sequential frame pairs, and 2) Prompt Alignment (abbreviated as Pro-Ali), which calculates the mean cosine similarity between a given text prompt and all frames in a video. Our evaluation, detailed in Table 2, includes an analysis of 116 video-prompt pairs from the DAVIS dataset. Notably, CoDeF [32] and Rerender [49] exhibit lower scores in both temporal consistency and prompt alignment, aligning with the findings from our user study. Interestingly, TokenFlow shows superior performance in maintaining temporal consistency. However, it is important to note that TokenFlow occasionally underperforms in modifying the video, leading to outputs that closely resemble the original input. Our approach not only secures the highest ranking in prompt alignment but also performs commendably in temporal consistency, achieving second place. B.2. Runtime breakdown We benchmark the runtime with a 512 × 512 resolution video containing 120 frames (4 seconds video with FPS of 30). Our runtime evaluation was conducted on a 512 × 512 resolution video comprising 120 frames, equating to a 4-second clip at 30 frames per second (FPS). Both our methods, FlowVid, and Rerender [49], initially create key frames followed by the interpolation of non-key frames. For these techniques, we opted for a keyframe interval of 4. FlowVid demonstrates a marked efficiency in keyframe generation, completing 31 keyframes in just 1.1 minutes, compared to Rerender’s 5.2 minutes. This significant time disparity is attributed to our batch processing approach in FlowVid, which handles 16 images simultaneously, unlike Rerender’s sequential, single-image processing method. In the aspect of frame interpolation, Rerender employs a reference-based EbSynth method, which relies on input video’s non-key frames for interpolation guidance. This process is notably timeconsuming, requiring 5.6 minutes to interpolate 90 non-key frames. In contrast, our method utilizes a non-referencebased approach, RIFE [21], which significantly accelerates the process. Two other methods are CoDeF [32] and TokenFlow [13], both of which necessitate per-video preparation. Specifically, CoDeF involves training a model for reconstructing the canonical image, while TokenFlow requires a 500-step DDIM inversion process to acquire the latent representation. CoDeF and TokenFlow require approximately 3.5 minutes and 10.4 minutes, respectively, for this initial preparation phase. This paper is available on arxiv under CC 4.0 license. (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction 2. Related Work 3. Preliminary 4. FlowVid 4.1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate Experiments 5.1. Settings 5.2. Qualitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons Abstract and Introduction Abstract and Introduction 2. Related Work 2. Related Work 3. Preliminary 3. Preliminary 4. FlowVid 4. FlowVid 4 .1. Inflating image U-Net to accommodate video .1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate 4.3. Generation: edit the first frame then propagate Experiments Experiments Experiments 5.1. Settings 5.1. Settings 5.2. Qualitative results 5.2. Qualitative results 5.3. Quantitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo We highly recommend looking at our demo web by opening the https://jeff-liangf.github.io/ projects/flowvid/ to check the video results. B. Quantitative comparisons B.1. CLIP scores Inspired by previous research, we utilize CLIP [36] to evaluate the generated videos’ quality. Specifically, we measure 1) Temporal Consistency (abbreviated as Tem-Con), which is the mean cosine similarity across all sequential frame pairs, and 2) Prompt Alignment (abbreviated as Pro-Ali), which calculates the mean cosine similarity between a given text prompt and all frames in a video. Our evaluation, detailed in Table 2, includes an analysis of 116 video-prompt pairs from the DAVIS dataset. Notably, CoDeF [32] and Rerender [49] exhibit lower scores in both temporal consistency and prompt alignment, aligning with the findings from our user study. Interestingly, TokenFlow shows superior performance in maintaining temporal consistency. However, it is important to note that TokenFlow occasionally underperforms in modifying the video, leading to outputs that closely resemble the original input. Our approach not only secures the highest ranking in prompt alignment but also performs commendably in temporal consistency, achieving second place. B.2. Runtime breakdown We benchmark the runtime with a 512 × 512 resolution video containing 120 frames (4 seconds video with FPS of 30). Our runtime evaluation was conducted on a 512 × 512 resolution video comprising 120 frames, equating to a 4-second clip at 30 frames per second (FPS). Both our methods, FlowVid, and Rerender [49], initially create key frames followed by the interpolation of non-key frames. For these techniques, we opted for a keyframe interval of 4. FlowVid demonstrates a marked efficiency in keyframe generation, completing 31 keyframes in just 1.1 minutes, compared to Rerender’s 5.2 minutes. This significant time disparity is attributed to our batch processing approach in FlowVid, which handles 16 images simultaneously, unlike Rerender’s sequential, single-image processing method. In the aspect of frame interpolation, Rerender employs a reference-based EbSynth method, which relies on input video’s non-key frames for interpolation guidance. This process is notably timeconsuming, requiring 5.6 minutes to interpolate 90 non-key frames. In contrast, our method utilizes a non-referencebased approach, RIFE [21], which significantly accelerates the process. Two other methods are CoDeF [32] and TokenFlow [13], both of which necessitate per-video preparation. Specifically, CoDeF involves training a model for reconstructing the canonical image, while TokenFlow requires a 500-step DDIM inversion process to acquire the latent representation. CoDeF and TokenFlow require approximately 3.5 minutes and 10.4 minutes, respectively, for this initial preparation phase. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

