yijunwang2 · Released, not on our provider
MOSS AnyPaint (2026): what it does, what you download, and how to run it
MOSS AnyPaint is an unofficial community adapter, created on Hugging Face in November 2026 under Apache-2.0, that adds arbitrary-mask inpainting, outpainting and image editing to the MOSS-Image-Base image model, runs in 8 steps on a CUDA GPU, and cannot be run on this site, which only chats with text models.
Use it now
Chat
This chat starts with Inkling 2 Small, not MOSS AnyPaint.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- yijunwang2
- Type
- Image
- Released
- No date given
- OpenRouter id
- None
Checked
MOSS AnyPaint: facts, with sources
- The card gives no release date. Hugging Face’s data for the repository yijunwang2/moss-anypaint shows it was created on November 13, 2026 (UTC) and last changed on November 14, 2026. The maker’s ComfyUI node repository was created on November 13 and its latest release, v0.2.0, is dated November 14, 2026. These are repository dates, not an announced launch.Source: huggingface.co · checked 2026-11-19
- The card says MOSS AnyPaint provides arbitrary-mask inpainting, outpainting and image editing for MOSS-Image-Base, the OpenMOSS image model. One runtime handles edits inside the picture, extending the canvas, several separate regions, narrow shapes, and masks that cross the edge of the source. It is a small add-on (a rank-32 LoRA adapter), not a new base model. The card calls it an unofficial community release that is not endorsed by the authors of the parts it builds on.Source: huggingface.co · checked 2026-11-19
- Input as the card describes it: one RGB source image, one binary mask (the size of the source or of the output canvas), an output canvas whose sides are multiples of 16, an optional box saying where the source sits on the canvas, and a prompt that describes the complete picture you want afterwards. White mask pixels are generated or edited; black pixels are kept. Everything outside the placed source counts as generated, so one input form covers both inpainting and outpainting.Source: huggingface.co · checked 2026-11-19
- Recommended settings on the card: 8 steps, CFG 1, shift 5, LoRA scale 1.0 for both adapters, frozen transformer and text weights in Quanto qfloat8, BF16 for the compute and task adapter, FP32 for the VAE and the distillation adapter, a reference image of at most 384 px on its longest side, and a 32-pixel redraw band at the outer seams of the source (0 px for masks inside the picture). The card says the unmasked pixels are restored after every step and that no hidden composite with the source is added at the end.Source: huggingface.co · checked 2026-11-19
- To run it you need Python 3.11 or newer and a CUDA GPU with BF16 support. The card gives the steps: download the repository with the Hugging Face CLI, install requirements.txt, then run example.py with a source image, a mask, a size, an optional box and a prompt. The first run downloads the base components and the 8-step distillation adapter. The card gives no GPU memory figure, and says the auto-generated Diffusers snippet on Hugging Face and ordinary LoRA importers do not work with this adapter.Source: huggingface.co · checked 2026-11-19
- The maker’s GitHub README for the ComfyUI nodes (MIT licence on GitHub) lists what to put in ComfyUI: a MOSS image base file (int8), a Qwen 3.5 9B text encoder file (int8) and a small FLUX.2 VAE file from the Comfy-Org/MOSS-Image repository, the AnyPaint LoRA, and the 8-step DMD2 LoRA loaded through a second ordinary Load LoRA node. The nodes are called Prepare, Encode and Model Patch, and then the standard KSampler runs with 8 steps, Euler, the simple scheduler and CFG 1.0.Source: github.com · checked 2026-11-19
- Licence: the card says the LoRA weights and the included runtime are released under Apache License 2.0, and the base model is also Apache-2.0. The maker’s notices say the 8-step distillation adapter, downloaded separately, is declared Apache-2.0 on its own model card (no separate licence file was present), and that the DiffSynth-Studio runtime is Apache-2.0. Training data is not included.Source: huggingface.co · checked 2026-11-19
- Responsible use, as the card states it: use reference images only with the subject’s consent or another lawful basis; do not use the model for deceptive imitation of real people, fraud, harassment, non-consensual intimate imagery, biometric identification, or misleading political or media content. Deployments should follow the upstream terms, apply suitable content moderation and disclose AI-generated content where the law requires it.Source: huggingface.co · checked 2026-11-19
- The card lists its own limits: a large generated share of the picture may add unrelated subjects or structures; very thin mask features can be coarsened; ambiguous prompts can change more of the surroundings near the redraw band; the mask must cover the whole object or edge to be replaced; small-object counts, exact positions and geometry vary by seed; unmasked pixels can still differ slightly after VAE decoding; and the model may repeat biases from the MOSS base model.Source: huggingface.co · checked 2026-11-19
- The card gives no benchmark scores or speed figures. It shows twelve cases, each with its source image, exact mask, result and prompt in the repository. The card says the sources are synthetic images made with a different base model and contain no real person, brand or existing character. We show none of these pictures on this page.Source: huggingface.co · checked 2026-11-19
- The weights can be downloaded without a fee from Hugging Face. We found no hosted price for MOSS AnyPaint on the pages we read, and OpenRouter has no price because it does not list it.Source: huggingface.co · checked 2026-11-19
- MOSS AnyPaint is not on OpenRouter’s model list (checked November 19, 2026). It is an image-editing adapter, and this site is a text chat, so this site cannot run it. A text model here can help you write the description of the picture you want, but it does not edit images.Source: openrouter.ai · checked 2026-11-19 · second-hand
Our table
The four parts you download to run MOSS AnyPaint
- The adapter file moss_anypaint_rank32.safetensors is 148,693,368 bytes. The maker’s runtime lock file records rank 32 and alpha 32 for it, and pins the base model, the distillation adapter and the DiffSynth-Studio runtime to exact revisions. The base MOSS-Image-Base repository holds 5,082,308,624 parameters by Hugging Face’s count.Source: huggingface.co · checked 2026-11-19
Parts and licences
MOSS AnyPaint parts: role, size and licence as stated by the maker
MOSS AnyPaint is only one piece. The card says it works together with the MOSS-Image-Base model, an 8-step distillation adapter and the DiffSynth-Studio code, and the first run downloads the other pieces. The table lists what each piece is for. This site cannot run image models; for small open-weight text models you can chat with, see Inkling 2 Small, and for other small models from Hugging Face we have read, see Aion 1.0 and SenseNova Loop 9B.
