Qwen-Image-2.1 Turbo Distill 8-Step LoRA (SVD Extract)

Post-hoc LoRA adapters extracted from the weight delta between Qwen/Qwen-Image-2.1-Turbo and Qwen/Qwen-Image-2.1 (base).

Which file to download

ComfyUI (recommended for most users): qwen_image_2.1_turbo_distill_comfy_8step_r64/r128/r256.safetensors (0.60 / 1.21 / 2.51 GB). Built for ComfyUI 0.38+; loads with zero LoRA key warnings. Use these in ComfyUI.
Diffusers: qwen_image_2.1_turbo_distill_8step_r64/r128/r256.safetensors (0.42 / 0.84 / 1.73 GB). The original unfused-MLP files. In ComfyUI 0.38+ these skip the MLP layers ("lora key not loaded" warnings), so ComfyUI users should pick the ComfyUI files.
r128 or r256 match the reference quality closely; r64 is smaller with slightly softer text.

Tested in ComfyUI 0.38: strength 1.0, 8 steps, euler / simple, CFG 1, 1024px.

For each weight matrix we compute ΔW = W_turbo − W_base and factor it with truncated SVD into low-rank LoRA pairs. Apply the LoRA to the base model to get turbo-like 8-step results without swapping in the full Turbo checkpoint.

Experimental. This is not an official Qwen release. The extraction is a low-rank approximation of the distillation delta, not an exact copy of Turbo.

Choosing a rank

r256 (1.73 GB): delta energy captured 46.4% overall / 40.4% median layer; per-layer rank 4 / 136.5 / 763 (min / median / max). Closest match to Turbo.
r128 (0.84 GB): 37.4% / 32.1%; per-layer rank 4 / 79 / 351. In between.
r64 (0.42 GB): 30.4% / 24.7%; per-layer rank 4 / 42 / 185. Tight disk/VRAM or LoRA stacking.

"Mixed" means the rank budget (average rank × 232 linear layers) is shared out across layers greedily by singular-value energy, so layers whose Turbo−base delta is more concentrated get higher rank (minimum 4). The two changed norm weights are stored exactly as full diffs. "Energy captured" is the share of ‖W_turbo − W_base‖²_F kept by the low-rank factors, measured during extraction.

All three ranks gave good 8-step results in our showcase, and generation speed was the same for each (about 13 s per 1024px image on an RTX 3090). Pick r64 if disk space or VRAM is tight, or if you're stacking it with other LoRAs. Pick r256 for the closest match to Turbo. r128 sits in between.

Visual comparison gallery

Rendered with Qwen-Image 2.1 base + this LoRA at 8 steps, seed 42, settings below. Each 2×2 sheet shows a header card (top-left), then Rank 256, Rank 128 and Rank 64 for the same seed and prompt.

How to use (ComfyUI)

Load the Qwen-Image 2.1 base diffusion model (not Turbo).
Apply one of the LoRA files to the model.
Sample with the official Comfy-Org Qwen-Image 2.1 Turbo template settings: text encoder node TextEncodeQwenImage21; guider CFGGuider, CFG 1; sampler euler via SamplerCustomAdvanced; sigmas (8 steps, ManualSigmas): 1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0
VAE: we recommend madebyollin/texture-fix-vae-for-qwen-image-2.1 (https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1), which reduces the checkerboard texture the stock VAE can leave.

Edits work the same way, with source images passed to TextEncodeQwenImage21 (image_1 … image_10).

No trigger words.

Caveats

Approximation, not identity. The adapters recover much of the base→Turbo shift, but they do not guarantee pixel-level parity with native Turbo.
Scheduler matters. Use the 8-step Turbo settings above when you evaluate.

License

These adapters are derived from Qwen-Image 2.1 weights and inherit the Qwen-Image 2.1 license (https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE). Non-commercial use only: no selling generated images, no generation services, no selling the model or merges. Non-commercial sharing is allowed.

Base models: Qwen/Qwen-Image-2.1 · Qwen/Qwen-Image-2.1-Turbo

Full card and gallery: https://huggingface.co/TheDivergentAI/Qwen-Image-2.1-Turbo-Distill-8step-LoRA