MiniMax-H3 × Z-Image — the spatial detail graft
Z-Image's eye for texture on MiniMax-H3's engine. These are drop-in replacements for the
standard MiniMax-H3 checkpoints: same identity, same voices, same speed, same VRAM, same
workflows — but sets and surfaces render noticeably richer. Peeling paint peels harder, rust
bleeds further, water carries more light. Measured on chained scenes, the extra detail stays
flat across joins (no per-shot sharpening creep: 0.99 high-band ratio over 3 joins vs 1.11
on stock).
What it looks like
(demo videos — three-shot continuous takes rendered with the files on this page)
How it works
Z-Image (Lumina2, a 6B image model with exceptional texture rendering) and MiniMax-H3 both use
per-head Q normalisation in attention. This graft transplants the magnitude profile of
Z-Image's spatial attention — how sharply it commits to fine texture — onto H3's later blocks
by rescaling q_norm weights. No retraining, no new knowledge, no architecture change: H3
keeps everything it knows and attends to texture the way Z-Image does. Early blocks are left
untouched (grafting them produces a lattice artifact in regular textures — measured, not
guessed), K normalisation and feed-forward are never touched.
This is the second marriage in this line:
Joy-LTX 2.5
put JoyAI-Echo's performance on LTX-2.5's engine by weight-delta transplant. Here the donor
is a different architecture entirely, so what crosses is attention
statistics rather than
weights - a mechanism of our own, block-gated and dose-controlled, verified against
same-seed baselines. A nod to TenStrip, whose H3 attention experiments sparked the
question of what an image model could donate.
Which file
Same picking rules as standard H3 GGUFs. curve = the current recommended bakes.
| file | fits | notes |
|---|
*-curve-zs05-Q8_0.gguf | 32 GB | closest to full precision |
*-curve-zs05-Q5_1.gguf | 24–32 GB | the 24 GB pick |
*-curve-zs05-Q4_0.gguf | 16–24 GB | the 16 GB pick |
| plain (non-curve) variants | as above | for workflows built on the original bakes |
fl2va vs ref2va: identical choice as stock H3 — ref2va when identity/voice must persist
(reference images, voice anchoring, identity bank), fl2va when a shot must land on a supplied
frame. Both chain.
Install
Drop the file where your H3 checkpoints live, pick it in your loader. That's the whole
install — every H3 workflow works unchanged, including the
MiniMax-H3 Multishot seamless-chain
canvases. comfy-native versions (bf16 / fp8 / int8 / w4a8 / nvfp4 and more):
see the companion repo.
Verified
- Same-seed equivalence against the runtime-patch implementation (differences at run-noise level)
- Identity holds on faces across variants; facial texture improves over stock
- 4-shot chained scenes: detail flat across joins, scripted events land, no artifacts
- Every file: tensor-level verification of the graft math after baking
Made by joeygambino. Questions: open a discussion — I answer.
