<h2 id="draw-any-anime-character-from-one-picture-no-per-character-lora-training.">Draw any anime character from one picture – no per-character LoRA training.</h2><p>Give MOSS-Image a reference picture as <em>Picture 1</em> and a prompt: CharRef v2 draws that character – face, hair, outfit and art style – in the new scene. It works on single images and on <strong>whole comic pages</strong> (MOSS JSON page layouts with panels and speech bubbles), so a character stays on-model page after page.</p><h3 id="how-to-use-(comfyui-0.42)">How to use (ComfyUI ≥ 0.42)</h3><p>The attached workflow has everything wired.</p><ol><li><p>Install the <a target="_blank" rel="ugc" href="https://github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit">ComfyUI-MOSS-Image-Ostris-Edit</a> node (reference latents + KV cache).</p></li><li><p>Load your character picture, keep the LoRA at 1.0.</p></li><li><p>Prompt: <strong>Same character and art style as Picture 1.</strong> + your scene (tags, sentences or a MOSS JSON page layout).</p></li><li><p>50 steps, CFG 3, euler / sgm_uniform, SkipLayerGuidance single layer 22.</p></li></ol><h3 id="tips">Tips</h3><ul><li><p>A few appearance words (hair style, eye colour, key outfit pieces) noticeably improve fidelity.</p></li><li><p>The outfit follows the reference – unless the prompt names other clothing, so you can change outfits by text.</p></li><li><p>Clean, large references (full body or bust, plain background) work best. Crop comic panels to the character, otherwise props and speech bubbles come along.</p></li></ul><h3 id="limitations">Limitations</h3><ul><li><p>Fine hair details can drift without appearance words. One reference per image (scenes with several characters keep only Picture 1 reliably).</p></li><li><p>Made for anime / illustrated characters: a photorealistic or 3D face is not kept reliably.</p></li></ul><h3 id="prohibited-use">Prohibited use</h3><p>Not for intimate or sexual images of real people without their consent, sexual depictions of minors, facial-recognition scraping or harmful manipulation (EU AI Act Art. 5). If you publish a picture that could pass as a real person or event, say that it is AI-generated (Art. 50(4)).</p><p>Also on Hugging Face: <a target="_blank" rel="ugc" href="https://huggingface.co/7err4/MOSS-Image-CharRef">7err4/MOSS-Image-CharRef</a>. Built for the "reference picture" option of <a target="_blank" rel="ugc" href="https://github.com/7err4/fumika">Fumika</a>, a local comic and image studio.</p>

MOSS-Image CharRef – character reference LoRA

<p>~1330 reference → target pairs (anime character art and synthetic renders, safe / suggestive only), LoRA rank 32, 1000 steps on MOSS-Image-Base bf16 with ai-toolkit edit mode (KV cache).</p><p>v2 rebalances the data of v1 (no more inflated proportions) and keeps outfits stable. Recipe idea from <a target="_blank" rel="ugc" href="https://huggingface.co/pleiades7/MOSS-Image-CharRef-PoC">pleiades7/MOSS-Image-CharRef-PoC</a>.</p><p>Showcase: CharRef v2 at 1.0, 50 steps, CFG 3, euler / sgm_uniform, SLG single layer 22, 832×1216, one seed per scene (no cherry-picking). The two sheets show the reference on the left.</p>

Files with SHA256: 1a77e98228242d7ca79dd01df737d751f211f2ae676f16cce664987dbf0bac05