MOSS 4-Step Distill LoRA Cuts GPU Time 14x With Better Detail
TimeTruth2490 · reddit · 2026-11-19
Reddit user lvladikov released a self-trained 4-step distillation LoRA for MOSS-Image SFT, mixing eDMD with LADD (latent adversarial diffusion distillation), compressing the teacher's 50-step × CFG 3 pipeline (100 calls) into 4 steps at CFG 1.
Key numbers:
- 26.6s vs 387.3s for 22 images at 1024×1024 — roughly 14× less GPU time
- Fine texture detail 1.22× the teacher with near-matched tone (0.7 L lightness delta)
- Strong text rendering: 14/15 text prompts perfect, beating both the teacher and akhaliq's 4-step LoRA (12/15)
- Trained across 12 resolutions (768×512 up to 1824×1024) on 34K prompts plus 2,877 word-spelling prompts (signs, posters, labels)
- 30 hours on a single NVIDIA B300; a 15K-sample checkpoint is released with a possible update ahead
Includes 8 ComfyUI workflows (local/cloud; T2I, layout boxes, image editing) and a recommendation to stack the community MOSS Realism Enhancer LoRA for photorealistic skin.
More from Multimodal
- WaveView Injects Warped Views Into Noise Phase, Beating NVS Baselines at 2.1x Speed — KIT-Karlsruhe · 2026-11-19
- GPT-6.2 Astra demo: a Corvette escapes a MacBook, rebuilt desk in Blender for the chase — paw_lean · 2026-11-19
- HeyGen creates instant voice + avatar from one filtered photo in 5 minutes — toolstelegraph · 2026-11-19
- Multiple Google employees tease "Gemini", hinting at an imminent new model release — haider1 · 2026-11-19
- Meta's AVTR adds visual imagination routing: +4.5 pts, 33% lower latency — meta · 2026-11-19
- MOSS-Image SFT releases 4-step distilled LoRA using eDMD + LADD for high-detail T2I and I2I edits — TimeTruth2490 · 2026-11-19