Original title: MOSS-Image-IDM-2-Steps
Based on MOSS-Image-SFT, progressively distilled through lodestones 8 steps and darask 0 4 steps, then trained into a 2-step, CFG-free LoRA using IDM (Intrinsic Distribution Matching) loss. The authors claim that on held-out dense text prompts, some prompts yield better results with 2 steps than with the 50-step teacher model; there is also an experimental editing version v2.
Why it matters MOSS can generate posters and menu-style images with text in just 2 steps; this is the community's first open implementation trained with IDM.
Hugging Face likes 42 · Hugging Face upvotes 42 · Hugging Face downloads 0
License: Commercial use allowed · ComfyUI: Mentioned in model card · VRAM: Not specified
Reasons
- The author explicitly states that the v2 editing version degrades general text-to-image performance; use v1 for text-to-image
- Includes an online Demo Space and grouped sample images
Usage
- For text-to-image, use v1 under runs/idm/export/000750/; for local attribute editing, use v2 from the repository root or adapter_v2/, with 2 steps and CFG disabled
Risks
- Community implementation; evaluation is based on the author's own small-scale comparisons; the two checkpoints serve different purposes, and using the wrong one will degrade performance