MOSS-Image 4-step distillation LoRA hits final 30K checkpoint with doubled training
TimeTruth2490 · reddit · 2026-11-20
The author released the final version of their MOSS-Image SFT 4-step distillation LoRA — checkpoint 30K, double the training of the initial 15K release, using an eDMD + LADD mix on cloud B300 GPUs. Supports T2I, T2I with layout boxes, and I2I edits, with text rendering optimized (14 of 15 text prompts correct at 1024×1024).
Improvements:
- Dark scenes now close to the 50-step teacher: L offset on the nine darkest prompts down from +2.8 to +0.6
- Doubled edges reduced from 1.12× to 1.00× of the teacher's
- Closer to teacher output overall at 1024×1024
Both LoRA files now hold chk00030000 under unchanged filenames (re-download needed); the ComfyUI file's built-in factor is now 1.12 — keep the loader at 1.0. The 15K files remain archived. Known limitations are on the model card; the author notes B300 cloud training is "mighty and very expensive."
Related event: MOSS-Image releases final 4-step distillation LoRA at 30K checkpoint(2 posts)→
More from Multimodal
- LTX 2.5 keeps crashing on AMD 9070 XT with driver timeouts while MiniMax H3 runs fine — ZestycloseReading504 · 2026-11-22
- Stripe residency dev draws all friends in JavaScript with GPT 6.2 Astra, results stun — gabrielchua · 2026-11-22
- Fable 5.2 plus WebGL demoed as a tool to "create anything" — hewarsaber · 2026-11-22
- Cinematic UFO folklore music video built with Suno and SLOPCANNON — DarkroomSignal · 2026-11-22
- Photographer crafts AI-generated digital muse Nous Girl on faux magazine cover — alexcovo_eth · 2026-11-22
- Turn any Google Maps location into a cinematic drone shot with AI — aziz4ai · 2026-11-22