Loop Forge

An anime character in layout boxes, on SFT

From MOSS-Image add-ons, tested locally

Model
MOSS-Image SFT (moss_image_sft_int8_convrot.safetensors)
Output
1024×1024
Sampling
25 steps, CFG 2.5, er_sde / simple
Seeds
7, 42, 101, 21, 300, 777
Seed 7
57.4 s, peak VRAM 10.4 GiB
Ran on
RTX 3060, 12GB (local)

What was wired in

  • Freya, anime character
    Picture 1
    Freya, anime characterAnime character drawn from the Freya cast photo with the Qwen-Image 2.1 edit template (not MOSS)Download PNG0.8 MB
Verdict

The board, the route number and the poster land where they should, the spelling holds and the character stays in her box. Her face is close on both models, but neither keeps it exactly the same, so I wouldn't count on it for the same character from one picture to the next.

The prompts

Download workflowDownload everything for this test (.zip, 0.8 MB)
Same character and art style as Picture 1.

Notes

The boxes say where things go and the picture says who she is. The boxes are drawn on each result. The other seeds use the same workflow with only the seed changed.

The CharRef card says it was trained on Base; the SFT run is the same graph on the SFT model.

7err4/MOSS-Image-CharRef on Hugging Face (v2)

Run it yourself

Download everything for this test (.zip, 0.8 MB). It holds the workflow, every input it loads and a README. Copy the inputs into ComfyUI/input, put the models below in place and drag the workflow onto ComfyUI.

FilePut it inSize
moss_image_sft_int8_convrot.safetensorsmodels/diffusion_models5.4 GB
moss_image_charref_v2.safetensorsmodels/loras148.7 MB
qwen3.5_9b_w4a8_gptq.safetensorsmodels/text_encoders7.3 GB
flux2-vae_small.safetensorsmodels/vae249.5 MB

No custom nodes. Every node is part of ComfyUI 0.43.1 (frontend 1.57.6, templates 0.11.100).