Sign In
Civitai turns 4. Pick up a hat.
homemodelsimagesvideos3D Modelshubsarticlescomicscruciblesupdatesshop

MOSS-Image Edit CharRef (early experiment)

0

Updated: Oct 20, 2026

conceptanimemoss-imageimage editingconsistent characterreference

Recently loaded

Download

1 variant available

bf16 SafeTensor

moss_image_charref_poc.safetensors

BF16, good balance • 138.43 MB

Verified:

Type
LoRA
Stats

267

0

Reviews

(23)

Published

Oct 20, 2026

Base Model

MOSS Image

Training
Steps: 2,000
Usage Tips
Strength: 1
Hash
AutoV271E65BA732
00a_flyer.png

Description

An experimental edit LoRA for MOSS-Image Base: Picture 1 is an anime character, the prompt describes a new scene and pose, and the model tries to draw the same character there.

It is not tied to any particular character: any anime character can be the reference. Bagel the fox and Kujou, in the gallery, are only examples.

⚠️ Very preliminary. A first experiment: 2,000 steps on about 830 image pairs from 16 characters, tested on only a handful of characters. Expect rough results: it usually picks up the broad look (hair colour, ears or horns, main outfit colours) but often gets face shape, exact hairstyle and outfit details wrong; the art style usually does not carry over; results vary a lot between seeds and characters. Anime characters only. Shared so others can try the idea; a properly trained version is planned.

Without this LoRA, MOSS-Image just repaints the reference and ignores the prompt (see the comparison image).

What you need (ComfyUI)

  • ComfyUI v0.42.0 or newer

  • Custom node ComfyUI-MOSS-Image-Ostris-Edit by ostris: github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit

  • From huggingface.co/Comfy-Org/MOSS-Image:

  • moss_image_base_int8_convrot.safetensors → models/diffusion_models (Base, not SFT)

  • qwen3.5_9b_int8_convrot.safetensors → models/text_encoders

  • flux2-vae.safetensors → models/vae

  • This LoRA → models/loras

  • The workflow: the Archive file on this page (zip), or drag the 00b gallery image into ComfyUI

How to use

  1. Reference: one character, large and clear (full body or upper body). A tiny character in a wide scene does not work.

  2. Prompt, three parts: Same character and art style as Picture 1., comma-separated tags for the scene, pose, framing and expression, and the character's own appearance tags (hair, eyes, outfit), for example from a WD14 tagger on the reference. The character tags keep the character much closer (see the comparison image).

  3. Settings: 30 steps, CFG 3, euler / simple, edit patch kv_cache on, skip layer 22, output about 1 MP, LoRA 1.0 (0.8–1.2).

  4. Leave the negative prompt empty.

Tip: upper body or cowboy shot framing keeps the face closer to the reference.

Limitations

Anime characters only, not for real people. One character and one reference. Fine identity details are often wrong. The reference's style usually comes out closer to MOSS's default anime look. Early experiment with very little testing.

Credits

OpenMOSS (MOSS-Image, Apache-2.0), ostris (AI Toolkit edit training and the ComfyUI node), Comfy-Org (model files).