Trying to make real pixel art, but stuck on VAE. Is the pixel space model the way to go?

The Forge / 工坊 Models / LoRA / 模型与 LoRA
1

I’m making assets for a game, and I need strict pixel art, not that blurry mess. Most models can’t do real pixels—the scaling is all over the place, some tiles are 14 pixels, some are 16, nothing lines up. After digging into it, the problem is the VAE: models don’t generate pixel-by-pixel; they first output a low-res latent space about 1/8 the size of the final image, then the VAE reconstructs it, and that step destroys pixel consistency.

I’m now running Flux base with a custom-trained LoRA, using a dataset of 8x downscaled pixel art. The results look almost perfect after shrinking.

I heard there are models that skip the VAE entirely and work directly in pixel space, like Chroma1-Radiance, but it’s obscure and hard to find info on how to run it or train LoRAs for it. Anyone here trained a pixel LoRA on a pixel-space model, feeding the dataset raw without scaling?

2

That analysis about VAE messing up pixel recovery is spot on — I’ve been wondering why that happens for ages.

3

I’ve been using that self-trained LoRA 8x upscale trick too, the results are legit.

4

The pixel space model is pretty niche, there’s really not much info out there.

5

Real pixel art and blurry pixel crap are two completely different things. Anyone who makes games knows that.