Community LoRA distills Krea 2 Turbo to 4 steps, cutting render time ~1.6x
TimeTruth2490 · reddit · 2026-08-20
A Reddit user released a 4-step distillation LoRA for Krea 2 Turbo, dropping the minimum usable step count from 8 to 4.
Unlike circulating Raw→Turbo extraction LoRAs (low-rank projections of an existing weight delta), this one is trained via distillation on top of Turbo, shortening its own schedule rather than reproducing it.
Usage: LoRA scale 1.0, 4 steps, CFG 0.0 (Turbo is CFG-free), timestep shift mu=1.15, with Turbo's own 4-sigma deployment grid. Running it on Raw is unsupported: results are malformed below 14 steps.
Measured at 1024×1024 on Apple Silicon (MLX, bf16): 8-step baseline takes 88.7s total; 4 steps + LoRA takes 54.5s — 1.6× faster, or 1.8× counting denoise alone. The LoRA's job is restoring the high-frequency detail a 4-step schedule loses. Training is ongoing; the latest file is a rolling pointer.
More from Multimodal
- FLUX 3 Upscaler hits Higgsfield: 50% faster and cheaper, native 2K/4K video — tisch_eins · 2026-10-16
- AI-generated "RiRi" images stun with photorealistic diamond-gown Rihanna looks — egeberkina · 2026-10-16
- SeedAudio 1.5 hits CapCut PC: one prompt generates dialogue, SFX and music — LearnWithBishal · 2026-10-16
- DLSS 5 Assetto Corsa demo stuns sim racers: 'better than reality' — CSProfKGD · 2026-10-16
- Video recast with fal agent: swap a character via one character sheet — adamho · 2026-10-16
- Dev open-sources Krea 2 pixel-art LoRA for 16x16 game icons with transparent backgrounds — Angaisb_ · 2026-10-16