Krea2 Turbo 4-step distill LoRA cuts generation time 1.6x while matching 8-step texture quality
TimeTruth2490 · reddit · 2026-09-07
Developer lvladikov released the final version of a 4-step distillation LoRA for Krea 2 Turbo, cutting the minimum usable step count from 8 to 4: 1024×1024 generation drops from 88.7s to 54.5s end-to-end (1.6x), and 1.8x on denoising alone.
Key details:
- Texture: fine-texture energy hits 1.03x the 8-step teacher at 1280/1440px, with fewer clipped highlights and shadows and skin texture at 0.97–0.99x.
- Distillation quality: velocity error to the teacher fell from 4.70e-02 (stock 4-step) to 2.59e-02 — 45% of the gap closed; a VLM judge preferred the teacher in only 6 of 45 side-by-side comparisons.
- Training: 13,750 teacher 8-step trajectories + 43,044 real-photo crops (LSDIR, Flickr2K), 78,000 samples total, prompt-aware critic scoring; 21 days on a single RTX 3090 with 16 recipe adjustments.
- Drop-in: plain LoRA weights for diffusers and ComfyUI plus a ready workflow, no custom nodes; keep guidance at 0.0, or dial strength to 0.6–0.9 if texture feels too strong.
More from Multimodal
- ByteDance preps real-time spatial video generation model, Zhang Yiming personally overseeing launch — pstAsiatech · 2026-09-07
- Engineer trains a bufo LoRA to auto-generate style-consistent Slack memes — emilyzsh · 2026-09-07
- Gemini's agentic video understanding cuts tokens by 88% and costs by 66% — patloeber · 2026-09-07
- GPT-6 Astra builds Rhino 3D canopy model, images and camera views end-to-end — Hailuo_AI · 2026-09-07
- MiniMax H3 two-step workflow beats one-step speed at same size, quality holds — Hot_Store_5699 · 2026-09-07
- Gemini Omni Flash 1.1 turns everyday objects into pets — prompts included — Able-Line2683 · 2026-09-07