MOSS-Image Base: 5B Text-to-Image Checkpoint for Fine-Tuning
MOSS-Image Base is the open 5B pre-training checkpoint of MOSS-Image, kept for LoRA and fine-tuning work with layout JSON control and ComfyUI support.
MOSS-Image Base
Text-to-ImageLayout ControlFine-TuningOpen WeightsThe pre-training checkpoint of MOSS-Image, published without post-training or distillation of inference parameters so the full training signal stays in the weights. OpenMOSS recommends it as the starting point for LoRA training and for anyone who wants to build on the model rather than run it as-is.
| Developer | OpenMOSS |
| Release Date | 2026-10-16 |
| Architecture | 5B single-stream DiT (42 layers) in the FLUX.2 latent space, Qwen3.5-9B text encoder, FLUX.2 VAE |
| Parameters | 5B |
| License | Apache-2.0 |
| Sampling | 50 steps or more, CFG 3 (range 3 to 6), shift 5, Euler, negative-branch skipping at layer 22 |
Capabilities
- Text rendering: legible text in English and Chinese, with automatic line breaks
- Layout JSON: place text boxes and objects by bounding box instead of describing positions in prose
- Prompt formats: natural language, tag sequences or layout JSON, with prompts up to 4K tokens
- Resolution range: 0.25M (512×512) to 4M (2K) pixels, tuned for 1M, 1.5M and 2M outputs
- LoRA ready: the checkpoint keeps the full training information, which is why it is the one OpenMOSS suggests for training
Notes
Because nothing was distilled, the base checkpoint needs the settings from the user guide to perform: 50 steps or more, CFG 3, shift 5 and negative-branch skipping at layer 22. Turning layer skipping off makes output more realistic but can collapse detail. Negatives the guide suggests are nsfw, worst quality, bad anatomy, watermark, jpeg artifact.
Related
Guides and workflows related to this model series.
Comments
Sign in with GitHub to join the discussion.