Krea 2 Turbo 2-step distillation LoRA hits 4x faster denoising with new checkpoint
TimeTruth2490 · reddit · 2026-10-03
The author released a new checkpoint (chk00041320) of the Krea 2 Turbo 2-step distillation LoRA, still in training.
- Steps drop from 8 to 2; denoising at 1024×1024 goes from 76.4s to 19.3s (4x faster). Fine-texture energy reaches 0.97–1.09x the teacher (stock Turbo at 2 steps: 0.40–0.65x).
- Works best for close-up subjects; small subjects still ghost or smear — use the 4-step LoRA there. Saved steps can instead buy resolution: renders up to 2048×2048, Krea's recommended max, though the base model duplicates subjects beyond that.
- The new checkpoint adds 9,720 training samples targeting duplicated subjects, blended poses, and faces in busy scenes, and is the first evaluated on a full 22-prompt sweep (7 new structure-testing scenes added). A blind rubric misses 2 of 352 items; a judge prefers the 8-step teacher in 12 of 66 cases, mostly on style.
More from Multimodal
- Degrading latents don't matter:分段 fresh latents + FL2V 接缝修复跑通长视频 — roychodraws · 2026-10-03
- A meta-prompt that turns any LLM into a YuE2 music prompt builder — Fabulous_Broad_115 · 2026-10-03
- Full prompt shared for a 15-second one-shot magical winter AI video — umesh_ai · 2026-10-03
- AI-Generated Game of Thrones Anime Trailer Wows With Cinematic Quality — azed_ai · 2026-10-03
- AI video experiment: 'Journey of a morning coffee' camera-movement prompt shared — umesh_ai · 2026-10-03
- 30 years of speech synthesis: IEICE webinar on statistical TTS now public on YouTube — heiga_zen · 2026-10-03