MOSS-Image now runs natively in ComfyUI: which files to download and what settings to use

A day after FLUX 3 Image Klein arrived, a second small open image model landed. On October 16, 2026 OpenMOSS released MOSS-Image in two checkpoints: MOSS-Image-Base and MOSS-Image-SFT, a version fine-tuned from Base on curated data. The project lists strong text rendering, layout control, prompts of up to 4K tokens and a compact 5B backbone among its highlights. The weights on Hugging Face are tagged Apache-2.0.

The practical news for local users is that ComfyUI picked it up within hours. Here is what you need to know to run it.

What the model is

ComfyUI’s support pull request describes it as a 5B single-stream DiT in the FLUX.2 latent space with a stock Qwen3.5-9B text encoder. The DiT is the only new component. The transformer config on Hugging Face shows 42 layers with a hidden size of 3072. Both checkpoints share the same architecture and pipeline. On OpenMOSS’s own whiteboard comparison, SFT produced clearer text grouping and a more organized layout than Base.

Two ways to run it in ComfyUI

Native support. PR #17095 (“Add MOSS-image support”) was merged into ComfyUI on October 16 at 16:06 UTC. A follow-up, PR #17099, was merged on October 17. The latest tagged release, v0.41.0, was published on October 15, before either merge. So for now you need an up-to-date ComfyUI checkout rather than the last release. The native path loads the Comfy-Org/MOSS-Image repackaged files:

Folder	File	Size
models/diffusion_models/	moss_image_sft_bf16.safetensors (or _base_)	10.16 GB
models/diffusion_models/	moss_image_sft_int8_convrot.safetensors (or _base_)	5.41 GB
models/text_encoders/	qwen3.5_9b_int8_convrot.safetensors	9.75 GB
models/text_encoders/	qwen3.5_9b_w4a8_gptq.safetensors	7.34 GB
models/vae/	flux2-vae.safetensors	0.34 GB

You need one diffusion model, one text encoder and the VAE. The smallest combination, the INT8 DiT plus the W4A8 text encoder plus the VAE, is about 13.1 GB on disk. Neither OpenMOSS nor Comfy-Org publishes a VRAM figure, so treat file size as a floor, not a guarantee.

OpenMOSS’s own plugin. Search for MOSS-Image in ComfyUI Manager, or copy comfyui/moss-image/ into ComfyUI/custom_nodes/. This path reads the original Hugging Face directory (transformer/, text_encoder/, tokenizer/, vae/) through a MOSS-Image HF Loader node. It does not download from a Hub ID, so fetch the repo first. That download is larger: the bf16 text encoder alone is about 16.8 GB, plus about 10.2 GB for the transformer.

Settings that matter

The user guide says Base has had no CFG distillation, so defaults matter:

• Steps: 50 or more. CFG 3.0 by default (3.0–6.0 range, higher for short prompts). Shift 5.0, Euler sampler.
• Skip layer guidance: the recommended two-pass mode skips layer 22 on the negative branch. The plugin’s example workflow ships with 50 steps, CFG 3, Shift 5, layer-22 skipping, Euler / sgm_uniform at 1024×1024.
• Resolution: 1M, 1.5M or 2M total pixels are recommended, and dimensions must be multiples of 16.
• Negative prompt: nsfw, worst quality, bad anatomy, watermark, jpeg artifact.
• Fine-tuning: keep the text encoder frozen if you train LoRAs on Base.

Layout prompts use pixel coordinates

MOSS-Image takes a JSON layout as the prompt, with a summary and a list of items that each have a label, caption, optional text and bbox. Boxes are [y_min, x_min, y_max, x_max] in absolute pixels relative to the canvas center, so a 1024×1024 canvas runs from −512 to +512 on each axis. That differs from FLUX 3 Klein’s 0–1000 scale. If you change the output size, recompute the boxes. OpenMOSS calls layout support experimental.

Should you try it?

If you render posters, signs or bilingual text, MOSS-Image is worth testing next to Klein and Ideogram 4, especially because the Apache-2.0 tag is simpler than Klein’s revenue-capped license. For a smaller local setup, our Bonsai Image 4B guide covers the same workflow. OpenMOSS says few-step (DMD) distillation is still in progress, so expect 50-step generation times for now.