MOSS–ImageJoin the waitlist

Every claim,
with its source.

This is an independent community site. Below, each factual statement we make is tied to a primary source, with an honest note on how we checked it. Checked October 18, 2026; sources change, so please read the originals.

Claims, status and primary sources
AreaStatement on this siteHow we checkedSources
ModelBase and SFT weights and code released on 2026-10-16Read in primary source
Model~5B-parameter DiT; Apache License 2.0; text-to-image pipeline tagRead in primary source
ModelUp to 4K-token text promptsRead in primary source
ModelJSON layout input: summary/items, bbox_format yxyx, bbox_coord_space abs_centralized (pixel, canvas-centred; ±512 on a 1024×1024 canvas); layout prompting labelled experimentalRead in primary source
ModelEnglish recommended (most training data); Chinese and other languages supported; explicit line breaks can improve text controlRead in primary source
ModelSFT is fine-tuned from Base on curated data; one published example shows clearer text groupingRead in primary source
ComfyUINative custom-node plugin; copy comfyui/moss-image into custom_nodes; MOSS-Image HF Loader needs a local model directory; 4096-token text-encoding capRead in primary source
ComfyUIREADME says integration into the official ComfyUI repository is being worked on. We do not claim built-in ComfyUI support.Read in primary source
ComfyUIComfy-Org publishes repackaged files (diffusion model, Qwen 3.5 9B text-encoder and FLUX.2 VAE variants) on Hugging Face. The text encoder and VAE are separate components: check their own licences.Read in primary source
HistoryUpgraded MOSS open-sourced 2023-04-21 by Fudan University’s NLP lab; plugins include search, image generation, calculator and equation solver (report dated 2023-05-08)Read in primary source
HistoryOriginal MOSS: 16B-parameter plug-in-augmented conversational model; code Apache-2.0, data CC BY-NC 4.0, model weights GNU AGPL-3.0; Shanghai AI Lab thanked for GPU supportRead in primary source
HistoryOpenMOSS describes itself as a research group under the Shanghai Innovation Institute, working with Fudan University and MOSI Intelligence; Prof. Xipeng Qiu leads; MOSS-Image is a pinned repositoryRead in primary source
HistoryThe MOSS-Image README links @Open_MOSS on X as the project accountRead in primary source
ComparisonMidjourney: style references, moodboards, personalization, editor (remix, inpaint, pan, zoom out), Edit Model with up to four reference images replacing Omni/Character Reference in V8.x, upscalers, video, ExploreConfirmed via search summary
ComparisonChatGPT: create images from prompts, edit uploaded or generated images by describing changes, transparent backgrounds, aspect-ratio pickerConfirmed via search summary
Not verifiedTechnical report PDF (linked in the README): linked here but not read in full; no benchmark or accuracy figure from it is quoted on this siteNot verified — not claimed
Not verifiedHugging Face trending rank, day-one ComfyUI mention, and specific posts by @MosiAI_Official: not confirmed, so not stated. The account is linked in the footer because it was part of the original research brief; the official repository links @Open_MOSS instead.Not verified — not claimed
Not verifiedImage quality, text accuracy, speed or VRAM needs: we have not run the model. The gallery is hand-authored SVG, not model output.Not verified — not claimed—

On “search summary”: Midjourney and OpenAI help pages refused automated fetches (HTTP 403), so we confirmed those features from search-engine summaries of the pages rather than reading the pages ourselves. Treat that comparison as indicative and check the linked pages. Spotted an error? Email createsun@foxmail.com and we will correct it.