41 hours on one H100: 3-step distill LoRA beats Krea 2 Turbo's own fine detail
TimeTruth2490 · reddit · 2026-10-29
An independent developer released a third step-distillation LoRA for Krea 2 Turbo, cutting image generation from 8 steps to 3 with 2.6x faster denoising (65.4s to 24.9s at 1024×1024).
Quality: fine texture hits 1.15× the 8-step teacher on 22 test prompts, above the 4-step LoRA's 1.04× (stock Turbo at 3 steps: 0.58×), with color at teacher level. Trained from 30k recorded teacher trajectories and 12k samples on a single cloud H100 in 41 hours, starting from the 4-step adapter's weights.
Method notes: distribution matching rather than trajectory imitation to avoid averaging into blur; eight critic heads (artifacts, real photos, layout, prompt agreement, lettering, large faces, etc.) on the frozen base's mid-network features; 11 training resolutions from 512×768 to 1822×1024. Rank-64 LoRA, drop-in for diffusers / ComfyUI / MLX with provided workflows including an MLX one for Apple Silicon.
More from Multimodal
- "Darkness": a stunning dusk city flyover AI video demo — sdrowegnarts · 2026-10-30
- Falcon: Streaming Audio-Visual Speaker Extraction Cuts Params 91.9% and Latency 11× — _akhaliq · 2026-10-30
- Open-source StableGen 0.4 turns prompts into rigged, textured characters inside Blender — sakalond · 2026-10-30
- Grok appears to have access to Suno's MCP connector, gated by early access — koltregaskes · 2026-10-30
- A visual guide to multimodal generation: 50+ figures, free bonus chapter — MaartenGr · 2026-10-30
- MOSS-Image vs SenseNova U1.5 Lite: head-to-head on text, layout, photo and comics — Negative-Whereas3307 · 2026-10-30