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
- Sophon-Realtime-v1, a full-duplex speech-to-speech model, trends on Hugging Face — adventists-ai · 2026-11-11
- Asked Fable 5.2 to 'go wild' making a game, got a Pong clone — and loved it — ariG23498 · 2026-11-11
- Western town built with Fable 5.2: ray-traced weathering ages every board — ravisparikh · 2026-11-11
- Diffusion Week 2026 releases all 10 lectures on video generation, world models and more — LucaAmb · 2026-11-11
- Running Wan 2.2 on a RTX 3080 12GB: 5B Breaks Down, 14B Works Surprisingly Well — fixxxer2606 · 2026-11-11
- Indie dev teases Z-Image-based model 'on par with latest closed-source models' — darlens13 · 2026-11-11