minimax h3 image edit
Surpassing banana2!
4B text encoder, W4A8 DiT, 6 steps, no CFG, up to 16 reference images.
On an RTX 3080 (20 GB) with 32 GB DDR4, a 2048×2048 image takes roughly 20–30 seconds.
minimax_h3_image_edit.json