7err4/MOSS-Image-CharRef

MOSS-Image CharRef v2

A character-reference LoRA for MOSS-Image Base: give it one picture of a character as Picture 1 and a prompt, and it draws that character – face, hair, outfit and art style – in a new scene. No per-character LoRA training needed.

It works on single images and on whole comic pages (MOSS JSON page layouts with panels and speech bubbles). It is the "reference picture" option of the Fumika app.

Usage (ComfyUI ≥ 0.42)

Needs the ComfyUI-MOSS-Image-Ostris-Edit node (reference latents + KV cache).

1. MOSSImageOstrisEditModelPatch (kv_cache on) → LoraLoaderModelOnly with this LoRA at 1.0
2. TextEncodeMOSSImageOstrisEdit with your reference as image1 and a prompt that starts with Same character and art style as Picture 1. followed by the scene (tags or sentences, or a MOSS JSON page layout)
3. negative: the same encoder with an empty prompt
4. 50 steps, cfg 3, euler / sgm_uniform, SkipLayerGuidance on single layer 22 (MOSS full mode)

Tips:

• A few appearance words in the prompt (hair style, eye colour, key outfit pieces) noticeably improve fidelity.
• Outfits follow the reference unless the prompt names other clothing – so you can change outfits by text.
• Clean, large references (full body or bust, plain background) work best.

Examples

Reference (left) → three new scenes, CharRef v2 at 1.0, MOSS-Image full mode (50 steps, cfg 3, sgm_uniform, SLG single layer 22), 832×1216, prompt = prefix + scene tags + appearance words. No cherry-picking: one seed per scene.

Training

• ai-toolkit edit mode (arch: moss_image, edit: true, kv_cache: true), LoRA rank 32 / alpha 32, AdamW8bit 1e-4, batch 1, 1000 steps, 1024 px, MOSS-Image Base bf16.
• ~1330 reference → target pairs: anime character art and synthetic renders of anime characters in many scenes, and renders of well-known adult anime characters with mixed body types (2 references + 8 scenes each). Safe / suggestive only, no nudity; captions = prefix + outfit / scene / pose tags, identity never named (it has to come from the picture).
• Recipe idea from pleiades7/MOSS-Image-CharRef-PoC (v1 of this LoRA showed inflated proportions; v2 rebalanced the data).

Limitations

• Fine hair details (e.g. a single coloured streak, exact ponytail) can drift without appearance words in the prompt.
• One reference per image: scenes with several characters only take the identity of Picture 1 reliably.
• Inherits MOSS-Image's limits (small text, hands).
• Trained on anime art: photo and 3D styles survive, but a photorealistic / 3D face is not kept reliably (outfit and hair are, the face drifts between scenes). Use it for anime / illustrated characters.
• A reference cropped from a comic panel carries its props over (door, mug, even the speech bubble) – crop to the character.

Prohibited use

Do not use this LoRA, alone or with other models, to:

• create intimate or sexually explicit images of an identifiable real person without their explicit consent;
• create child sexual abuse material or any sexual depiction of minors;
• build or extend facial recognition databases by untargeted scraping of facial images;
• manipulate people subliminally or deceptively in ways that cause them or others harm.

These are prohibited practices under Art. 5 of the EU AI Act (as amended in 2026). If you publish a picture that resembles a real person, place or event and could pass as authentic (a deep fake), you must disclose that it is AI-generated (Art. 50(4)).