Fixing the Friction: Building the "Missing Link" Between GenAI and Professional Art Direction

Yifei. I'm a commercial filmmaker and VFX Art, and I was I'm also a, cofounder of, this visual studio you're seeing on the SWE page. It's called MediumRare Imagery. It's a studio based in Shanghai. It's like a production house. And I, moved to this country, this city last November. And, I, I have been loving this city ever since. And, and now I run this company remotely from here. And as you can probably tell, I'm not SQLite, Frontend engineering or coding Co-founder. I'm what they call the AI coder. So that that meets, trust me, I came about this close to accelerate accidentally showing you my tool from local HINTS 3,000. And, so, Date AI gonna show you, this tool I built. It's called SousVide. It's basically an AI image, image editing a AI image generation tool. I built Sous Vide, to tackle some real life problems, from my friends and colleagues, back in Shanghai, in this very niche industry. The the the the they're from they're from, like, advertising agencies and GraphQL studios and also, like, ecommerce endianness, they need to wrap their lead around how AI world, but most of the AI image tools we see nowadays are, are lead. Should I say that? And, AI, those developers and platforms, they they promote those commercially sexy features, to sell subscription to use. But, in real life scenario or, like, in professional, AI, only, according to my coordination, CLI AI about 15 of the features are actually used. The other 85%, I mean, 15% of the, functions are actually used. The other 85% are, merely UX SQLite or or should I say, like, distraction. And so, my solution to build this, sous vide is, like, I want to I want to build something that is, like, very minimalist. And SWE don't do, like, infinite canvas AI most of the, platforms that does Date. And we don't want you to get lost in the infinite canvas. And we don't do, like, Nano base GraphQL because we think that's a little bit overwhelming when you're trying to create for real. And we use a clean, masonry layout Banana simple drag and drop based logic. So we try what we try to do is, like, trying to 2026 trying to give the the remains canvas back to the creator, trying to give the, freedom of creation, back to the creator himself. Okay. Now I'm gonna do a little bit of, demoing and, let's see. So AI I said, everything is basically drag and drop based. You can basically drag and drop from anywhere to everywhere, like, AI sorry. I I don't have a picture here. Maybe maybe I'll try this 1. Yeah. You can you can drag and drop to the ref image box. You can drag this tool, you you can Drag from, Pinterest Date, and you can drag this down to the prompt box to refer to it. You can drag it into the describe image box to have to interrogate the image. So everything is very simple. It's drag and drop based. AI like when we use Nano Banana, the 1 thing got me a headache is, like, trying to refer to different image. Like, I want I want this boy from image a and the the, lighting style from image b and something something from image c. That always tool me a headache Esther that's why I built this, like, drag and drop Date. So when you whenever you release, the image will be, will world be properly referred to as, like, image a, image b, or image c in here. So let's, let's try something, like, let's try something from, Intercom. Maybe AI this 1. And, how about how about how about this 1? So in this scenario, I'm gonna say, like, oh, sorry. AI to, try to tweak image b with the AI loss image a. So it's GenAI, and, once 1 sad thing about APIs, like, is it takes more or less 1 minute to GenAI. So I'll I'll try to demo the next feature while we're generating the first 1. Okay. The the next feature, I think worth noting is, like, how it builds a AI background remover into the tool. It's SWE it's a web browser based background remover, so it's AI. It's not it doesn't cost you anything ex except a little bit of computational power. SWE, let let's try, for example, something I I I tried here. This is AI a J. Cole poster. Let let's look for maybe, like, Kanye West. Oh, this is this is good. Oh, here is also something worth mentioning that is, like, you can drag and drop Pinterest image Direction to the tool, which cannot happen with with, like, majority of the tools, nowadays because these are not Tinkerers. These are links. So AI build this this little something to go deep down into the link to extract the Tinkerers. Easy for you to steal AI from other artists. Audio, yeah, let's SWE, this is the first generation we were talking tool. Let's let's do a recap. We what we did is what we did is we refer we refer to the 2 Imagery separately and we give clear instruction of how we want the generation to