nynxz/Qwen-Image-2.1

Qwen-Image 2.1 Turbo LoRA (153 MB)

A small LoRA that turns the Qwen-Image 2.1 base model into the Turbo model: 8 steps, CFG 1. It is extracted from the difference between Comfy-Org's qwen_image_2.1_bf16 and qwen_image_2.1_turbo_bf16.

It is 153 MB. The official extracted turbo LoRA (qwen_image_2.1_turbo_lora_avg_rank_178_bf16) is 913 MB, and both give the same results.

File	Size
loras/qwen_image_2.1_turbo_lora_fro_153mb_bf16.safetensors	153 MB

Same prompt and seed in every column: 1024×1024, 8 steps, CFG 1, euler / simple.

Usage (ComfyUI)

• Load the base diffusion model qwen_image_2.1_bf16.safetensors, not the turbo one.
• Add LoraLoaderModelOnly with this file at strength 1.0.
• Sample with 8 steps, CFG 1, euler / simple.
• Text encoder and VAE are the standard Qwen-Image 2.1 ones (qwen3vl_8b, qwen_image_2.1_vae).

How it was made

1. SVD of every weight delta (turbo − base) for all 200 linear layers.
2. Global Frobenius allocation. Each kept singular direction costs out + in parameters and removes σ² of squared error. Directions are taken across all layers, in order of error removed per parameter, until the size budget is spent. Ranks run from 9 to 75 per layer (mean 26.5), against roughly 180 everywhere in the official LoRA.
3. Exact patches for the small, sensitive layers. The latent input projection (img_in), the output projection (proj_out) and the timestep embedder's first linear layer are stored as exact .diff weights, as are the changed norm scales. Together they add about 3 MB.

The format is standard ComfyUI: lora_up / lora_down + alpha, plus .diff.

How close it is

Weight space, as a fraction of the base weights (whole model):

	Error vs Turbo weights
Turbo's change from base	0.238% (the full delta)
Official LoRA, 913 MB	0.186%
This LoRA, 153 MB	0.209%
bf16 rounding already in the Turbo file	0.166%

The Turbo finetune is full-rank, and most of it sits at bf16 rounding scale. Both LoRAs keep the structured part and drop the remainder.

In renders (same prompt and seed, PSNR):

Comparison	PSNR
This LoRA vs the official LoRA	20.4 dB
Official LoRA vs the Turbo model	18.5 dB
This LoRA vs the Turbo model	16.6 dB

The two LoRAs are closer to each other than either is to the Turbo model. Any extracted LoRA drifts slightly from the full Turbo model in small details, such as an extra lamp or the amount of mist. Quality, sharpness and text rendering match.

If you need the exact Turbo behaviour, use the full qwen_image_2.1_turbo_bf16 model.

License

Derived from Qwen-Image 2.1 and Qwen-Image 2.1 Turbo; same license (qwen-research).