Deep Fish

it’s honestly kind of wild how fun it is to just iterate on a character identity with ai watching something evolve over time from a rough idea into something that actually feels coherent is a totally new kind of creative loop without tools like fal this just wasn't possible. and even wasn't possible until around a year ago in a reasonable amount of time / effort I never really bought the whole ai is going to replace hollywood / artists thing that feels like such a narrow read of what’s actually going on what does feel real is this shift in how we express and transmit ideas. gen ai isn’t about making content faster it’s about making ideas more sticky, more personal, and more direct the ai slop shit is just a phase like most things but underneath that, something way bigger is happening to me learning feels different. creative expression feels different. the barrier between thinking something and showing someone that thing is collapsing into the same point that’s the interesting part to me. expanding who gets to create and how ideas move between people it feels a new medium entirely a more personal way of communicating

If you want to build a node editor with react flow you should check out the Deep Fish repository. Left is showcased open source react flow component vs. deep fish's on the right. github.com/noahgsolomon/deep…

Deep Fish is now Open Source! @deepfishlol My goal was to build the web version of comfyui that had the extensibility of touch designer. I ended short so graph execution is not recursive and remains client directed. Go drop it a star and if anyone needs help plz dm :D

going to open source @deepfishlol this weekend

This is what i've been working on for the PAST 2 MONTHS!! Fal Sandbox where you can run models on the same input and see the (shareable) generations in one view! fal.ai/sandbox Hope you guys like it (another 20 trillion to fal) 🥹🫶

Sunsetting development on @deepfishlol. When a side project begins to feel more like a chore than an expression of curiosity I think that's when it's the best time to move on. I've began to work on a heart rate / hrv wristband monitor last week and want to move more in this direction (intersection of health monitoring hardware and data analysis) for my side projects through the next year. If u want to help and live in sf lmk :D

I love my stupid pfp

Real time persistence maxxing 🐟⋆｡𖦹°🫧˖ there's like 0 local state for workflows not directly anchored to server state now (much needed improvement!). (for workflows ✅ for flows 🔜) changelog 08/31 - @deepfishlol :D

User downloaded .deepfish files for workflows are going to be encoded in characters like ⋅˚ ଳ ˚ 𓆝⋆｡˚ 𓇼 ⋅˚ ଳ ˚ 𓆝⋆｡˚ 𓇼 so when u open the file it’s just an ocean in text I think it’s pretty cool 😸

Character design in vintage anime style using Dreamina 3.1

It’s always such a warm feeling when u see an app u have rank for new keywords on google

Going to give a shot at making a short anime episode 100% AI this weekend that doesn't feel AI. If you'd like to try too here's the way of doing so with consistent style, consistent voices, limited artifacts in video scenes, and captions that i'm thinking of from models that are all on @fal (minus RVC right now): Each scene needs to be grounded with a start and end frame. We can get the first frame from Nano Banana from google. However for text to image, it's not the best model for aesthetics. So I'll use Dreamina 3.1 to get the initial character designs and environments and overall style/aesthetic from text prompts and from then on use Nano banana to make variations of those characters and environments. For video generation we could either do a start+end frame video interpolation approach but the problem with that for dialog heavy scenes is it's not entirely automatable to guarantee generated audio and interpolated video match pacing or even length actually. And even if they match on both fronts, then you need to use an external lipsync model which always has artifacting (and is trained for realism so idek how it would do for like a dragon talking for example, I imagine pretty bad). The best approach for the video side I think is start frame into veo 3 is the best approach. Then with that audio+video output, transcribe who is speaking when in the video with the eleven labs transcription segment api and then isolate the vocals during those time segments, and run the vocals through a rvc model for the voice you want to use for that character (for consistent voice) and combine the rvc outputted vocal with the vocal removed audio and replace the audio for that scene with the voice remapped audio. And for captioning, there's a transcription + video editing model on fal to add captions to the video.

My 3 favorite text to image models for vintage anime aesthetic are Qwen Image, Dreamina and Wan 2.2 I made a flow for my 3 favorite anime models you can go try for free with different characters and backgrounds. link below

Nano Banana is really awesome but I'm not impressed with it's aesthetics for text to image. I still think the best text to image model for aesthetics in general is wan 2.2 right now

I found a cool use case for nano banana prompt below

if 2 comparable things overlap a lot in what you can see but don’t actually share much under the hood then the only honest way to talk about their difference is just to point at what comes out of them. the observable shape of most of these image editing models for example is basically the same with different quality magnitudes (which are temporal qualities so don't count as the observable shape). the only way to compare these models in the same classification is to unify the user interface and let all or a subgroup be able to be used at once on the same user query. I think this idea of converging interfaces for classes of "things" applies to a lot outside of just different emerging generative media operation categories. Like software generally. The only way to have ur users return again and again who are interested in ur product is to make something truly novel either in its idea or the emotions it evokes in ur users. This is why I'm so bullish on making silliness and whimsy a first class citizen in design decisions now bc it's the easiest way to alter the emotion evoked by ur product without changing the fundamental interface of the tool ur offering to ur users which u want to be grounded in their already existing habits.

this, next.js and nettspend <3