Arekkusul/krea-2-raw-distill-lora

Krea 2 Raw — Single-Stage Step-Distillation LoRA

A LoRA adapter for krea/Krea-2-Raw (the undistilled 12B single-stream flow-matching DiT behind Krea 2) that halves the sampling step count — 52 steps → 26 steps — while reproducing the full-quality composition and detail of the 52-step baseline, verified on held-out prompts never seen during training.

This was trained on a from-scratch conversion of a Krea 2 Raw checkpoint, for anyone in the same situation: you have a converted Raw checkpoint but not the official, separately and expensively trained Turbo checkpoint, and want a faster path without needing that separate model.

What this is NOT

• Not Krea 2 Turbo. Turbo is a genuinely separate, TDM-distilled model (8 steps, CFG-free) — not something a LoRA can reach from Raw. This LoRA is a single-stage, ~2x step reduction, not Turbo's ~6.5x jump. Don't expect Turbo-grade speed or quality.
• Not exhaustively validated. Verified on 2 held-out prompts (a coffee-cup still life, a detailed bicycle/brick/ivy scene) across two training runs (6-image and 12-image datasets). That's real evidence the method works, not a comprehensive quality guarantee across all content types, styles, or resolutions.

Method

Single-stage progressive distillation (Salimans & Ho 2022, "Progressive Distillation for Fast Sampling of Diffusion Models"), adapted to run on a single Apple Silicon Mac with no second trainable model copy resident in memory:

1. A real training photo's VAE latent (x0) + random noise define a point x_i on the flow-matching trajectory at a random noise level.
2. The frozen base model itself (no LoRA) takes two small reference Euler steps from x_i forward — this is the "teacher" signal, computed fresh every training step, no separate teacher model needed.
3. The LoRA-adapted model takes one big Euler step covering the same span — the "student."
4. Loss = MSE between the two, backprop only through the student's LoRA factors.

This method was designed after five published few-step-distillation methods (TDM, Hyper-SD, RAPM, LCM-LoRA, Progressive Distillation's full cascade) were researched and ruled out as infeasible on consumer Apple Silicon hardware — each hits a memory wall (needing 2-3 resident model copies), a compute wall (hundreds of GPU-hours), or is unvalidated for flow-matching architectures.

Training details

• Base model: a krea/Krea-2-Raw conversion (12B, hidden=6144, 28 single-stream blocks, GQA 48Q/12KV heads).
• LoRA target modules: to_q, to_k, to_v, to_out.0 on all 28 transformer blocks (112 targets total) — the standard PEFT attention surface.
• Rank: 16, alpha: 16.
• Optimizer: AdamW, learning rate 1e-4, gradient norm clipped to 1.0.
• Training data: 12 self-generated 1024x1024 images spanning varied subject matter and texture (portrait, animal, food, landscape, interior, still life) — no external/copyrighted training images.
• Steps: 800, batch size 1, ~110-130s/step on an Apple M4 Pro (64GB unified memory). Final training loss: 0.0015 (bounded, noisy-but-healthy throughout — flow-matching loss varies naturally by sampled noise level, not a sign of instability).
• Hardware: entirely trained on a single Mac mini M4 Pro, no cloud GPU.

Usage

Diffusers/PEFT-format LoRA (standard {path}.lora_a/{path}.lora_b naming, alpha/rank metadata) — load with any Krea 2 Raw-compatible LoRA loader. Recommended: 26 steps, guidance scale 3.5 (matches training conditions) — other step counts/guidance are untested.

License

Inherits the Krea 2 Community License from the base model (krea/Krea-2-Raw) this LoRA is a derivative of — free for personal/non-commercial use, with the base license's content-filtering obligation for deployers. Full text: LICENSE.pdf in this repo (copied verbatim from the base model).

Known limitations

• Trained on a small (12-image) dataset — may not generalize to all content types equally well.
• Verified on 2 held-out prompts only.
• The base krea/Krea-2-Raw model has a separate, pre-existing content-dependent artifact at reduced step counts (unrelated to this LoRA — confirmed via isolation testing to occur, and be worse, even WITHOUT this LoRA). This LoRA does not cause it and did not worsen it in testing, but it's a base-model characteristic worth knowing about.