FlowVid: Taming Imperfect Optical Flows: Webpage Demo and Quantitative Comparisons

FlowVid: Taming Imperfect Optical Flows: Conclusion, Acknowledgments and References

(1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu);
(2) Bichen Wu, Meta GenAI and Corresponding author;
(3) Jialiang Wang, Meta GenAI;
(4) Licheng Yu, Meta GenAI;
(5) Kunpeng Li, Meta GenAI;
(6) Yinan Zhao, Meta GenAI;
(7) Ishan Misra, Meta GenAI;
(8) Jia-Bin Huang, Meta GenAI;
(9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com);
(10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com);
(11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction
2. Related Work
3. Preliminary
4. FlowVid
4.1. Inflating image U-Net to accommodate video
4.2. Training with joint spatial-temporal conditions
4.3. Generation: edit the first frame then propagate


Experiments


5.1. Settings
5.2. Qualitative results
5.3. Quantitative results
5.4. Ablation study and 5.5. Limitations
Conclusion, Acknowledgments and References
A. Webpage Demo and B. Quantitative comparisons 5.4. Ablation study Condition combinations We study the four types of conditions in Figure 6(a): (I) Spatial controls: such as depth maps [38]. (II) Flow warped video: frames warped from the first frame using optical flow. (III) Flow occlusion: masks indicate which parts are occluded (marked as white). (IV) First frame. We evaluate combinations of these conditions in Figure 6(b), assessing their effectiveness by their winning rate against our full model which contains all four conditions. The spatial-only condition achieved a 9% winning rate, limited by its lack of temporal information. Including flow warped video significantly improved the winning rate to 38%, underscoring the importance of temporal guidance. We use gray pixels to indicate occluded areas, which might blend in with the original gray colors in the images. To avoid potential confusion, we further include a binary flow occlusion mask, which better helps the model to tell which part is occluded or not. The winning rate is further improved to 42%. Finally, we added the first frame condition to provide better texture guidance, particularly useful when the occlusion mask is large and few original pixels remain. Different control type: edge and depth We study two types of spatial conditions in our FlowVid: canny edge [5] and depth map [38]. Given an input frame as shown in Figure 7(a), the canny edge retains more details than the depth map, as seen from the eyes and mouth of the panda. The strength of spatial control would, in turn, affect the video editing. For style transfer prompt ’A Chinese ink painting of a panda eating bamboo’, as shown in Figure 7(c), the output of canny condition could keep the mouth of the panda in the right position while the depth condition would guess where the mouth is and result in an open mouth. The flexibility of the depth map, however, would be beneficial if we are doing object swap with prompt ’A koala eating bamboo’, as shown in Figure 7(d); the canny edge would put a pair of panda eyes on the face of the koala due to the strong control, while depth map would result in a better koala edit. During our evaluation, we found canny edge works better when we want to keep the structure of the input video as much as possible, such as stylization. The depth map works better if we have a larger scene change, such as an object swap, which requires more considerable editing flexibility. v**-prediction and** ϵ**-prediction** While ϵ-prediction is commonly used for parameterization in diffusion models, we found it may suffer from unnatural global color shifts across frames, as shown in Figure 8. Even though all these two methods use the same flow warped video, the ϵ-prediction introduces an unnatural grayer color. This phenomenon is also found in Imagen-Video [17]. 5.5. Limitations Although our FlowVid achieves significant performance, it does have some limitations. First, our FlowVid heavily relies on the first frame generation, which should be structurally aligned with the input frame. As shown in Figure 9(a), the edited first frame identifies the hind legs of the elephant as the front nose. The erroneous nose would propagate to the following frame and result in an unsatisfactory final prediction. The other challenge is when the camera or the object moves so fast that large occlusions occur. In this case, our model would guess, sometimes hallucinate, the missing blank regions. As shown in Figure 9(b), when the ballerina turns her body and head, the entire body part is masked out. Our model manages to handle the clothes but turns the back of the head into the front