Regarding the most discussed AI image models recently, Nano Banana is undoubtedly the focus. With features like “character consistency,” it has gone viral on social media platforms, even triggering a new wave of excitement in the image generation sector.
However, the real value of Nano Banana goes beyond just being “user-friendly”. How did the team refine the model? What challenges did they encounter during the productization process? What is the future ceiling for image generation models—higher resolution or better understanding of user intent?
This issue of 「Yunqi Technology π」 features firsthand insights and deep reflections from Nicole Brichtova and Oliver Wang, core researchers of the Nano Banana team.
This article is reprinted from “Founder Park” (excerpted)
Original title: Nano Banana Core Team: The quality of image generation is nearly at its peak, the next step is to make the model understand user intention
Podcast source: Unsupervised Learning: Redpoint’s AI Podcast
01Higher resolution is the most requested feature by usersHost: Before the official release of “Nano Banana,” what uses did you think would be the most popular and exciting internally? Now that it has been launched, is the actual situation what you expected?Nicole: For me, the most exciting aspect is actually “character consistency,” being able to see myself in different scenarios. For example, I created a whole set of slides featuring me in a “wanted poster” and as an archaeologist, covering all the professions I dreamed of as a child. We even made a set of email templates that included my facial image, as well as those of other team members, for easy reference when developing new models.Host: In the AI field, that is definitely the highest honor.Nicole: And these images are very personalized, which excites me. So I was looking forward to the “character consistency” feature from the start because it allows people to “imagine themselves” in a whole new way, which was difficult to achieve before. This has indeed become one of the features users are most interested in.We found that people turn their images into “mini statues,” which is a very popular use case. However, one scenario surprised me: many people use it to colorize old photos, which holds significant emotional meaning for users. For example, “Now I can finally see what I looked like as a baby,” or “I can see what my parents looked like back in the day from these black-and-white photos.” This feedback is truly heartwarming.Host: After a product becomes popular, it will definitely receive countless feature requests. What do users most often ask for? What do you think will be the next milestone for image models?Nicole: The most frequent request we receive on Twitter is for “higher resolution,” as many professional users mention this. Currently, the model’s resolution is 1K, and users want it to be higher. Additionally, “transparent backgrounds” are also a high-frequency request because they are very practical for professional scenarios. These two are the most common requests I’ve seen, besides “better text rendering effects.”Host: Everyone is curious, what has led to such a significant improvement in the model’s performance?Oliver: I think the situation is actually quite “simple”; there isn’t a single factor that determines everything. The key is to refine all the details, continuously optimize the “technical solutions,” and the team has been researching this issue for a long time. To be honest, we were a bit surprised by how successful this model has been. We knew it was a good model and were looking forward to its launch, but we didn’t expect the response to be so strong. For example, after we released it on the “Arena” platform, not only did it receive high ratings, but what concerned me more was that so many users flocked to LM Arena to use this model that we had to repeatedly increase the “queries per second” (QPS) to support the load. This completely exceeded our expectations and made us realize for the first time: “Oh, this thing is really special, and many people need it.”Host: I think that’s one of the most interesting aspects of the entire AI ecosystem: as developers, you have a certain understanding of the models you build, but only when they are launched to the market and subjected to public scrutiny can you truly understand their “potential.”Host: “Nano Banana” has already gone viral online. Besides that, what other developments in the AI image field do you think are worth paying attention to but currently receive little attention?Nicole: I think it’s the “factual dimension of images.” For example, people use “Nano Banana” to create infographics or upload photos of Niagara Falls for the model to annotate information. As a demonstration, the results look decent, but upon closer inspection, you will find: the text has garbled characters, the information is inaccurate, and there are repetitions. This direction currently has few people paying attention, but I believe it will continue to improve in the future.Oliver: This is actually very similar to the development of LLMs. For example, when GPT-1 and GPT-2 first came out, people found them “interesting” and used them to write creative content, where the range