MOSS-Image is a 5B-parameter DiT text-to-image model from the OpenMOSS team, released under Apache 2.0 in two checkpoints — MOSS-Image-Base and MOSS-Image-SFT. Both landed on Hugging Face on 11 October. The GitHub repository went public on 16 October, and the ComfyUI integration is not a third-party wrapper: it lives in comfyui/ inside the official repo.
What distinguishes it is the input side. The model card describes three accepted prompt forms: natural-language prompts, tag sequences, and JSON layouts that control text placement and object arrangement. The repository description is blunter about the priority — “high-quality text rendering, layout generation, and efficient fine-tuning.”
Why layouts matter more than they sound
Most open image models give you one lever: a string. If you want a sign in the upper left reading something specific, you write that in prose and hope. Bounding-box conditioning turns that into an instruction rather than a suggestion, which is the difference between a model you can prompt and a model you can lay out.
The showcase assets in the repository include a layout-bboxes example, and the base-versus-SFT comparison the team published is specifically a whiteboard scene — the SFT checkpoint produces “clearer text grouping and a more organized whiteboard layout.” That is a narrow claim about a narrow thing, and it is the right kind of claim to make: text and arrangement are where open image models have been weakest and where a 5B model has a real shot at being competitive with something far larger.
The quick-start prompt in the model card is a refrigerator with a handwritten sticky note in Chinese. Both checkpoints are tagged en and zh, and bilingual text rendering is a harder problem than English alone.
The week the community got to it first
The interesting part of this release is not the release. It is what happened to it in five days.
| Repo | What it is | Live |
|---|---|---|
OpenMOSS-Team/MOSS-Image-Base | Base checkpoint, Apache 2.0 | 11 Oct |
OpenMOSS-Team/MOSS-Image-SFT | Fine-tuned checkpoint | 11 Oct |
Comfy-Org/MOSS-Image | Single-file repackaging for ComfyUI | 16 Oct |
lodestones/MOSS-Image-SFT-8-steps-dmd | Community 8-step DMD distillation | 16 Oct |
The Comfy-Org repackaging passed 90,000 downloads within a day of appearing. The distillation — a DMD step-reduction of the SFT checkpoint by an independent uploader — is at a little over 12,000. Neither of those is the team’s own work, and both are the reason the model is now reachable from a stock ComfyUI install on modest hardware.
That pattern is becoming the standard shape of an open image release: the lab ships weights, Comfy-Org ships the single-file version, and somebody distils it before the week is out.
Running it
The official workflow is deliberately unoptimised, and the README says what it is tuned for:
The workflow is ready to run with 50 steps, CFG 3, Shift 5, and negative-branch skipping at layer 22, using Euler / sgm_uniform at 1024×1024.
Fifty steps is a lot. If that is the reference configuration, the eight-step distillation is doing most of the work of making this usable interactively, and the GGUF and w4a8 quantisations circulating alongside it are doing the rest.
Installation is a directory copy — comfyui/moss-image/ into ComfyUI/custom_nodes/ — or python comfyui/install.py --comfy-root /path/to/ComfyUI from the repo root. The nodes output standard MODEL, CLIP and VAE, so the model slots into native sampling and ComfyUI’s own memory management rather than bringing its own.
For diffusers users, the pipeline is custom_pipeline="pipeline_moss_image" with trust_remote_code=True, and it wants diffusers>=0.38, transformers>=5.3 and PyTorch 2.9 or later.
What is not settled
The team published a technical report as a PDF in the repo rather than on arXiv, so there is no peer commentary and no comparison table anybody outside the project has reproduced. The base-versus-SFT comparison is a single image pair. There is a Spaces demo, which is the fastest way to find out whether the text rendering holds up on your own strings before you download 5B parameters twice.
Apache 2.0 on both checkpoints is the part that will matter in six months. It is the licence that lets the repackaging and the distillation above exist without anyone asking permission, which is exactly what happened.
Sources
- OpenMOSS-Team/MOSS-Image-Base — Hugging Face
- OpenMOSS-Team/MOSS-Image-SFT — Hugging Face
- OpenMOSS/MOSS-Image — GitHub
- MOSS-Image for ComfyUI — official plugin README
- MOSS-Image project homepage
- Comfy-Org/MOSS-Image — single-file repackaging