| Part | Role | Size | Licence stated |
|---|---|---|---|
| MOSS AnyPaint adapter | The task adapter: arbitrary-mask inpainting, outpainting and editing | 148,693,368 bytes, rank 32 | Apache-2.0 |
| MOSS-Image-Base | The base text-to-image model from OpenMOSS, downloaded separately | 5,082,308,624 parameters (Hugging Face’s count) | Apache-2.0 |
| Base 8-step DMD2 adapter | Distillation adapter by another author, kept unmerged, in FP32 | 1,189,176,800 bytes (lora_dmd2.safetensors) | Apache-2.0 declared on its model card |
| DiffSynth-Studio | Runtime code the model forward is derived from, pinned to one revision | Not stated on the pages we read | Apache-2.0 |
Source: huggingface.coSource: huggingface.coSource: huggingface.coSource: huggingface.co
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to MOSS AnyPaint you can use now
Inkling 2 Small
Thinking Machines Lab$0.30 / $1 (per million tokens)
- Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
- A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
- Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14
Qwen3.8 Flash
Alibaba$0.15 / $0.47 (per million tokens)
- Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13
FAQ
MOSS AnyPaint FAQ
What is MOSS AnyPaint?
MOSS AnyPaint is a small add-on (a LoRA adapter) for MOSS-Image-Base, the OpenMOSS image model. It lets you edit only the part of a picture that you mark with a mask, extend a picture past its edges, or do both in one request. The Hugging Face repository was created on November 13, 2026; the card gives no launch date. It is a community release by the Hugging Face user yijunwang2, not an OpenMOSS product.
What do I need to run MOSS AnyPaint?
The card asks for Python 3.11 or newer and a CUDA GPU with BF16 support. You download the adapter (148,693,368 bytes), and the first run also downloads the MOSS-Image-Base parts and the 8-step distillation adapter. The card gives no GPU memory figure, so we cannot say which graphics cards are enough. It will not work through the auto-generated Diffusers snippet or an ordinary LoRA loader; use the included example.py or the ComfyUI nodes.
How do I give MOSS AnyPaint a mask?
Use a black and white image. White pixels are repainted and black pixels are kept. The mask can be the size of the source picture or of the final canvas, and any area outside the placed source is repainted automatically, which is how outpainting works. In the prompt, describe the whole picture you want afterwards, not only the changed part. The card warns that the mask must cover the whole object or edge you want replaced.
How do I use MOSS AnyPaint in ComfyUI?
The maker’s GitHub page for the ComfyUI nodes says to clone the node repository into custom_nodes, put the AnyPaint LoRA in the loras folder, and load the MOSS base files, the text encoder, the VAE and the 8-step DMD2 LoRA with the ordinary loaders. Then connect Prepare, Encode and Model Patch before the standard KSampler (8 steps, Euler, CFG 1.0). Padding stays at 0 for inpainting; to extend the picture, set the side you want in pixels.
Can I use MOSS AnyPaint commercially?
The card says the LoRA weights and runtime are Apache-2.0, and the base model and the other parts are listed as Apache-2.0 too, which is a permissive licence. The card also says deployments must follow the upstream terms and that you should add content moderation and disclose AI-generated content where required. We are not lawyers; read the licence files in the repository before building a product on it.
What are the limits of MOSS AnyPaint?
By the card: a large repainted share may bring in unrelated objects, very thin mask features can come out coarser, vague prompts can change more of the surroundings, and the number and exact position of small objects vary with the random seed. Unmasked pixels can also differ slightly after decoding. The card gives no benchmark numbers, so there is no score to compare with other editing models.
Can I try MOSS AnyPaint on this site?
No. It is an image-editing adapter and it is not on OpenRouter’s model list, so this site cannot run it. This site is a text chat. You can use a small text model to draft the picture description for your prompt: on this site a typical reply to Inkling 2 Small costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit.
Is this the official MOSS AnyPaint website?
No. This page is unofficial and not affiliated with or endorsed by yijunwang2 or OpenMOSS. Facts link to the pages we read.