of acceptable answers was quite broad. But now, people rarely use LLMs for creativity; they use them more for “information retrieval,” “conversations,” and even “seeking emotional companionship.” I believe image models may also undergo a similar transformation: from “creative tools” to “information retrieval tools,” and in the future, people may even converse with video models when they need companionship. This trend is likely to emerge in the future.Nicole: Moreover, the model should become “more proactive.” Currently, users must actively state, “I want to generate an image,” but what if the query itself “requires image assistance”? In fact, we are already accustomed to this kind of “proactive adaptation” in search engines, where the system automatically returns “text + image” or pure image results based on the needs during the search. I look forward to future models being more proactive and intelligent: flexibly using different modalities (text, images, etc.) to interact based on user questions.Host: Is there a story behind the name “Nano Banana”?Nicole: We have a PM on our team named Nana. At that time, she was still working overtime at 2:30 AM, and it was then that she came up with this name. It sounded fun, so everyone kept using it. Now it has even become a “semi-official name” since “Gemini 2.5 Flash Image Model” is indeed a bit cumbersome.Host: Yes, this name is very successful; even Google’s executives have posted banana emojis on Twitter, indicating that this name has “resonated with people.”Nicole: If there is any insight into brand promotion, it is that “the name should ideally be paired with a suitable emoji,” making it easier for people to remember.02 From “Toy” to Productivity ToolBecause it integrates the “world knowledge” of LLMsHost: You have solved the major challenge of “character consistency.” In your view, what will be the next “frontier breakthrough” for image models?Oliver: I believe the most exciting aspect of this model is that you can start to make “more complex requests” to it. In the past, you might need to describe the details of the desired image clearly, but now you can “seek help” just like talking to an LLM. For example, someone might use it like this: “I want to rearrange my room, but I don’t know how to do it; give me some suggestions,” and the model can provide reasonable proposals, such as “Based on your room’s color scheme, these pieces of furniture would fit well.”For me, what is truly interesting is how to integrate the “world knowledge” from LLMs into image models, allowing the generated images to genuinely assist users, such as showcasing solutions users hadn’t considered or answering their “information retrieval needs.” For example, if a user asks, “How does this thing work?” the model can directly generate a diagram, labeling that “this is how it works.” I believe this will be a very important application direction for this type of model in the future.Host: How much can image models benefit from the advancements in LLMs? And as LLMs continue to develop, will this trend of benefiting continue?Oliver: Of course, they can benefit, and almost 100% of this is due to the “world knowledge” of LLMs. In fact, the official name of this model is “Gemini 2.5 Flash Image Model,” and “Nano Banana” is just a more interesting nickname.Oliver: I even wonder how much of our success is due to the catchy name “Nano Banana.” But it is indeed a model from the Gemini series, so you can interact with it just like you would with Gemini; it understands everything Gemini understands. I believe integrating image models with language models is a crucial step in enhancing the model’s practicality and functionality.Nicole: You may remember that two or three years ago, if you wanted the model to generate an image, you had to describe it very specifically, such as “a cat sitting on a table, with the background looking like this, and these colors.” But now, it’s not that complicated, largely because the performance of language models has improved significantly.Host: Yes, you no longer have to “sneakily perform prompt conversion” like before. The previous “little trick” was: you input a sentence, and the system would convert it into a detailed prompt of 10 sentences to ensure the model could generate content accurately. But now, the model’s complexity is high enough to directly understand simple prompts, which is truly exciting.03The future interaction will definitely be multimodalRecognizing user intent is particularly criticalHost: From a product perspective, the user base of “Nano Banana” is actually very diverse. There are experts who know exactly what they want to do, as well as many ordinary users facing the “blank canvas dilemma.” Can you talk about how you design products for these two completely different types of users?Nicole: First of all, users on LM Arena, including developers, are very professional, familiar with how to use these tools, and can come up with new scenarios we hadn’t anticipated. For example, someone might turn objects in photos into “holograms,” which we hadn’t trained for and didn’t expect the model to excel at, but it performed well.For ordinary consumers, “simplifying operations” is crucial. For