MiniMax-H3 × Z-Image: Spatial Detail Graft for H3 Video
A community graft transplants Z-Image's spatial attention onto MiniMax-H3 via q_norm rescaling: richer sets and textures with flat detail across joins, drop-in ComfyUI checkpoints.
What it is
Z-Image (Lumina2, a 6B image model known for exceptional texture rendering) and MiniMax-H3 both use per-head Q normalisation in attention. The graft rescales the q_norm weights of H3's later blocks so they attend to fine texture the way Z-Image does, without retraining, new knowledge, or architecture changes. Early blocks are left untouched (grafting them produces a lattice artifact in regular textures, measured rather than guessed), and K normalisation and feed-forward layers are never modified.
This is the second "marriage" in the author's line: Joy-LTX 2.5 transferred JoyAI-Echo's performance onto LTX-2.5 by weight-delta transplant. Here the donor is a different architecture entirely, so what crosses is attention statistics rather than weights, block-gated and dose-controlled, verified against same-seed baselines.
What it looks like
Demo videos, three-shot continuous takes rendered with the files on this page:
Measured on chained scenes, the extra detail stays flat across joins: a 0.99 high-band ratio over 3 joins vs 1.11 on stock, i.e. no per-shot sharpening creep.
ComfyUI usage
Drop the file where your H3 checkpoints live and pick it in your loader. Every H3 workflow works unchanged, including the MiniMax-H3 Multishot seamless-chain canvases. The ComfyUI-ready builds (bf16 / fp8 / int8 / w4a8 / nvfp4) load with the plain Load Diffusion Model node on ComfyUI 0.32+; the GGUF repo offers curve bakes for smaller cards:
| file | fits |
|---|---|
*-curve-zs05-Q8_0.gguf | 32 GB |
*-curve-zs05-Q5_1.gguf | 24-32 GB |
*-curve-zs05-Q4_0.gguf | 16-24 GB |
fl2va vs ref2va: identical choice to stock H3, ref2va when identity and voice must persist (reference images, voice anchoring, identity bank), fl2va when a shot must land on a supplied frame. Both chain. On RTX 30/40 cards, the author reports the GGUF builds run 4-8x faster than any 4-bit ComfyUI-ready arm on Ampere.
Verification
The author reports same-seed equivalence against the runtime-patch implementation, identity holding on faces across variants (facial texture improves over stock), 4-shot chained scenes landing scripted events without artifacts, and tensor-level verification of the graft math after baking for every file.
Comments
Sign in with GitHub to join the discussion.