Hacker News Reader

Comment by starwaver | Hacker News Reader

I like the idea of node based image editing. It reminds me of when learning how to write shader code for games and GLSL was waaaay too over my head until I discovered node based shaders, and I didn't have to wrangle with coding and instead focusing experimenting with different nodes.

However soon creating a "shader that works" was no longer an issue but how to create X effect using shaders was my next blocker, and luckily there were ton of YouTube tutorials on these, which was very helpful, but this continues to be a pain point even now

Since now we are in the age of AI, would it be possible to prompt something like "create me a workflow to take image A, a concept art of a character and convert into into a walking animation sprite sheet with 16 frames for each animation walking up, down, left, right and all diagonal directions" and have it not only generate the result, but a workflow to create the result so it can be edited and tweaked.

View on HN

Oh yea I know what you mean. There are several parallels here with shader nodes for sure. We've been thinking about a voyager/agent-style approach where an agent can start to learn "skills" where skills are individual blocks. Each skill represents a certain function applied to an image and based on a specific instruction set we should be able to craft a sequence of actions that will lead to that result.

One way to leverage that is building the graphs via a prompt, but another way might be to not think of the workflow as a pre-constructed graph at all. Rather perhaps we build dynamic graphs whenever you ask for a certain action - like a conversational image editing interface.

So you say something like make the woman's hair purple. We apply segmentation to the hair, and then add a puple color overlay exactly to that area.

I prototyped a system one year ago which created node-based, editable scene graphs in-browser with the help of an LLM-powered agent. It allowed the user to quickly construct complex 3D scenes.

It was some amazing tech with a ton of applications, but sadly leadership had other plans and pivoted to a highly derivative, slapped-together AI sex bot.

Theoretically yes, with a few limitations.

The walking animation is going to be a lost cause without specific inputs. We can do ControlNet stuff to make a character match a pose, and you can supply a series of poses that represent the walking animation.

On some level it seems silly to try and get anything to generate the workflow to do that. What you really want is a workflow to generate an image off of a pose, and then pass in the poses you want. Side tangent, I don’t know why the ai generation community has decided “workflow” is what they’re going to call “functions”?

After that your problem is that the results will be kind of meh. And that’s the brunt of where it’s at right now. You can make assets that satisfy descriptive conditions. But you can’t demand they be good. And you can’t demand they be consistent across different drawings. Can you hire an artist to fix your generated directionally correct assets? Yeah, maybe. Sounds depressing and error prone though.

Check out ComfyUI for a much more advanced and open source version of this.

https://github.com/comfyanonymous/ComfyUI

Not really an apples-to-apples comparison. ComfyUI is for diffusion-focused workflows, this is not.

Plus you don't need a local GPU for this. I realize this is a Pro for some Con for others, so there can be different products in the market serving different needs.

Reply on news.ycombinator.com