example, when you open the Gemini app now, you will see banana emojis everywhere; we did this because we found that many people, upon hearing “banana” (referring to the model), couldn’t find it in the app because there was no obvious entry point before. We also collaborated with creators to showcase some use cases in advance, providing examples that link directly to the Gemini app, where clicking would auto-fill the prompt. I believe we still have a lot to do in terms of “initial interface guidance,” such as providing visual guidance.Additionally, when editing images, we might consider adding “gesture controls” so that users don’t have to rely entirely on prompts. Sometimes, even if you want a specific effect, you need to write a long prompt, which doesn’t feel natural for most consumers. I use the “parent test” to validate products; if my parents can use it easily, then it qualifies. But we haven’t reached that standard yet, so we still have a long way to go. However, the core idea is actually “show more, teach less”: provide users with examples they can easily replicate, making sharing simple. As Oliver often says, there is no “magical single solution”; it requires multifaceted efforts.Oliver: “Social sharing” is actually the key to solving the “blank canvas dilemma.” When people see content created by others using the model, they easily think, “I can also try putting myself, my friends, or my pets in it.” This kind of “imitative creation” is an important way for “Nano Banana” to spread.Host: Currently, interaction is mainly through text. In the long term, what other “design interfaces” can make it easier for people to interact with the model? Are there any exciting ideas in this regard?Nicole: I think we are just beginning to scratch the surface of “interaction possibilities.” Ultimately, I hope all “modalities” (text, images, voice, etc.) can merge into an “intelligent interface” that automatically selects the most suitable interaction method based on the task you want to accomplish.For example, we are already moving towards “LLMs not only outputting text but also generating images or visual explanations when users need them.” Voice interaction also has great potential because it is a very natural way for people, but no one has truly solved the problem of “how to integrate voice interaction into user interfaces” yet. We still mainly rely on “inputting text,” and perhaps we can combine it with “gestures,” such as wanting to remove an object from an image, which should be as simple as erasing it on a draft. How to seamlessly switch between different interaction modalities based on task requirements is a direction I am very interested in, and there is still much exploration space in this area.Additionally, I think the idea of “generating a ‘directly usable product’ with a short prompt” is overhyped. In reality, generating content requires a lot of iterative optimization; even the content shared on social platforms requires a lot of effort to refine into the final effect. So this expectation of “one-step completion” is somewhat unrealistic; the future interaction interface (UI) is currently underestimated. How to integrate various modalities (text, images, voice, etc.) to make it easier for ordinary people to use these models and understand their capabilities, while allowing the models to adapt to specific workflows, is a direction whose value has not been fully recognized.Host: What are the limitations currently facing the “voice interaction interface”?Nicole: I think part of the reason may be “prioritization”; we are currently focused on enhancing the core capabilities of the model. However, voice technology has indeed made significant progress in recent years, so I believe someone will soon start exploring “the combination of voice and image models,” and our team may also work on this aspect.I have been pondering what this interaction interface might look like. I think the core of the problem lies in how to recognize user intent and how to switch to different interaction modes based on user intent and the tasks they actually want to accomplish, as user needs are often unclear. Moreover, this could lead the interface back to a “blank canvas” state, so how to show users “which operations are feasible” is a significant challenge in itself.We find that users tend to assume that chatbots “can do anything” when using them, as you can communicate with them like talking to a person. However, in reality, explaining to users “what the bot cannot do” is very difficult; especially when the tool’s capabilities are already very powerful, clearly demonstrating “what it can do” is also not easy. Therefore, I believe the key lies in clarifying the boundaries of the problem and designing the interface to make it clear to users “which operations are feasible,” ultimately helping them accomplish almost everything they want to do.04 Active testing from real users is the best way to evaluate modelsHost: Let’s talk about “model evaluation.” Besides publicly testing on the LM Arena platform, how do you conduct regular evaluations? What insights do you have on “how to judge and measure the quality of models”?Oliver: In fact, the advancements in language models and