face, which would be confusing if displayed in a video. This paper is available on arxiv under CC 4.0 license. (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction 2. Related Work 3. Preliminary 4. FlowVid 4.1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate Experiments 5.1. Settings 5.2. Qualitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons Abstract and Introduction Abstract and Introduction 2. Related Work 2. Related Work 3. Preliminary 3. Preliminary 4. FlowVid 4. FlowVid 4 .1. Inflating image U-Net to accommodate video .1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate 4.3. Generation: edit the first frame then propagate Experiments Experiments Experiments 5.1. Settings 5.1. Settings 5.2. Qualitative results 5.2. Qualitative results 5.3. Quantitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo and B. Quantitative comparisons 5.4. Ablation study Condition combinations We study the four types of conditions in Figure 6(a): (I) Spatial controls: such as depth maps [38]. (II) Flow warped video: frames warped from the first frame using optical flow. (III) Flow occlusion: masks indicate which parts are occluded (marked as white). (IV) First frame. We evaluate combinations of these conditions in Figure 6(b), assessing their effectiveness by their winning rate against our full model which contains all four conditions. The spatial-only condition achieved a 9% winning rate, limited by its lack of temporal information. Including flow warped video significantly improved the winning rate to 38%, underscoring the importance of temporal guidance. We use gray pixels to indicate occluded areas, which might blend in with the original gray colors in the images. To avoid potential confusion, we further include a binary flow occlusion mask, which better helps the model to tell which part is occluded or not. The winning rate is further improved to 42%. Finally, we added the first frame condition to provide better texture guidance, particularly useful when the occlusion mask is large and few original pixels remain. Condition combinations Different control type: edge and depth We study two types of spatial conditions in our FlowVid: canny edge [5] and depth map [38]. Given an input frame as shown in Figure 7(a), the canny edge retains more details than the depth map, as seen from the eyes and mouth of the panda. The strength of spatial control would, in turn, affect the video editing. For style transfer prompt ’A Chinese ink painting of a panda eating bamboo’, as shown in Figure 7(c), the output of canny condition could keep the mouth of the panda in the right position while the depth condition would guess where the mouth is and result in an open mouth. The flexibility of the depth map, however, would be beneficial if we are doing object swap with prompt ’A koala eating bamboo’, as shown in Figure 7(d); the canny edge would put a pair of panda eyes on the face of the koala due to the strong control, while depth map would result in a better koala edit. During our evaluation, we found canny edge works better when we want to keep the structure of the input video as much as possible, such as stylization. The depth map works better if we have a larger scene change, such as an object swap, which requires more considerable editing flexibility. Different control type: edge and depth v**-prediction and** ϵ**-prediction** While ϵ-prediction is commonly used for parameterization in diffusion models, we found it may suffer from unnatural global color shifts across frames, as shown in Figure 8. Even though all these two methods use the same flow warped video, the ϵ-prediction introduces an unnatural grayer color. This phenomenon is also found in Imagen-Video [17]. 5.5. Limitations Although our FlowVid achieves significant performance, it does have some limitations. First, our FlowVid heavily relies on the first frame generation, which should be structurally aligned with the input frame. As shown in Figure 9(a), the edited first frame identifies the hind legs of the elephant as the front nose. The erroneous nose would propagate to the following frame and result in an unsatisfactory final prediction. The other challenge is when the camera or the object moves so fast that large occlusions occur. In this case, our model would guess, sometimes hallucinate, the missing blank regions. As shown in Figure 9(b), when the ballerina turns her body and head, the entire body part is masked out. Our model manages to handle the clothes but turns the back of the head into the front face, which would be confusing if displayed in a video. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

