TheDivergentAI/krea2-turbo-distill-lora

Krea-2 Turbo Distillation LoRA (SVD Extract)

Post-hoc LoRA adapters extracted from the weight delta between krea/Krea-2-Turbo and krea/Krea-2-Raw.

For each 2D weight matrix in the Krea-2 transformer, we compute ΔW = W_turbo − W_raw and factor it with truncated SVD (torch.svd_lowrank, q=256) into low-rank lora_A / lora_B pairs. The goal is to approximate turbo behavior on the Raw base model without swapping the full 24 GB checkpoint.

Experimental. This is not an official Krea release. Distillation is not purely low-rank, and turbo inference also depends on scheduler settings (8 steps, CFG=0, mu≈1.15).

Files in this repo

File	Rank	Size	Notes
krea2_turbo_distill_r64.safetensors	64	~0.47 GB	Smallest; rougher fit
krea2_turbo_distill_r128.safetensors	128	~0.94 GB	Recommended starting point
krea2_turbo_distill_r256.safetensors	256	~1.87 GB	Closest fit; largest file
extraction_report.json	—	~15 KB	Per-layer reconstruction metrics

Each LoRA file contains 530 tensors (265 layers × lora_A + lora_B).

Reconstruction quality

Approximation error for the 2D weight delta (lower is better):

Rank	Global recon error	Mean singular energy captured
64	46.1%	86.8%
128	41.5%	93.6%
256	36.6%	100%*

*Energy is computed from the top-256 singular components returned by svd_lowrank(q=256).

Worst-fit layers tend to be text/timestep MLP projections (txtmlp.*, tmlp.*, tproj.*). See extraction_report.json for per-layer details.

Visual comparison gallery

Side-by-side rank comparisons on Krea-2 Raw with turbo distill LoRA at 8 steps, CFG 0, mu 1.15. Each grid shows the prompt and metadata (top-left), then Rank 256, Rank 128, and Rank 64 outputs for the same seed and settings.

Panel	Content
Top-left	Prompt + generation settings
Top-right	Rank 256 output
Bottom-left	Rank 128 output
Bottom-right	Rank 64 output

Chroma Aperture

Rocket Launch Exhaust · 9:16

Designer Toy Figure · 1:1

Vintage Analog Collage · 5:4

Anime Portrait Smile · 3:4

Ocean Wading Illustration · 9:21

Light Spill

Tree and Dog Landscape · 16:9

Portrait with Lilies · 4:5

Harvest Mouse Macro · 3:2

Sailor Girl Motion · 2:3

Coastal Convertible Sunset · 4:3

Split Spectrum

Stone Guardian Ruin · 9:16

Jungle Fox Tapestry · 21:9

Retro Chrome Spaceface · 16:9

Gold Ribbon Portrait · 2:3

Menacing Jester Fantasy · 1:1

Analog Echo

Fashion Editorial Crimson · 4:5

Ink Faces Landscape · 3:4

Vintage Anime Crowd · 3:2

Windy Anime Portrait · 4:3

Moody Close-Up Portrait · 1:1

Signal Grid

Turbo Distill Keynote Hero · 3:4

Rank Ladder Laboratory · 3:4

Eight-Step Horizon · 3:4

Neural Condenser Array · 3:4

Raw Versus Turbo Split · 3:4

How to use

1. Load the Krea-2-Raw transformer (not Turbo) with ComfyUI or HuggingFace diffusers.
2. Apply one of the LoRA files above on the diffusion transformer.
3. Generate with turbo-style settings:
   • Steps: 8
   • CFG / guidance scale: 0
   • Timestep shift mu: 1.15 (recommended for turbo)

Start with krea2_turbo_distill_r128.safetensors. Use r256 if you need a tighter weight approximation; use r64 only if VRAM or file size is constrained.

Key format

Keys follow the ComfyUI Krea2 LoRA convention:

diffusion_model.blocks.0.attn.wq.lora_A.weight
diffusion_model.blocks.0.attn.wq.lora_B.weight

LoRA alpha equals rank (64, 128, or 256 respectively).

Caveats

• Approximation, not identity. These adapters recover part of the Raw→Turbo weight shift; they do not guarantee pixel-level parity with native Turbo.
• Scheduler matters. Turbo expects few-step, CFG-free sampling. Match turbo settings when evaluating.
• Official Krea workflow. Krea recommends training LoRAs on Raw and running them on Turbo. These adapters explore making Raw behave more like Turbo via an extracted weight delta.

Method

• SVD low-rank extraction on (W_turbo − W_raw) per 2D layer
• Source checkpoints: krea/Krea-2-Raw, krea/Krea-2-Turbo

License

These adapters are derived from Krea-2 weights and inherit the Krea-2 community license. See Krea licensing for commercial use terms.

Citation

If you use Krea-2, please cite the Krea team:

@misc{krea-2-2026,
  author = {Sangwu Lee and Erwann Millon and Le Zhuo and others},
  title = {{Krea 2}},
  year = {2026},
  howpublished = {\url{https://www.krea.ai/blog/krea-2-technical-report}}
}

Base models:

• krea/Krea-2-Raw
• krea/Krea-2-Turbo