visual language models have brought about a benefit: a “feedback loop” has formed, allowing us to use the intelligence of language models to evaluate the content they generate. This creates a virtuous cycle that can simultaneously promote progress in both language models and image models, which is very exciting. Ultimately, users themselves are the “final standard for judging whether an image meets their needs.” Therefore, allowing users to input their prompts to use the model, like on the LM Arena platform, is actually the best way to evaluate the model.Nicole: “Aesthetics” is also very important. Oliver is quite modest; in fact, she is the person in the team with a “high sensitivity to image details” and can immediately spot whether the image output is good or has defects. We have several such members in our team who, after the model training is completed, will first conduct a large amount of “manual initial screening” to determine whether the model’s output is qualified.Returning to your question about “evaluation methods,” we receive user feedback from many channels (including X) to understand “which features are useful and which are not.” We then adjust the evaluation criteria to ensure that “useful features do not degrade” while focusing on optimizing “useless features” that the community wants to improve.05 Meeting the demand for “aesthetics” is challenging, requiring deep contextual interaction at the prompt levelHost: Among the “experienced users” you have encountered, are there any particularly impressive use cases?Oliver: My personal favorite use case among experienced users? I have spent most of my career working with video-related tasks, so I am particularly interested in video tools and creative tools. I found that when “Nana Banana” is used in conjunction with video models like Voe3, it can become a practical tool for creating AI-generated videos, helping you brainstorm ideas and plan shots more quickly. Interestingly, this resembles the production process in the film industry: first using a “storyboard” to outline the story and shots, and now users are also using this method to create more coherent and longer video content.Nicole: I was surprised to see someone use it in “actual architectural workflows.” For example, starting from blueprints, first generating effects similar to 3D models (without actually building a 3D model), and then further iterating into design drawings. This greatly shortens the “tedious repetitive steps” in the workflow, allowing people to focus on the “creative, interesting, and personally enjoyable aspects.” Moreover, I didn’t expect its “out-of-the-box” performance in such scenarios to be so good.Host: Just like quickly building a “basic framework” using image models across various fields.Nicole: Another scenario is “generating website UI through code.” Previously, the process from “inputting a prompt” to “generating website code” always felt abrupt to me, as it lacked an “iterative design” step, making it impossible to quickly modify design plans. But now, we can finally iterate on the design before generating the code, ensuring that we are satisfied before generating the code.Host: This is simply the future workflow. After all, if the generated code does not meet your aesthetic standards or completely deviates from your expectations, then the computational power spent on “generating code” would be wasted, right? This approach makes much more sense.Nicole: And it is also more enjoyable. As Oliver said, people will naturally integrate new technologies into existing workflows. Although the rapid progress of LLMs can generate websites directly from prompts, which is astonishing, I believe spending more time on the “design iteration” step to ensure the final effect meets one’s aesthetic standards will be more enjoyable for users.Host: How far have we progressed in this direction?Oliver: The demand related to “aesthetics” is actually quite difficult to meet because it requires deep personalization to provide useful suggestions. Moreover, I believe that on a technical level, “personalization” itself is still being continuously optimized. Therefore, we still have some distance to go before we can “precisely understand user needs.” However, I think that through “a small amount of clarification” and “dialoguing with the model,” this is one of the features I look forward to the most, and the situation will improve. You will be able to communicate with the model like in a chat thread, gradually refining your needs to ultimately get the image you want.Host: Do you think “personalization” will remain at the “prompt level”? For example, achieving it through dialogue and context? Or will everyone have their own dedicated “aesthetic model” in the future?Oliver: I think it will mostly stay at the “prompt level.” For example, based on the personal preferences you have previously communicated, the model can make decisions that better meet your needs. At least I hope so. After all, if everyone has to maintain their own model, that sounds quite troublesome. So this may be a direction for future development.Nicole: But I do believe that different people will have distinctly different “aesthetic preferences,” and a certain degree