FlowVid: Taming Imperfect Optical Flows: Ablation Study and Limitations

(1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu);
(2) Bichen Wu, Meta GenAI and Corresponding author;
(3) Jialiang Wang, Meta GenAI;
(4) Licheng Yu, Meta GenAI;
(5) Kunpeng Li, Meta GenAI;
(6) Yinan Zhao, Meta GenAI;
(7) Ishan Misra, Meta GenAI;
(8) Jia-Bin Huang, Meta GenAI;
(9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com);
(10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com);
(11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction
2. Related Work
3. Preliminary
4. FlowVid
4.1. Inflating image U-Net to accommodate video
4.2. Training with joint spatial-temporal conditions
4.3. Generation: edit the first frame then propagate


Experiments


5.1. Settings
5.2. Qualitative results
5.3. Quantitative results
5.4. Ablation study and 5.5. Limitations
Conclusion, Acknowledgments and References
A. Webpage Demo and B. Quantitative comparisons 5.3. Quantitative results User study We conducted a human evaluation to compare our method with three notable works: CoDeF [32], Rerender [49], and TokenFlow [13]. The user study involves 25 DAVIS videos and 115 manually designed prompts. Participants are shown four videos and asked to identify which one has the best quality, considering both temporal consistency and text alignment. The results, including the average preference rate and standard deviation from five participants for all methods, are detailed in Table 1. Our method achieved a preference rate of 45.7%, outperforming CoDeF (3.5%), Rerender (10.2%), and TokenFlow (40.4%). During the evaluation, we observed that CoDeF struggles with significant motion in videos. The blurry constructed canonical images would always lead to unsatisfactory results. Rerender occasionally experiences color shifts and bright flickering. TokenFlow sometimes fails to sufficiently alter the video according to the prompt, resulting in an output similar to the original video. Pipeline runtime We also compare runtime efficiency with existing methods in Table 1. Video lengths can vary, resulting in different processing times. Here, we use a video containing 120 frames (4 seconds video with FPS of 30). The resolution is set to 512 × 512. Both our FlowVid model and Rerender [49] use a key frame interval of 4. We generate 31 keyframes by applying autoregressive evaluation twice, followed by RIFE [21] for interpolating the non-key frames. The total runtime, including image processing, model operation, and frame interpolation, is approximately 1.5 minutes. This is significantly faster than CoDeF (4.6 minutes), Rerender (10.8 minutes) and TokenFlow (15.8 minutes), being 3.1×, 7.2×, and 10.5 × faster, respectively. CoDeF requires per-video optimization to construct the canonical image. While Rerender adopts a sequential method, generating each frame one after the other, our model utilizes batch processing, allowing for more efficient handling of multiple frames simultaneously. In the case of TokenFlow, it requires a large number of DDIM inversion steps (typically around 500) for all frames to obtain the inverted latent, which is a resource-intensive process. We further report the runtime breakdown (Figure 10) in the Appendix. This paper is available on arxiv under CC 4.0 license. (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction 2. Related Work 3. Preliminary 4. FlowVid 4.1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate Experiments 5.1. Settings 5.2. Qualitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons Abstract and Introduction Abstract and Introduction 2. Related Work 2. Related Work 3. Preliminary 3. Preliminary 4. FlowVid 4. FlowVid 4 .1. Inflating image U-Net to accommodate video .1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate 4.3. Generation: edit the first frame then propagate Experiments Experiments Experiments 5.1. Settings 5.1. Settings 5.2. Qualitative results 5.2. Qualitative results 5.3. Quantitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo and B. Quantitative comparisons 5.3. Quantitative results User study We conducted a human evaluation to compare our method with three notable works: CoDeF [32], Rerender [49], and TokenFlow [13]. The user study involves 25 DAVIS videos and 115 manually designed prompts. Participants are shown four videos and asked to identify which one has the best quality, considering both temporal consistency and text alignment. The results, including the average preference rate and standard deviation from five participants for all methods, are detailed in Table 1. Our method achieved a preference rate of 45.7%, outperforming CoDeF (3.5%), Rerender (10.2%), and TokenFlow (40.4%). During the evaluation, we observed that CoDeF struggles with significant motion in videos. The blurry constructed canonical images would always lead to unsatisfactory results. Rerender occasionally experiences color shifts and bright flickering. TokenFlow sometimes fails to sufficiently alter the video according to the prompt, resulting in an output similar to the original video. User study Pipeline runtime We also compare runtime efficiency with existing methods in Table 1. Video lengths can vary, resulting in different processing times. Here, we use a video containing 120 frames (4 seconds video with FPS of 30). The resolution is set to 512 × 512. Both our FlowVid model and Rerender [49] use a key frame interval of 4. We generate 31 keyframes by applying autoregressive evaluation twice, followed by RIFE [21] for interpolating the non-key frames. Pipeline runtime The total runtime, including image processing, model operation, and frame interpolation, is approximately 1.5 minutes. This is significantly faster than CoDeF (4.6 minutes), Rerender (10.8 minutes) and TokenFlow (15.8 minutes), being 3.1×, 7.2×, and 10.5 × faster, respectively. CoDeF requires per-video optimization to construct the canonical image. While Rerender adopts a sequential method, generating each frame one after the other, our model utilizes batch processing, allowing for more efficient handling of multiple frames simultaneously. In the case of TokenFlow, it requires a large number of DDIM inversion steps (typically around 500) for all frames to obtain the inverted latent, which is a resource-intensive process. We further report the runtime breakdown (Figure 10) in the Appendix. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis: Quantitative result

