@AIWarper MiniMax said the sweet spot is 16MP. Also, you can use the larger VAE from PrunaAI/P-H3-Image. Just gotta coerce the 5D (B, C, 1, H, W) tensor into the 4D (B, C, H, W).
I just did this, but it wasn't quite a drop-in replacement; I had to add an adapter. Both VAEs share the MiniMax-H3 video encoder and its 24-channel latent space (same
`latents_mean` / `latents_std`, same ImageNet pixel convention), but they differ in two ways:
* The decoder: P-H3 is a 36-layer, 2048-wide ViT; MiniMax is a 24-layer, 1024-wide one;
* The tensor layout: MiniMax decodes `(B, C, H, W)`, P-H3 treats an image as a one-frame
video and decodes `(B, C, 1, H, W)`.
Here are before & after photos. Not a terribly noticeable improvement.
@Tech2Wild I added these two flags to my SGLang-Diffusion server script, and H3-Image now consumes only 69GB, fitting comfortably on one 6000 Pro:
--text-encoder-quantization mxfp8
--component-quantization-ignored-layers.text_encoder self_attn visual
@hellorob I believe you'll need two RTX Pro 6000s, one for the massive text encoder and one for everything else. I'm gonna quantize as little as possible at MXFP8 to try to fit it on one.