of “personalization” is essential at this level. For example, when you search for sweaters on Google’s “shopping tab,” you receive many recommendations, but you actually hope to “align with your aesthetic” and even “combine with the clothes you already have in your wardrobe” to see which new clothes can match. I hope this demand can be realized through the “context window of the model,” such as feeding the model images of clothes in the wardrobe to recommend matching styles. I am very excited about this direction and hope to achieve it. Of course, perhaps some additional “aesthetic control” will be needed at the “model level,” but I guess this may be more applicable in “professional workflows.”Host: So, do you think the future will be a universal model that can handle all scenarios with precise prompts? Or will there be more specialized models, such as those specifically designed for “futuristic styles” or certain specific styles?Nicole: I have always been surprised by how broad the range of use cases supported by “off-the-shelf models” is. But as you said, in some “consumer-facing scenarios,” such as quickly sketching the appearance of an item in a room, its performance is already quite good; however, once it comes to “more advanced functional requirements,” such as producing final products for marketing or design workflows, it needs to be combined with other tools to truly unleash the model’s potential and become practical.06 The future key lies in enhancing the model’s “expressiveness”and filling the capability gapHost: Let’s broaden our perspective and talk about the entire “image model field.” Since the emergence of Stable Diffusion and Midjourney, the development speed in this field has been like riding a rocket. What do you think have been the “key milestones” in image generation models over the past two to three years?Oliver: It has indeed been a “rocket-like development.” When I first started working in this field, Generative Adversarial Networks (GANs) were still the mainstream method for image generation, and we were all amazed by the effects of GANs, but they could only generate images within a very limited range. For example, they could generate decent-looking human faces, but only “frontal faces.” Later, models that could “generalize generation” and be “completely controlled by text” began to emerge, but initially, they were small in scale, and the generated images were quite blurry. However, at that time, we realized: “Wow, this thing is going to change everything,” and everyone began to invest energy into research. But no one could have predicted how quickly it would progress. I believe there are two reasons behind this: first, many top teams are tackling these challenges, and second, the drive of “healthy competition.” When seeing other teams release outstanding models, everyone gets motivated, such as “Midjourney was far ahead for a long time, and the results were astonishing,” prompting us to ponder, “How did they do it? Why are the results so good?”Additionally, the emergence of Stable Diffusion as an open-source model has revealed the “potential of the developer community,” showing that so many people want to develop new things based on these models. This has undoubtedly been another “explosion point.” However, to be honest, working in this field is both interesting and somewhat “frustrating”: on one hand, models are advancing rapidly, while on the other hand, user expectations are continuously rising. Now users complain about some “minor issues,” but you think to yourself, “Oh my, do you know how much effort we put into optimizing this model? A year ago, the generated images were completely unrealistic, and everyone was amazed at that.” I must say, human “aesthetic fatigue” towards new technologies comes very quickly.Host: Why was Midjourney able to be the “benchmark” in this field for so long? It seemed to be the industry standard for quite a while.Oliver: I believe Midjourney figured out “how to conduct subsequent training on models” earlier than other teams, especially “how to generate stylized and artistic images through subsequent training.” This is precisely their core advantage, focusing on “allowing users to control image styles” and ensuring that “regardless of what content is generated, the visual effects are outstanding.” At that time, this was crucial: because if you could focus the generation range on the “beautiful images” niche, you could achieve better results in that area. Starting from “focusing on high-quality stylized images” was a very good strategy for them. Later, all models, including Midjourney (such as Flux, GPT image models, etc.), began to “broaden the generation range,” and now they can generate more categories of images while maintaining high quality.Host: What has allowed models to “broaden the generation range” and no longer be limited to generating those filtered high-quality images?Oliver: There are many reasons. First, we all figured out “what the training data should look like”; second, the scale and computing power of the models are naturally growing, and things that were previously impossible can now be achieved due to “increased scale.”Host: The progress in image models has been