Brian. And, oh, sorry, I I just accidentally AI the that prompt. Okay. This is AI a picture of Kanye West. Now let's look for maybe something like a, cool background. Yeah. Let let let's just let's just keep it simple with, like, a halftone background AI here. So what what AI I build SQLite, like I said, a very highlighting. It's removed background feature. It's not supposed to be, like, pixel perfect. It's no contest to, to AI doing it with, orthodox author, like in Photoshop or anything, but it gets the job done. And, I think most importantly, this is for, a rough art directing or and scene blocking because we're we're using AI to generate the final image anyway. So, I think the most important thing is to keep the workflow clean. Meets, myself, I don't I don't really enjoy AI switching back and forth between like Professional tools and AI tools and back and forth again. So if I can do this in 1 Night take Fin 1 tool, I think it's it's pretty cool. And Oh, sorry. Does it do it? Yeah, usually usually it takes about AI 10 or 20 seconds. Oh lead, yeah, there it goes. Okay, you Sonnet does City much a terrible job, it remains the lead as world, but that really doesn't matter. And you can press option and release, you see it's snapped onto the other picture, then you can do a basic, basic blocking SWE you can really fine tune, you can really art direct how you want the final image generation to be. Lead me let me cheat a little bit by copying this, prompt here. Okay. Here we go. And by the way, we we do have, like, we do have Nano Banana and GPT Imagery 2026 embedded, on, sous Code. So you you're feel free to try both to see how they differ. And, while we're generating tool, while while we're waiting for the generation, we we can also move on to the next next function, which is AI, the, intent extraction. SWE, basically, you can drag a picture here Fin a AI image box to have it AI to have it, like, form a Scripting. But this is not AI some cutting edge technology. This is AI everyday use. But, the thing is is, like I said, it's streamlined. It's intuitive. It's very simple. And also AI build this, element SWE, with a little CLI. It world extract the 3 key elements from the prompt. We we we did see a long ass prompt in the prompt box. Right? It's really a headache, especially for some someone like me who's not from a English speaking background to look at Analog ass prompt. So so here City does the extraction for you. It says AI a stylized portrait of a man resembling Kanye West. You can, you can just change it to anyone you want to and, hit switch. Lead say Lebron James maybe. Yeah, you see here? Sonnet does the switch for use, So it's pretty simplify. And, yeah, we we we might want to wait for this 1 while we show another 1. AI, I'll I'll just show you I'll Agent show 1 of the 1 of the generation I've already done. Yeah. Here it goes. This is a feature for for me personally. This is, like, 1 of the most used feature. It's, it generates 3 by 3 collage storyboard for for a Senior. As a filmmaker, this helps a lot, but AI don't take credit from this. I mean, Nano Banana does the heavy lifting and, it can keep the AI consistent and the character consistent. So when when you generate this, you you you want you might want something to, like, split it up real quick. So this little tool I AI don't know how to categorize it. That's why I put it here in the dev toolbox. It scale collage splitter. So it React through your prompt when it sees, like, meets something like 3 by 3 or 3 times 3. So it instantly it knows you want a 3 by 3, Cuppari, so you HINTS, you can add a little bit coding and you hit Processing. And the AI images 1 by 1 will fall, onto your canvas SWE you can tweak them, manually, each 1, or you can up Presenter upscale them with, with the help of Nano Banana. And let's look at the generation agents. Yeah. Here SWE go. It's not bad, isn't it? This is AI, the, Kanye West with, his discography and the key events and everything. Yeah. That's basically it. Oh, yeah. And we have APIs, project agents, we we use attack based, product project management. I think this works best if you have like lots of Product to go through. And, like I said, I'm a AI Folder before this. My only programming experience is like in Esther Effects, in Houdini, I did a little bit of a Python and JavaScript. So it's really amazing that nowadays AI can do this entirely with Vibe Coding. I'm really grateful. And, I would like to share what I learned from building this but AI these heavy hurdles for me may be like everyday routines for you AI. So I don't think I need to go through this anymore. So that's that's SousVide. Thank you for your time. If you Banana play around and you want some free credits, just come to me and we can talk about City. Okay? Thank you very much.