(1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu);
(2) Bichen Wu, Meta GenAI and Corresponding author;
(3) Jialiang Wang, Meta GenAI;
(4) Licheng Yu, Meta GenAI;
(5) Kunpeng Li, Meta GenAI;
(6) Yinan Zhao, Meta GenAI;
(7) Ishan Misra, Meta GenAI;
(8) Jia-Bin Huang, Meta GenAI;
(9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com);
(10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com);
(11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction
2. Related Work
3. Preliminary
4. FlowVid
4.1. Inflating image U-Net to accommodate video
4.2. Training with joint spatial-temporal conditions
4.3. Generation: edit the first frame then propagate


Experiments


5.1. Settings
5.2. Qualitative results
5.3. Quantitative results
5.4. Ablation study and 5.5. Limitations
Conclusion, Acknowledgments and References
A. Webpage Demo and B. Quantitative comparisons 5.2. Qualitative results In Figure 5, we qualitatively compare our method with several representative approaches. Starting with a per-frame baseline directly applying I2I models, ControlNet, to each frame. Despite using a fixed random seed, this baseline often results in noticeable flickering, such as in the man’s clothing and the tiger’s fur. CoDeF [32] produces outputs with significant blurriness when motion is big in input video, evident in areas like the man’s hands and the tiger’s face. Rerender [49] often fails to capture large motions, such as the movement of paddles in the left example. Also, the color of the edited tiger’s legs tends to blend in with the background. TokenFlow [13] occasionally struggles to follow the prompt, such as transforming the man into a pirate in the left example. It also erroneously depicts the tiger with two legs for the first frame in the right example, leading to flickering in the output video. In contrast, our method stands out in terms of editing capabilities and overall video quality, demonstrating superior performance over these methods. We highly encourage readers to refer to more video comparisons in our supplementary videos. This paper is available on arxiv under CC 4.0 license. (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction 2. Related Work 3. Preliminary 4. FlowVid 4.1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate Experiments 5.1. Settings 5.2. Qualitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons Abstract and Introduction Abstract and Introduction 2. Related Work 2. Related Work 3. Preliminary 3. Preliminary 4. FlowVid 4. FlowVid 4 .1. Inflating image U-Net to accommodate video .1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate 4.3. Generation: edit the first frame then propagate Experiments Experiments Experiments 5.1. Settings 5.1. Settings 5.2. Qualitative results 5.2. Qualitative results 5.3. Quantitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo and B. Quantitative comparisons 5.2. Qualitative results In Figure 5, we qualitatively compare our method with several representative approaches. Starting with a per-frame baseline directly applying I2I models, ControlNet, to each frame. Despite using a fixed random seed, this baseline often results in noticeable flickering, such as in the man’s clothing and the tiger’s fur. CoDeF [32] produces outputs with significant blurriness when motion is big in input video, evident in areas like the man’s hands and the tiger’s face. Rerender [49] often fails to capture large motions, such as the movement of paddles in the left example. Also, the color of the edited tiger’s legs tends to blend in with the background. TokenFlow [13] occasionally struggles to follow the prompt, such as transforming the man into a pirate in the left example. It also erroneously depicts the tiger with two legs for the first frame in the right example, leading to flickering in the output video. In contrast, our method stands out in terms of editing capabilities and overall video quality, demonstrating superior performance over these methods. We highly encourage readers to refer to more video comparisons in our supplementary videos. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis: Qualitative Results

(1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu);
(2) Bichen Wu, Meta GenAI and Corresponding author;
(3) Jialiang Wang, Meta GenAI;
(4) Licheng Yu, Meta GenAI;
(5) Kunpeng Li, Meta GenAI;
(6) Yinan Zhao, Meta GenAI;
(7) Ishan Misra, Meta GenAI;
(8) Jia-Bin Huang, Meta GenAI;
(9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com);
(10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com);
(11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction
2. Related Work
3. Preliminary
4. FlowVid
4.1. Inflating image U-Net to accommodate video
4.2. Training with joint spatial-temporal conditions
4.3. Generation: edit the first frame then propagate


