Qwen Image 2.1 2-step distill LoRA hits v1.5, enabling realtime editing on a single GPU
linoy_tsaban · x · 2026-10-31
Developer linoytsaban released v1.5 of their 2-step distilled LoRA for Qwen-Image-2.1, with improved prompt adherence and edit fidelity; the realtime demo Space has been updated accordingly.
Technical details:
- Trained with DMD2 (on-policy distribution matching, no regression/GAN loss), compressing the base model's 20-step sampling into 2 forward passes while staying close to teacher quality, competitive with the 8-step Turbo variant
- LoRA rank 32, alpha 32, attention-only (q/k/v/out); fixed sigmas [1.0, 0.5], truecfgscale=1.0
- Multi-aspect (1 MP) training, 2000 steps on a single H100
- Runs in realtime on ZeroGPU (half an RTX Pro 6000), enabling interactive live editing; multi-reference composition ("place the cat from Picture 1 on the beach from Picture 2") still works at 2 steps
Weights are open on Hugging Face under the qwen-research license; loading v1.5 only requires pointing weightname at the new checkpoint, with all other inference settings unchanged.
More from Multimodal
- GVOID tool appears to run WAN 2.7 with new Cinematic mode and start-end frame interpolation — MillsLittlePistachio · 2026-10-31
- Prompt template: manga-style character sheet with multi-angle views and expressions — azed_ai · 2026-10-31
- Glif demos end-to-end brand content with hundreds of castable AI models — Puzzleheaded-King584 · 2026-10-31
- Claude Opus drives Blender MCP to build 3D models straight from 3DGS — janusch_patas · 2026-10-31
- FastSDCPU Halloween release adds Fibo 2.0 Edit Turbo image editing via OpenVINO on CPU — simpleuserhere · 2026-10-31
- d1-2B decision model skims 1864 public-domain films to answer yes/no questions — vanstriendaniel · 2026-10-31