Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:<ul><li>☕ <a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Buy me a coffee on Ko-fi</a></li><li>💛 <a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">Sponsor on GitHub</a></li><li>🔁 <a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a> (recurring)</li><li>⚡ Or right here: the Civitai tip button on this page sends Buzz directly</li></ul> Z-Image's eye for texture on the Joy-LTX 2.5 engine. This merge exists for one thing: surfaces, textures, fine details and foliage. Hay holds individual strands at mid-distance. Hedges render as brambles and leaves instead of green mass. Ground litter stays physical, fabric keeps its weave, skin keeps its unretouched grain. Identity, motion, sound and speed stay exactly what Joy-LTX 2.5 already was - same loaders, same settings, same seeds.Honest calibration: the comparison images on this page are same-seed pairs at matched timestamps, and that's where the difference is obvious - pause and look. In motion it reads as "richer" rather than dramatic. Measured: +8-23% fine-detail energy on matched frames against stock v2.<h2 id="what-is-this-actually">What is this, actually</h2>Z-Image is a 6B image model with exceptional fine-texture rendering. Its spatial attention profile - how sharply each depth of the network commits to fine detail - is measured offline and transplanted onto Joy-LTX 2.5 by rescaling the engine's per-block attention normalization by small factors (about ±10% at the shipped dose). No retraining, no new knowledge, no Z-Image weights at inference - the engine keeps everything it knows and attends to texture the way Z-Image does. This is the third marriage in this line: Joy-LTX 2.5 put JoyAI-Echo's performance on LTX-2.5's engine, MiniMax-H3 x Z-Image proved the texture graft on H3, and this page brings that graft to the LTX engine.<h2 id="which-file">Which file</h2>RTX 30 / 40 - take a GGUF (needs ComfyUI-GGUF; on Ampere these run 4-8x faster than the emulated 4-bit comfy-native arms):<ul><li><code>Q8_0</code> - 21.2 GB - 32 GB cards, closest to full precision.</li><li><code>Q5_K_M</code> / <code>Q6_K</code> - 14.8-16.5 GB - the 24 GB picks.</li><li><code>Q4_K_M</code> - 13.2 GB - the 16 GB pick.</li><li><code>Q3_K_M</code> / <code>Q2_K</code> - 7.4-9.9 GB - low-VRAM, visible quality tax.</li></ul>RTX 50 - take a comfy-native (stock Load Diffusion Model, ComfyUI 0.32+):<ul><li><code>comfy-int8</code> - 20 GB - the fastest file on Blackwell, 32 GB cards.</li><li><code>comfy-fp8</code> - 20 GB - the fp8 twin.</li><li><code>comfy-w4a8</code> / <code>nvfp4</code> - ~11-12 GB - the 16 GB family (nvfp4 is Blackwell-native, emulated elsewhere; w4a8 is the quality pick).</li><li><code>comfy-mix4x8</code> - 13.8 / 17.0 GB - VRAM-fitted mixes.</li></ul>The master: <code>bf16</code> - 39.1 GB - quantize your own cuts from it. Every format lives on the Hugging Face repo pair (joeygambino) if it isn't attached here.<h2 id="install">Install</h2>Put the file where your Joy-LTX 2.5 checkpoints live. Pick it in your loader. Done - every Joy-LTX 2.5 workflow works unchanged, including the Multishot AV-extend canvases.<h2 id="links">Links</h2>All formats: the GGUF and comfy-native repos on Hugging Face (joeygambino). The H3 version of this graft: the MiniMax-H3 x Z-Image page. Questions: comment here - I answer.

Files with SHA256: d26d9455850540429216e95e8a104684e4de366ec834997649ed96c155f70e4a