t8star/Moss-Image-Comfy

MOSS-Image for ComfyUI

English | 简体中文

Independent DiT (UNET), text encoder and VAE files for ComfyUI MOSS-Image T8. Supports MOSS-Image Base and SFT with the original OpenMOSS inference pipeline. This is a community repackaging: tensor names, shapes, dtypes and raw data are preserved; configuration is embedded in safetensors metadata.

Files

Component	File	Size
Base DiT, BF16	diffusion_models/moss-image-base-bf16.safetensors	10.16 GB
SFT DiT, BF16	diffusion_models/moss-image-sft-bf16.safetensors	10.16 GB
Shared full Qwen3.5 TE, BF16	text_encoders/moss-image-qwen35-bf16.safetensors	16.79 GB
Shared FLUX.2 VAE, FP32	vae/moss-image-vae.safetensors	0.25 GB

Download each selected weight together with its matching .assets directory. These contain integrity manifests and the TE tokenizer. Base and SFT use the same TE and VAE; both variants together total about 37.4 GB. Components are never merged into one checkpoint.

Keep tokenizer/ limited to its three shipped files. Extra files and subdirectories are rejected by the loaders.

Quick start

1. Install ComfyUI MOSS-Image T8 through ComfyUI Manager (publisher t8star, package moss-image-t8) or follow the manual installation instructions.
2. From the ComfyUI directory, download models into models/ with the directory structure preserved:

hf download t8star/Moss-Image-Comfy --local-dir models --include "diffusion_models/*" --include "text_encoders/*" --include "vae/*"

3. Restart ComfyUI and drag a SFT workflow PNG or Base workflow PNG onto the canvas. JSON workflows are also available.

For SFT only, replace "diffusion_models/*" with "diffusion_models/moss-image-sft-bf16.safetensors*".

Defaults: 1024×1024, 50 steps, CFG 3, shift 5, seed 42, empty negative prompt, unconditional skip layer 22 (zero-based). TE/DiT use BF16 and VAE decoding uses FP32. Both variants passed actual ComfyUI canvas generation at these settings on a 24 GB RTX 5090 Laptop GPU. See validation.

Provenance and integrity

• OpenMOSS-Team/MOSS-Image-Base, revision 5f40f07a304541840b6019d2d34d30f966756787.
• OpenMOSS-Team/MOSS-Image-SFT, revision daf706d2e1c7a6369c17b6d1658671aee6db1ebe.

All 1,936 tensors were compared byte for byte with the source shards, including the complete TE vision weights. Configurations and tokenizer files were checked against the pinned upstream content IDs. Each .assets/manifest.json records the full weight SHA256 and provenance; SHA256SUMS covers all distributed weights and assets.

Models and pipeline by OpenMOSS; ComfyUI integration and repackaging by T8star. Apache-2.0; see NOTICE. This repository contains component files for the linked integration; use its dedicated loaders and official-pipeline sampler.

Links

Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI bundle · Hugging Face