Krea 2 Turbo distilled to 4 steps: 1.6x faster, closer to the 8-step teacher
RisingSayak · x · 2026-11-10
lvladikov released a 4-step distillation LoRA for Krea 2 Turbo, cutting the minimum usable step count from 8 to 4 at 1.6x end-to-end speed, with a Diffusers snippet in the model card.
- Retrained on 133k samples (1.7x the first release) with five critic heads: grid patterns, ghosting and flat-area grain all closer to the 8-step teacher; no over-sharpening at 1440px.
- A new 3-step LoRA takes it to 3 denoising steps at 3/4 of the 4-step LoRA's time, with fine texture at 1.26x the teacher's on 22 test prompts.
- Origin story: the author planned 'a few more' training samples and ended up adding 55k and reworking the critic architecture.
More from Multimodal
- 10 open-source GitHub repos for voice AI: from local ASR to zero-shot voice cloning — Shruti_0810 · 2026-11-10
- Fable 5.2 generates animated nature scene with ~70k trees from one prompt — Merzmensch · 2026-11-10
- MOSS-Image roundup: Control LoRAs, 461 new characters, and 20-minute LoRA training on a 5090 — optimisticalish · 2026-11-10
- StreamEval from Tsinghua, NVIDIA, CMU and Berkeley exposes gaps in streaming video generation across 38 models — jiqizhixin · 2026-11-10
- Reddit hot take: World Labs' deterministic camera control beats rolling the dice with AI video — Spiritual_Skirt_9312 · 2026-11-10
- LightX2V releases 4-step Qwen-Image-2.1 Lightning LoRA, trained on 16 Huawei Ascend 910B NPUs — RisingSayak · 2026-11-10