Introduction

We already covered Qwen-Image-2.1 landing as open weights with ComfyUI-friendly tooling. The follow-up question was obvious: can you keep that quality without waiting on a full 40-step roll? On Sep 24, 2026, Viggle answered with Qwen-Image-2.1-viggle-turbo v0.2.1 — a Distribution Matching Distillation (DMD) student that does text-to-image and instruction editing (1–3 reference images) in about six transformer passes, with no CFG, and roughly 5× faster end to end.

The model card is trending hard on Hugging Face (~12k downloads). Grab it here: Viggle/Qwen-Image-2.1-viggle-turbo.

What Turbo Actually Ships

v0.2.1 is the step-700 checkpoint of the same LoRA run that shipped v0.2 at step 600 a day earlier. Prefer these files:

  • Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors — rank 256 / alpha 256, bf16, ~1.3 GB (what the demo Space runs)
  • Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors — rank 128 cut (~680 MB) used by the ComfyUI workflows

Load the LoRA on top of Qwen/Qwen-Image-2.1 at runtime — do not merge it into bf16 (lossy). Text encoder, VAE, and processor stay on the base repo.

The 6-Step Schedule That Matters

Sample with 6 steps and this exact sigma ladder:

sigmas=[1.0, 0.9375, 0.875, 0.75, 0.5, 0.25]
true_cfg_scale=1.0

No negative prompt. Keep LoRA strength at 1.0. Use the shipped scheduler (shift_terminal: null) — the base 0.02 terminal shift wrecks the last step.

Viggle’s held-out checks say v0.2.1 stays near base diversity (~0.98×) with 0% composition drift vs the 40-step teacher on their 96-request set. Small dense text is still where the base model wins; 8 steps help on those prompts. Complicated multi-ref edits (face swaps, heavy identity locks) can still fall short of 40-step Qwen.

ComfyUI and Diffusers Paths

ComfyUI (tested around 0.37.0 with native Qwen-Image-2.1 support): copy comfyui/viggle_turbo.py into custom_nodes/, then drag in the shipped t2i / edit workflows. Two custom nodes matter:

  • Viggle Turbo Sigmas — the 6-step schedule with resolution-dependent shift (pair with euler + BasicGuider, no CFG)
  • Viggle Turbo LoRA (unmerged) — runtime W x + B A x so you avoid stock merge loss into bf16/int8

Drop the r128 (or r256) LoRA into loras/; base diffusion / text-encoder / VAE files follow Comfy-Org’s usual Qwen-Image-2.1 layout.

Diffusers needs a pinned git install for QwenImage21Pipeline, then load_lora_weights(...) for the v0.2.1 file and the same STEPS / SIGMAS on every call. Editing takes image=[...] with up to three refs; order maps to <image1> / <image2> in the prompt.

License Caveat

This is a Qwen RESEARCH LICENSE derivative: fine for research and evaluation; commercial use needs a separate licence from the Qwen licensor. Read LICENSE / NOTICE on the card before you ship anything client-facing.

Original Source

https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo

Viggle, Qwen-Image-2.1-viggle-turbo v0.2.1 (2026-09-24). Base: Qwen/Qwen-Image-2.1. Prior ArtRealmAI coverage of the open weights: Qwen-Image 2.1 open weights + ComfyUI.

Conclusion

Turbo is the practical sequel to Qwen-Image-2.1: same family, far fewer steps, LoRA adapters plus ComfyUI nodes ready to try. Stick to the 6-step sigma schedule, keep the LoRA unmerged, and remember the research licence if money is on the line. For creators who already run Qwen 2.1 locally, this is the “make it snappy” drop worth pinning.

—Aurelia ♡