ComfyUI MOSS-Image T8

ComfyUI MOSS-Image T8

English | 简体中文

Run MOSS-Image Base and SFT in ComfyUI with the pinned, unmodified official inference pipeline. DiT (UNET), text encoder and VAE have separate files and loaders. Base and SFT share the full Qwen3.5 text encoder and FLUX.2 VAE.

Models · Comfy Registry · Validation · Upstream

Install

In ComfyUI Manager, search for ComfyUI MOSS-Image T8 (publisher t8star, package moss-image-t8), install and restart ComfyUI.

For manual installation, run from your ComfyUI directory using the Python that runs ComfyUI:

git clone https://github.com/T8mars/Comfyui-Moss-Image-T8 custom_nodes/Comfyui-Moss-Image-T8
python -m pip install -r custom_nodes/Comfyui-Moss-Image-T8/requirements.txt

Requires Python 3.10+, PyTorch 2.9+, Diffusers 0.38–0.41 and Transformers 5.3+. Restart after installing dependencies.

Models

Download from t8star/Moss-Image-Comfy into ComfyUI/models, preserving the directory structure. This command downloads both variants and their shared components:

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

hf is installed by huggingface_hub; use its executable in your ComfyUI Python environment. To download SFT only, replace "diffusion_models/*" with "diffusion_models/moss-image-sft-bf16.safetensors*".

ComfyUI/models/
├── diffusion_models/
│   ├── moss-image-base-bf16.safetensors
│   ├── moss-image-base-bf16.safetensors.assets/manifest.json
│   ├── moss-image-sft-bf16.safetensors
│   └── moss-image-sft-bf16.safetensors.assets/manifest.json
├── text_encoders/
│   ├── moss-image-qwen35-bf16.safetensors
│   └── moss-image-qwen35-bf16.safetensors.assets/
│       ├── manifest.json
│       └── tokenizer/{tokenizer.json,tokenizer_config.json,chat_template.jinja}
└── vae/
    ├── moss-image-vae.safetensors
    └── moss-image-vae.safetensors.assets/manifest.json

Keep each .assets directory beside its weight file. Each DiT is 10.16 GB, the shared TE is 16.79 GB and VAE is 0.25 GB; both variants together need about 37.4 GB. Original tensor data and precision are preserved. To export directly from the pinned upstream repositories, run python scripts/prepare_models.py --models-dir <ComfyUI/models> --variant both from a source checkout (about 75 GB including source cache).

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

Workflows

Drag a JSON or PNG below onto the ComfyUI canvas. PNGs are actual generated images with the complete workflow embedded. The three model loaders are preselected for the filenames above.

| Model | Canvas workflow | Image with embedded workflow | |---|---|---| | SFT | JSON | PNG | | Base | JSON | PNG |

Six dedicated nodes cover DiT, TE and VAE loading, text encoding, official sampling and VAE decoding. Decoded output is standard ComfyUI IMAGE; packed latents use MOSS_IMAGE_LATENT and connect to this package's decoder. API workflows are included as *-api.json.

Defaults: 1024×1024 · 50 steps · CFG 3 · shift 5 · seed 42 · empty negative prompt. The unconditional branch skips layer 22 (zero-based); extra SLG and VAE tiling are off. TE/DiT use BF16; VAE decoding uses FP32. Width and height must be positive multiples of 16.

Models are released between encoding, sampling and decoding. auto follows ComfyUI's device policy and reclaims its cached models before checking available memory. The full TE can fall back to CPU and run more slowly. Tested on Windows with a 24 GB RTX 5090 Laptop GPU; this is a tested configuration, not a minimum VRAM claim.

Both workflows passed real canvas Run-button generation at 1024×1024 / 50 steps and all eight JSON/PNG import checks. All 1,936 tensors match the official source shards byte for byte. Tests cover pipeline parity, sampling, file integrity, device selection, cancellation, verification tools and the Windows launcher. See validation, latest review and source notes.

Credits and license

Models and inference pipeline by OpenMOSS: Base and SFT. This community integration and model repackaging are maintained by T8star. Apache-2.0; upstream revisions and attribution are recorded in NOTICE.

Links

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