Experiments


5.1. Settings
5.2. Qualitative results
5.3. Quantitative results
5.4. Ablation study and 5.5. Limitations
Conclusion, Acknowledgments and References
A. Webpage Demo and B. Quantitative comparisons 5. Experiments 5.1. Settings Implementation Details We train our model with 100k videos from Shutterstock [1]. For each training video, we sequentially sample 16 frames with interval {2,4,8}, which represent videos lasting {1,2,4} seconds (taking videos with FPS of 30). The resolution of all images, including input frames, spatial condition images, and flow warped frames, is set to 512×512 via center crop. We train the model with a batch size of 1 per GPU and a total batch size of 8 with 8 GPUs. We employ AdamW optimizer [28] with a learning rate of 1e-5 for 100k iterations. As detailed in our method, we train the major U-Net and ControlNet U-Net joint branches with v-parameterization [41]. The training takes four days on one 8-A100-80G node. During generation, we first generate keyframes with our trained model and then use an off-the-shelf frame interpolation model, such as RIFE [21], to generate non-key frames. By default, we produce 16 key frames at an interval of 4, corresponding to a 2-second clip at 8 FPS. Then, we use RIFE to interpolate the results to 32 FPS. We employ classifier-free guidance [15] with a scale of 7.5 and use 20 inference sampling steps. Additionally, the Zero SNR noise scheduler [27] is utilized. We also fuse the self-attention features obtained during the DDIM inversion of corresponding key frames from the input video, following FateZero [35]. We evaluate our FlowVid with two different spatial conditions: canny edge maps [5] and depth maps [38]. A comparison of these controls can be found in Section 5.4. Evaluation We select the 25 object-centric videos from the public DAVIS dataset [34], covering humans, animals, etc We manually design 115 prompts for these videos, spanning from stylization to object swap. Besides, we also collect 50 Shutterstock videos [1] with 200 designed prompts. We conduct both qualitative (see Section 5.2) and quantitative comparisons (see Section 5.3) with state-of-the-art methods including Rerender [49], CoDeF [32] and TokenFlow [13]. We use their official codes with the default settings. This paper is available on arxiv under CC 4.0 license. (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). (1) Feng Liang, The University of Texas at Austin and Work partially done during an internship at Meta GenAI (Email: jeffliang@utexas.edu); (2) Bichen Wu, Meta GenAI and Corresponding author; (3) Jialiang Wang, Meta GenAI; (4) Licheng Yu, Meta GenAI; (5) Kunpeng Li, Meta GenAI; (6) Yinan Zhao, Meta GenAI; (7) Ishan Misra, Meta GenAI; (8) Jia-Bin Huang, Meta GenAI; (9) Peizhao Zhang, Meta GenAI (Email: stzpz@meta.com); (10) Peter Vajda, Meta GenAI (Email: vajdap@meta.com); (11) Diana Marculescu, The University of Texas at Austin (Email: dianam@utexas.edu). Table of Links Abstract and Introduction 2. Related Work 3. Preliminary 4. FlowVid 4.1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate Experiments 5.1. Settings 5.2. Qualitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons