AI materials. any progress recently?
sublime tool
Does anyone know if anything cool appeared in AI textures/ materials field? Ai normal and height map from photo?
I tried recent Sampler and see no difference , still same blurry , soft, unspecific normal maps . Any better alternatives recently? I sort of lost my hope and stopped to invest my time but still wonder if anyone follow it up closely.
I have an impression some texture providers started to use Sampler instead of actual hi-res phtoscanning and you figure it out only after buying.
Replies
The problem is they mostly rely on illegally-scraped datasets. If you're OK with the ethical and legal issues of that (I'm very much not cool with it) then have at it.
https://pypi.org/project/imaginairy-normal-map/
https://github.com/brycedrennan/imaginAIry?tab=readme-ov-file#depth-map-control
There are ethically sourced apps though, for example
https://hugotini.github.io/deepbump
1. Smart materials are king, and honestly the smaller and more selective your library, the better
2. You already have a infinite supply on (scanned) textures essentially, although some things might be rare
The one use I can see is a full texturing pass on an asset, but then again thats "standard workflow" which personally I think is outdated, dosnt scale with project size well and is unflexible. COD is now going 300 GB apparently and many games have to make a lot of cuts in memory
(On the other hand, we will see ML compression soon based on some research which might cut texture size by 5-10x apparently)
I assume the real niche with AI materials is in junior artists or programmers doing 1-2 man games and just cobbling something together with standard workflow or such. Although I could see some sort of pipeline where you generate a height map for a decal or something like that or starting point for a designer graph. Maybe I just dont see the vision yet, anything is possible, but the more I get better, the more I value extremely compact and extremely polished libraries, which feels like the opposite.
Edit: Ok, Hand painted or special styles, thats definitely a big case but then really requires consistency.
These are things that can't be done perfectly with generalised standard algorithms as they require intuition about the content of the image. ML models can provide that intuition.
the nice thing about reducing the problem space to focused tasks like this is that you can reasonably collect/create the training data required to make a decent model without stealing shit off people.