Viggle/Qwen-Image-2.1-viggle-turbo

Viggle‑Turbo v0.2.1 is a distilled version of Qwen‑Image‑2.1 that produces images in six sampling steps. The student model runs about five times faster than the 40‑step teacher while keeping quality that is competitive with the base, and on some prompts the outputs are preferred. Reported diversity is 0.98 × the base model and composition drift is zero, meaning the generated scenes stay faithful to the prompt. The system can perform both text‑to‑image generation and image editing; leaving the reference slots empty triggers pure generation, while filling one to three slots enables editing, composition or style transfer. A demo space lets developers preview results without GPU access, showing pre‑rendered examples that use a fixed seed, the six‑step schedule and the exact prompt text. The model can also be loaded with a LoRA adapter (rank 256) on top of the base transformer, or with the full fine‑tuned transformer from version 0.1, depending on the chosen configuration. Known limits include difficulty with complex multi‑reference edits, face swaps and identity‑preserving changes, where the 40‑step base still outperforms the distilled version. Very small dense text may print less cleanly at six steps, though eight steps improve this. The current release uses the v0.2.1 LoRA weights; future updates may adjust diversity and sharpness further.