dennygod/moss-image-workflow

MOSS-Image workflow models

This repository contains the model files needed by one ComfyUI workflow: MOSS-Image SFT (int8) with the 4-step acceleration LoRA plus two content LoRAs, sampled at 768×768.

The folder layout mirrors a ComfyUI models/ directory, so a download lands in the right places without rearranging anything.

Files

Path	Size	Role
models/diffusion_models/Moss/moss_image_sft_int8_convrot.safetensors	5.04 GiB	Diffusion model (UNETLoader)
models/text_encoders/qwen3.5_9b_moss_image_sft_w4a8_gptq.safetensors	5.06 GiB	Text encoder (CLIPLoader, type moss_image)
models/vae/flux2-vae.safetensors	0.31 GiB	VAE
models/loras/moss/moss_image_sft_4_steps_lora_v2_600_avg_rank_72_bf16.safetensors	0.30 GiB	4-step LoRA (strength 1.0)
models/loras/moss/PenisLoraMossNoTriggerwords.safetensors	0.08 GiB	Content LoRA (strength 0.2)

All files are unmodified copies; see SHA256SUMS.txt for checksums and SOURCES.md for origin and licensing.

Restore

Download into the ComfyUI installation root, so the repository's models/ prefix becomes <ComfyUI>/models/:

hf download dennygod/moss-image-workflow --repo-type model --local-dir "<ComfyUI root>"

If ComfyUI reads models from another directory through extra_model_paths.yaml, download anywhere and point base_path at the downloaded models folder:

comfyui:
    base_path: <path to downloaded models folder>
    diffusion_models: diffusion_models/
    text_encoders: text_encoders/
    loras: loras/
    vae: vae/

Workflow requirements

• ComfyUI build that supports the Flux2 latent (EmptyFlux2LatentImage) and the moss_image text-encoder type in CLIPLoader.
• rgthree-comfy custom nodes for Lora Loader Stack (rgthree); StringReplace ships with core.
• Sampling used with this workflow: 4 steps, CFG 1.5, euler / simple, 768×768, denoise 1.0.

Notes

• The repository holds no prompts, seeds, or any personal content — model files only.
• The two LoRAs are optional. Loading only the 4-step LoRA works but changes the output style.