significant, but I am currently uncertain whether we are “only left with 10% of improvement space” or if “three years from now, we will look back and think, ‘How could we have thought those models were so good, that’s ridiculous.'” What is your take on this issue? And now that the generated images are already quite good, I can’t even imagine what the “next tenfold improvement” will look like.Oliver: I believe we still have a long way to go. Not to mention other application scenarios, just in terms of “image quality,” there is enormous room for improvement. I think key progress will be reflected in the “expressiveness of the model”: currently, we can perfectly generate certain content, and the generated images are almost indistinguishable from real images; however, once we exceed the range of “commonly generated content by users,” the image quality drops sharply. For example, prompts that require “more imagination” or “fusion of multiple concepts” often yield poor results. Therefore, I believe future models may exhibit this trend: “the best image quality now and the best image quality a few years from now may not differ much; but the ‘worst image quality’ now will be significantly worse than the ‘worst image quality’ a few years from now.” We will make the model more practical and applicable in a wider range of scenarios. Moreover, we find that the broader the model’s applicability, the more use cases users can discover, and the more useful the model itself will become.07 In future workflows, traditional tools and AI models will coexist for a long timeHost: You provide both models and APIs; how do you determine which features are suitable for inclusion in a general chat tool like Gemini, and which are better left to other specialized products to implement?Nicole: I believe the positioning of these two types of scenarios is completely different. We find that users use Gemini for “rapid iteration”; for example, someone on our team wants to redesign their garden, so they first generate an effect diagram in Gemini to imagine possible appearances, and then collaborate with a landscape designer to refine and realize that idea. Therefore, Gemini is more like “the first step in creative conception” and is rarely used as “the final production tool.”However, for advanced users (such as developers), they will build more complex tools, linking multiple models together for use, which is a more precise and complex “multi-tool collaboration process.” The advantage of chatbots lies in “helping you kickstart creativity and providing inspiration,” and they can support many “interesting and easily shareable” scenarios, such as sharing creative results with family and friends. I believe this positioning will remain because advanced users with higher demands will always tend to use “more visual” or “more professional” tools.Host: How should the “editing workflow” be integrated into this? Using AI to generate initial ideas is great, but to refine a piece from 95% to 100%, do you think we will still need to rely on traditional editing tools? Or will the entire workflow change?Oliver: I think this largely depends on the type of user. Some users have “pixel-level precision requirements” for their results, and for these needs, we must integrate the model with existing tools (such as various Adobe products); while some users need “inspirational prompts” and do not have such strict requirements for the results, for them, quickly generating ideas in a chatbot is sufficient. Therefore, both application scenarios are important.Nicole: Regarding “pixel-level control,” I learned about a case just two days ago: when creating advertisements for different products or brands, the “direction of the model’s gaze” can significantly impact the message conveyed by the advertisement, as the audience’s attention is guided by the model’s gaze. Achieving such fine control is difficult with a chatbot. Therefore, for such users and scenarios, we will still need “professional tools” and “extremely high precision control capabilities” in the future.Oliver: Ultimately, the key lies in “which needs can be clearly described in language and which cannot.” Language is well-suited for conveying “macro ideas,” but if you want to move an element “three pixels to the left,” describing it in words feels awkward. Therefore, I believe that “traditional tools” and “AI models” will coexist for a long time.Host: Yes, if we observe the complete workflow of professional artists or creators, we find that they often struggle to describe their operations precisely in words; much of it is done “by feel.” In Google’s internal discussions, what products or businesses do you most look forward to seeing this image model implemented in?Nicole: I see many directions. First is the creative field, such as “photo applications,” where editing directly in the photo library would be very convenient. For example, I often need to create birthday cards from family photos a few times a year, and if I could do that directly in the photo application, it would be very convenient.Additionally, “knowledge-based scenarios” also hold great potential. In various Google products, if a 5-year-old child wants to understand “photosynthesis,” but