MOSS-Image-Base

Use it to generate images from prompts with fewer sampling steps than the base model’s 50-step configuration.

Availability

Checkpoint: update 12000 (EMA), float32 weights

Pricing model

Not established

Deployment

Load locally with the base MOSS-Image-Base transformer

Privacy

Not established

Licence

Apache License, Version 2.0

• It supports natural-language prompts, tag sequences, and JSON layouts. ^[1]

  It supports natural-language prompts, tag sequences, and JSON

• The quick start requires an NVIDIA GPU with CUDA support. ^[1]

  Use Python 3.10 or later and an NVIDIA GPU with CUDA support.

• The model weights and bundled inference code are released under the Apache License, Version 2.0. ^[1]

  The model weights and bundled inference code are released under the

• Code and model weights for MOSS-Image-Base are publicly available. ^[2]

  Code and model weights are publicly available.

• MOSS-Image supports layout control for text and visual elements. ^{[1] [2]}

  - 📐 **Layout control:** Control the placement of text and visual elements.

• MOSS-Image accepts text prompts of up to 4K tokens.
• MOSS-Image generates anime and photorealistic images.
• The MOSS-Image series is available through the MOSI API Platform.
• The distillation targets 8 network evaluations from a 50-step configuration. ^[3]
• The recommended setup uses the LoRA with the stock MOSS-Image-Base transformer. ^[3]
• Recommended sampling uses CFG 1 and 4–8 steps. ^[3]
• The checkpoint is update 12000 (EMA) with float32 weights. ^[3]

  Checkpoint: update 12000 (EMA), float32 weights.

• The transformer must be loaded and run in fp32 for the data-space sampler. ^[3]

  Precision: the distillation delta is smaller than bf16's rounding step, so `transformer/` must be loaded AND run