Community-Built LoRA Cuts Krea 2 Turbo's Image Generation Steps in Half
A community-trained LoRA cuts Krea 2 Turbo from 8 to 4 sampling steps, delivering roughly 1.6x faster image generation with minimal quality loss.
- New LoRA halves Krea 2 Turbo sampling from 8 to 4 steps while preserving detail.
- Delivers roughly 1.6x wall-clock speedup at 1024x1024, measured on Apple Silicon in bf16.
- Trained via progressive distillation with Turbo acting as its own teacher.
- Rank 64 bf16 adapter touches 228 modules including time embedding, not just transformer blocks.
- Trained on a single RTX 3090 using NF4-quantized frozen base; released weights are unquantized bf16.
- Available on Hugging Face with a ready ComfyUI workflow and hosted demo Space.
Krea 2 Turbo is a distilled, guidance-free variant of Krea's flagship text-to-image model that runs at 8 steps instead of a full diffusion schedule. A new community-trained LoRA called Krea2-Turbo-Distill-4step-LoRA halves that to just 4 steps while attempting to preserve the fine texture and edge detail that a shorter trajectory normally sacrifices.
Most Krea 2 LoRAs in circulation are extractions: a low-rank projection of the weight difference between Krea 2 Raw and Krea 2 Turbo. Applied to Raw, they reproduce Turbo's output. They are delivery mechanisms for a model that already exists, and they cap out at Turbo's 8 steps. This adapter is trained from scratch via distillation and assumes Turbo is already loaded underneath.
Half the steps, real wall-clock savings
LoRAs add per-step overhead, so the real question with any step-reduction adapter is whether it actually saves time. Measured on Apple Silicon at 1024×1024 in bf16, 4 steps with the LoRA runs about 1.6× faster than the 8-step baseline: 54.5 seconds against 88.7 seconds, saving roughly 39% of wall-clock time. Denoise-only speedup is closer to 1.8×.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