there are no suitable visual materials online, the model could generate a dedicated diagram, opening up many new scenarios and opportunities for “personalized visual learning,” as many people are “visual learners.”Oliver: I think “office collaboration (Workspace)” is also a great direction. For example, in PowerPoint and Google Slides, people may be able to create “more engaging presentations” in the future, rather than the monotonous “text lists.”Host: When I started working, I did consulting, and if I had this feature back then, it would have been fantastic. I understand the pain of spending a lot of time adjusting formats.Nicole: Previously, when making slides, I had to first sketch on a whiteboard to determine titles and chart placements (for example, “put this dataset’s chart on the left”). If I could feed these requirements to an LLM and let it handle these tedious tasks, that would be incredibly exciting.Oliver: You could even “take a photo of the content on the whiteboard” and let the model recognize it.08 The future will see all teams moving towards the direction of “universal models”Host: What is the relationship between image models and video models? Are their developments independent, or do they borrow from each other? Is there much interaction between these two fields?Oliver: Their connection is very close. I believe that in the future, all teams will move towards the direction of “universal models (Omni Models)” that can handle multiple tasks. These models have many advantages and may become mainstream in the long run, although I am not sure. But it is certain that many techniques we have learned in the field of image generation will be applied to video generation models, and vice versa. This is also one of the reasons why the video generation field can develop rapidly. The entire industry has already mastered the problem-solving approaches for these issues. So I see them as “close partners” that will share many technologies and may even “merge together” in the future.Host: When you mention “technology,” do you mean that the core technical frameworks behind image and video models are similar?Nicole: From a workflow perspective, people often use these two types of models “complementarily.” For example, if you are a filmmaker, the early creative iteration often starts with organizing ideas in an LLM, then quickly generating frame images in an image model, which is faster and more cost-effective, before finally entering the video production phase. Therefore, even from the perspective of “workflow and usability,” these two types of models have complementarity. Additionally, many problems they need to solve are similar, such as “consistency”; whether in images or videos, it is necessary to ensure consistency of characters, objects, and scenes. However, the video field is more complex because it requires maintaining this consistency across multiple frames.Host: What do you think are the core issues that need to be addressed in the video model field moving forward?Oliver: I believe the first is to give video models “the same level of controllability as the latest image models,” which will greatly impact the development of the video field and is a direction worth paying attention to. Secondly, video teams are also continuously optimizing “resolution” and “long-term consistency.” Of course, “having the same character appear in multiple scenes” is also one of the most urgent needs of users. Therefore, the future development direction is clear: moving towards “longer and more coherent video content.”Host: Will the market landscape of image models ultimately evolve to be dominated by a few leading players, similar to the LLM field?Oliver: That’s a good question. So far, I believe there is still a possibility for “small teams to produce top models” in the image field. We have seen many small labs develop outstanding models. I hope this situation continues because the participation of small teams will make this field more vibrant.However, as I mentioned earlier, the “world knowledge reserve” and “practicality enhancement” of image models heavily rely on “scale effects,” especially the scale of LLMs. Therefore, I speculate that in the future, teams capable of training LLMs or teams that can endow image models with rich world knowledge may dominate the image field. We have seen some large labs in China also launching excellent image models, which is similar to the trends in the LLM field. So I believe that the image field will also see the emergence of such leading players in the future.Host: For image models, is there a significant disadvantage to using “the most advanced open-source models” compared to using “cutting-edge closed-source LLMs”?Oliver: That’s a great question. I think the answer largely depends on “the future development of open-source models,” as the changes in the open-source field are very rapid. About a year ago, “using open-source models” seemed like a safe choice, but now the situation may not be so clear. However, I am also uncertain about the future direction of open-source models; there is still a strong possibility that they will continue to develop and support more small labs in training high-quality image models.