Abstract and Introduction Abstract and Introduction 2. Related Work 2. Related Work 3. Preliminary 3. Preliminary 4. FlowVid 4. FlowVid 4 .1. Inflating image U-Net to accommodate video .1. Inflating image U-Net to accommodate video 4.2. Training with joint spatial-temporal conditions 4.2. Training with joint spatial-temporal conditions 4.3. Generation: edit the first frame then propagate 4.3. Generation: edit the first frame then propagate Experiments Experiments Experiments 5.1. Settings 5.1. Settings 5.2. Qualitative results 5.2. Qualitative results 5.3. Quantitative results 5.3. Quantitative results 5.4. Ablation study and 5.5. Limitations 5.4. Ablation study and 5.5. Limitations Conclusion, Acknowledgments and References Conclusion, Acknowledgments and References A. Webpage Demo and B. Quantitative comparisons A. Webpage Demo and B. Quantitative comparisons 5. Experiments 5.1. Settings Implementation Details We train our model with 100k videos from Shutterstock [1]. For each training video, we sequentially sample 16 frames with interval {2,4,8}, which represent videos lasting {1,2,4} seconds (taking videos with FPS of 30). The resolution of all images, including input frames, spatial condition images, and flow warped frames, is set to 512×512 via center crop. We train the model with a batch size of 1 per GPU and a total batch size of 8 with 8 GPUs. We employ AdamW optimizer [28] with a learning rate of 1e-5 for 100k iterations. As detailed in our method, we train the major U-Net and ControlNet U-Net joint branches with v-parameterization [41]. The training takes four days on one 8-A100-80G node. Implementation Details During generation, we first generate keyframes with our trained model and then use an off-the-shelf frame interpolation model, such as RIFE [21], to generate non-key frames. By default, we produce 16 key frames at an interval of 4, corresponding to a 2-second clip at 8 FPS. Then, we use RIFE to interpolate the results to 32 FPS. We employ classifier-free guidance [15] with a scale of 7.5 and use 20 inference sampling steps. Additionally, the Zero SNR noise scheduler [27] is utilized. We also fuse the self-attention features obtained during the DDIM inversion of corresponding key frames from the input video, following FateZero [35]. We evaluate our FlowVid with two different spatial conditions: canny edge maps [5] and depth maps [38]. A comparison of these controls can be found in Section 5.4. Evaluation We select the 25 object-centric videos from the public DAVIS dataset [34], covering humans, animals, etc We manually design 115 prompts for these videos, spanning from stylization to object swap. Besides, we also collect 50 Shutterstock videos [1] with 200 designed prompts. We conduct both qualitative (see Section 5.2) and quantitative comparisons (see Section 5.3) with state-of-the-art methods including Rerender [49], CoDeF [32] and TokenFlow [13]. We use their official codes with the default settings. Evaluation This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

Stories, News & Mentions about XFLOW

Xflow’s Stories on HackerNoon

Xflow's News & Mentions