Viggle/Qwen-Image-2.1-viggle-turbo — new model trending #30 on Hugging Face
Viggle releases a six-step distilled LoRA of Qwen-Image-2.1 that runs about five times faster.
Viggle released Qwen-Image-2.1-viggle-turbo v0.2.1, a Distribution Matching Distillation student of Qwen-Image-2.1 for text-to-image and instruction-driven editing with one to three reference images. The recommended rank-256 LoRA is 1.3 GB in bf16 and samples in six steps on a fixed sigma schedule, with no classifier-free guidance, versus 40 steps for the base model and about five times faster end to end. On a 96-request held-out set, v0.2.1 shows 0.98 times the base model's sample diversity and 0% composition drift, while dense text and complicated multi-reference edits can still fall short. A 680 MB rank-128 variant and ComfyUI workflows are also published.
- Six-step LoRA needs no classifier-free guidance and is about 5× faster.
- Recommended rank-256 adapter is 1.3 GB; rank-128 cut is 680 MB.
- Held-out diversity is 0.98× the base model with 0% composition drift.
- Dense text and complex multi-reference edits still lag the 40-step base.
- Vendors
- Viggle
- Products
- ComfyUI
- Organizations
- ViggleHugging Face
From the original article
12 min read at huggingface.cotext-to-image · 47,873 downloads · 214 likes · license other
By Viggle · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.