Download
3 variants available
153
487
14.2K
MiniMax H3 is licensed by MiniMax under the MiniMax H3 Community License Agreement. That agreement’s Applicable Territory excludes the European Union, the United Kingdom, the Republic of Korea and the United States of America. Your use of H3 and of any H3 derivative is subject to that agreement and its Acceptable Use Policy.
MiniMax H3
Z-Image's eye for texture on MiniMax-H3's engine. Drop-in replacements for the standard 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.
And the extra detail stays flat across chained shots - measured on the same seed, the graft ran a 0.99 texture ratio over three joins where stock drifted to 1.11. More detail, no per-shot sharpening creep from the graft.
What is this, actually
Z-Image is a 6B image model with exceptional fine-texture rendering. Both it and MiniMax-H3 normalise attention queries per head - so the shape of how sharply each attention head commits to fine detail can be transplanted between them by rescaling those normalisation weights. That's the whole trick: no retraining, no new knowledge, H3 keeps everything it knows and attends to texture the way Z-Image does. Early blocks are deliberately left untouched (grafting them creates a weave artifact in tile-like textures - we measured, so you don't have to). This is the second marriage in this line - Joy-LTX 2.5 put JoyAI-Echo's performance on LTX-2.5's engine the same way, donor to engine. The statistics-transplant mechanism here is our own; a nod to TenStrip, whose H3 attention experiments sparked the question of what an image model could donate.
Which file
Two versions on this page - fl2va (default) chains and lands on supplied frames; ref2va adds reference images, voice anchoring and the identity bank for characters that must persist. Every format is attached to its version:
RTX 30 / 40 - take a GGUF (needs ComfyUI-GGUF; the Multishot pack's loader takes both formats):
curve-zs05-Q8_0- 21.5 GB - 32 GB cards, closest to full precision.curve-zs05-Q5_1- 15.2 GB - the 24 GB pick.curve-zs05-Q4_0- 11.5 GB - the 16 GB pick.plain (non-curve)
Q5_1/Q4_0(andQ3mixon fl2va) - for workflows built on the original bakes.
RTX 50 - take a comfy-native (stock Load Diffusion Model, ComfyUI 0.32+):
comfy-int8- 21 GB - the fastest file on Blackwell, 32 GB cards.comfy-fp8- 21 GB - the fp8 twin (ref2va also shipsfp8e5m2).comfy-w4a8/nvfp4/mxfp8- ~11-12 GB - the 16 GB family (ref2va also shipsw4a4).int8_convrot(ref2va) - the Lightricks-style convrot build.
The master: ref2va pruned zs05 bf16 - 40.2 GB - quantise your own cuts from it.
Install
Put the file where your H3 checkpoints live. Pick it in your loader. Done - every H3 workflow works unchanged, including the MiniMax-H3 Multishot seamless-chain canvases.
Links
All formats: the GGUF and comfy-native repos on Hugging Face (joeygambino). Workflows: the MiniMax-H3 Multishot page. Questions: comment here - I answer.