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.<h2 id="what-is-this-actually">What is this, actually</h2>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.<h2 id="which-file">Which file</h2>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):<ul><li><code>curve-zs05-Q8_0</code> - 21.5 GB - 32 GB cards, closest to full precision.</li><li><code>curve-zs05-Q5_1</code> - 15.2 GB - the 24 GB pick.</li><li><code>curve-zs05-Q4_0</code> - 11.5 GB - the 16 GB pick.</li><li>plain (non-curve) <code>Q5_1</code> / <code>Q4_0</code> (and <code>Q3mix</code> on fl2va) - for workflows built on the original bakes.</li></ul>RTX 50 - take a comfy-native (stock Load Diffusion Model, ComfyUI 0.32+):<ul><li><code>comfy-int8</code> - 21 GB - the fastest file on Blackwell, 32 GB cards.</li><li><code>comfy-fp8</code> - 21 GB - the fp8 twin (ref2va also ships <code>fp8e5m2</code>).</li><li><code>comfy-w4a8</code> / <code>nvfp4</code> / <code>mxfp8</code> - ~11-12 GB - the 16 GB family (ref2va also ships <code>w4a4</code>).</li><li><code>int8_convrot</code> (ref2va) - the Lightricks-style convrot build.</li></ul>The master: <code>ref2va pruned zs05 bf16</code> - 40.2 GB - quantise your own cuts from it.<h2 id="install">Install</h2>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.<h2 id="links">Links</h2>All formats: the GGUF and comfy-native repos on Hugging Face (joeygambino). Workflows: the MiniMax-H3 Multishot page. Questions: comment here - I answer.

MiniMax-H3 x Z-Image - richer sets and textures, same identity (the spatial detail graft)

The default variant. Chains seamlessly, lands on supplied first/last frames.<ul><li><code>MiniMax-H3-fl2va-curve-zs05-Q8_0.gguf</code> - 21.5 GB - 32 GB cards.</li><li><code>MiniMax-H3-fl2va-curve-zs05-Q5_1.gguf</code> - 15.2 GB - 24 GB cards.</li><li><code>MiniMax-H3-fl2va-curve-zs05-Q4_0.gguf</code> - 11.5 GB - 16 GB cards.</li></ul>Plain (non-curve) bakes and comfy-native formats (bf16 / fp8 / int8 / w4a8 / nvfp4) are on Hugging Face.

Files with SHA256: 4f09561efd591fb9d8981a6cd8a6a9ce3abed162369b86207d916dc81948aa30