MMOSS Image AI

MOSS Image generator

A lightweight form for the MOSS-Image model. It talks directly to OpenMOSS's official Hugging Face Space, so results match the official demo. The Generate button on this page is the way to make an image. Settings below explain the values OpenMOSS recommends.

MOSS-Image generator (simple form)

Sends your prompt straight from your browser to the official Space through Gradio's public API. We never see your prompt or image. Free ZeroGPU quota and queue rules are Hugging Face's, not ours.

Advanced settings (CFG, seed, negative prompt)

Beginners can leave these. CFG stays at the official default of 3. Free ZeroGPU time is shared, so a queue is normal. Pro pay-per-run is not available yet.

Negative prompt

Default taken from the official user guide.

Open official Space ↗

Anime woman waiting at a rainy Tokyo crosswalk. OpenMOSS published MOSS-Image output, assets/images/007
Crosswalk scene (SFT-style published output, assets/images/007). Source: OpenMOSS assets/images/007. © OpenMOSS Team, Apache-2.0. We resized the long edge to at most 1400px and saved WebP. Not made on this site.
Black and white film still of a detective by a window. OpenMOSS published MOSS-Image output, assets/images/015
Noir interior (published output, assets/images/015). Source: OpenMOSS assets/images/015. © OpenMOSS Team, Apache-2.0. We resized the long edge to at most 1400px and saved WebP. Not made on this site.
Diagram of the settings the generator form sends to the Space
Diagram of the settings this form sends. Not a generated image.

How the free generator works

  1. Your browser connects to the public Space OpenMOSS-Team/MOSS-Image using Gradio's JavaScript client. This site's server is not involved.
  2. Hugging Face queues the request and runs it on ZeroGPU, shared GPUs lent to the Space for a few seconds at a time.
  3. GPU time counts against your daily ZeroGPU quota: about 2 minutes if you are not signed in, 5 minutes for a free Hugging Face account, and 40 minutes for PRO. A 50-step image at 1024×1024 uses a meaningful share of the anonymous allowance.
  4. When the quota runs out the form says so. Signing in on the official Space page gives a larger allowance. The quota resets 24 hours after the first use.

Settings explained

SettingOfficial recommendationWhat changing it does
ModelBase or SFTSFT is fine-tuned and tends to organise text more cleanly. Base is the starting point OpenMOSS describes for LoRA training.
Resolution1M, 1.5M or 2M pixelsWidth and height must be multiples of 16. Dense text needs more pixels than a 512×512 draft.
Steps50 for best resultsFewer steps are faster and use less quota, and they soften fine text.
CFG3.0 (range 3–6)Hidden under Advanced. 3 is usually enough for long prompts.
Shift5.0A scheduler value. The form keeps it at 5.
Layer skippingNegative-branch skip at layer 22The form keeps the official default.
Negative promptnsfw, worst quality, bad anatomy, watermark, jpeg artifactThe tags recommended in the user guide.
Seed—Fix a seed to compare prompt edits. The same seed on different hardware may still differ slightly.

Prompt tips

  • Quote every piece of visible text exactly, and put real line breaks where lines should break.
  • English descriptions work best. The text to render can be Chinese, English, or mixed.
  • Describe hierarchy: which line is the large title, which is small print, and where each block sits.
  • Need exact positions? Use the layout JSON builder. This form accepts its output and the matching size.
  • Ready-made examples are in the prompt library.

When the form does not work

  • Quota used up — the official Space link under the form lets you sign in to Hugging Face for a larger allowance.
  • Network error — some networks block huggingface.co or *.hf.space.
  • Space restarting — wait a few minutes after an OpenMOSS update.

Sources

Facts checked 2026-10-18.