Krea 2 Turbo 4-Step Distillation LoRA: 8 Steps Down to 4
A community distillation LoRA cuts Krea 2 Turbo from 8 steps to 4 with texture at teacher parity and a prompt-aware critic. Plain LoRA weights, no custom nodes needed.
Why a LoRA, and why it was hard
Step distillation usually demands changes concentrated in the timestep pathway: an analysis of Krea's own Raw-to-Turbo delta shows time_embed.linear_2 moves about 5.5x more than any block linear. A LoRA that froze the timestep path would withhold exactly the weights the task needs most, so this LoRA trains the timestep pathway as well, while staying a plain rank-64 LoRA.
The current checkpoint (chk00042000) adds two upgrades over earlier ones:
- Texture at teacher parity: fine-detail energy is 1.00x the 8-step teacher's at 1280x1280 and 1.02x at 1440x1440, matched band-for-band across the frequency spectrum (the previous checkpoint measured 0.88x and 0.82x).
- Prompt-aware critic: training scores images against their prompts, so adherence is pressured directly rather than inherited.
 |  |  |
|---|---|---|
| Turbo 8 steps | Turbo 4 steps: detail suffers | 4 steps with the LoRA: texture restored |
ComfyUI usage
The repo ships _comfyui-suffixed files alongside the diffusers weights:
krea2_turbo_4step_rank_64_lora_latest_comfyui.safetensorsfor normal use (a rolling pointer to the newest accepted checkpoint)krea2_turbo_4step_rank_64_lora_chk00042000_comfyui.safetensorsto pin this exact checkpoint
Load it on top of Krea 2 Turbo, run 4 steps instead of 8, and keep guidance at 0.0. It is a drop-in plain LoRA: no custom nodes, no patched sampler, no code. Twelve trained resolutions from 512x512 to 1440x1440 are published with per-resolution sweeps in the repo.
The gains even reach the out-of-spec 2-step preview mode, where the hollow mid-frequencies of the old checkpoint are visibly filled in.
Comments
Sign in with GitHub to join the discussion.