<h3 id="v3-4step">v3-4step</h3><ul><li>euler / beta</li><li>4 steps</li><li>CFG 1</li></ul><h3 id="v3">v3</h3><ul><li>Res_2s / bong_tangent or euler / beta</li><li>35 steps</li><li>CFG 5.5</li><li>I like 1.5MP resolutions but any Z-Image supported resolution work well</li></ul><h3 id="v2">v2</h3>Continue fine tuning from V1, focusing on anatomy with some success, there are still certain poses and positions that need work. My goal is to achieve anatomy accuracy without fully lobotomizing the model into being nothing but goon material. Still recommend similar settings as shown below. The model handles 1.5-2.0 MP generations quite well and I've been experimenting with the <a target="_blank" rel="ugc" href="https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill">z-image-fun-lora-distill </a>in the workflow, works quite well I haven't settled on the best settings between using the lora and not to get similar results between the two, but definitely throw it in and speed up those workflows.<h3 id="future-goals:">Future goals:</h3><ul><li>Continue tuning for anatomy, additional nsfw concepts, specific art styles, better camera tech. understanding (eg. camera types, lighting techniques, iso, aperture)</li><li>Add GGUF, FP8 versions</li><li>Distilled version once satisfied with finetuned base</li></ul><h3 id="additional-info:">Additional Info:</h3>I added some training images that were solely for negative prompting, specifically Pony and Flux style images, to help combat that AI look. In the negatives try using these:<pre><code>negatives: aig3n, ponystyle, fluxface,</code></pre>It's subtle, but seems to work with some generations.Cheers and Enjoy :)<h3 id="v1">v1</h3>The first iteration of many I hope. A full finetune of <a target="_blank" rel="ugc" href="https://github.com/Tongyi-MAI/Z-Image">Z-Image</a> on my personal curated dataset using <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/musubi-tuner">Musubi-Tuner</a> on <a target="_blank" rel="ugc" href="https://www.runpod.io/">Runpod</a>There are still some anatomy/detail features lacking as you'll find. But, just like the base model, it's quite versatile. The finetune is currently still running as of 2/4/26 and plan to have FP8/GGUFs versions.I have not done any testing when it comes to LoRas, so please, if you find something, like always, my comments and dm's are open and I appreciate any input.Recommended settings or at least some that have worked for me in ComfyUI and check showcase images for captions/settings.<ul><li>dpmpp_sde/simple or beta</li><li>er_sde/simple or beta</li><li>30 - 40 steps</li><li>CFG 4-5</li><li>HRF 1.5 latent upscale: 12 steps, CFG 4, heunpp2/sgm_uniform or dpmpp_sde/simple</li><li>Denoise 0.55-0.65 depending on scheduler</li><li>Upscaling: USDU using <a target="_blank" rel="ugc" href="https://civarchive.com/models/2201950/daf-zit">DAF-ZIT</a> (Z-Image-Turbo)</li><li>2x, 5 steps, Denoise .20, tiles 1280, ddim/ddim_uniform</li></ul><h3></h3>

Files with SHA256: 4f1369ab9348cfe77f3940ac633c27b9e3b9121e28940843551b73deeeca3d38