Krea2 Turbo 2-step distill LoRA hits 4.2x faster denoising at near-teacher quality
TimeTruth2490 · reddit · 2026-09-26
A new checkpoint (chk31600) of the Krea 2 Turbo 2-step distillation LoRA cuts denoising from 8 steps to 2: 1024×1024 renders drop from 81.4s to 19.5s (4.2x faster), with fine-texture energy at 0.95–1.08x of the teacher (stock Turbo at 2 steps: 0.39–0.57x).
Where it works: close-up subjects (portraits, single objects) hold up at 2 steps; small subjects — faces in crowds, distant objects — still come out ghosted, where the author's 4-step LoRA remains the pick. Saved steps can trade for resolution up to 2048×2048; beyond that the base model duplicates subjects.
Training: reuses 13,750 teacher trajectories from the 4-step project (zero new teacher runs), 31,600 samples in the 2-step stages, 18 days on a single RTX 3090. Rank 64 across 228 modules, works with stock Euler in diffusers/ComfyUI/MLX, no custom nodes. A blind rubric misses 1 of 240 prompt-conditioned items; a judge prefers the 8-step teacher in only 11 of 45 comparisons.
More from Multimodal
- PixVerse R2 lets you walk through AI-generated worlds that react in real time — Aiden_Tech_Ai · 2026-09-26
- PixVerse opens R2 real-time world model: one base model, distilled straight to real-time — kimmonismus · 2026-09-26
- PixVerse R2 gallery: interactive film branches are generated live, not pre-written — kimmonismus · 2026-09-26
- Hands-on with PixVerse R2: WASD-controlled worlds and interactive films generated in real time — kimmonismus · 2026-09-26
- PixVerse R2 hands-on: WASD controls with live frame-by-frame world generation — kimmonismus · 2026-09-26
- Pexo turns hours of research into a finished 30s explainer video in 15 minutes — omarsar0 · 2026-09-26