2-step DMD2 LoRA brings Qwen Image 2.1 to real-time generation on ZeroGPU
multimodalart · x · 2026-10-24
Hugging Face's linoytsaban released Qwen-Image-2.1 Realtime V1: a 2-step DMD2 distillation LoRA that runs in real-time on ZeroGPU (half an RTX Pro 6000), with the live-editing demo unsped up.
- DMD2 (on-policy distribution matching, no regression/GAN loss) with teacher CFG guidance
- LoRA rank 32/alpha 32, attention-only (toq, tok, tov, toout.0)
- Inference: 2 steps, fixed sigmas [1.0, 0.5], truecfgscale=1.0
- Trained 2000 steps on a single H100 (1MP multi-aspect data)
- Quality competitive with 8-step Turbo and 40-step base; instruction editing and multi-reference composition still work at 2 steps
- Requires the diffusers dev branch (QwenImage21Pipeline); LoRA and Space are open-sourced
More from Multimodal
- Tencent's Open-Source World Model Hy-Nomas: 720p/24fps, Runs on a Single RTX 4090 — mishig25 · 2026-10-24
- Kling 4.0 goes live: 30-second clips, 15 references and 10 keyframes per scene — miilesus · 2026-10-24
- Kimi K3.1 turns heads with a detailed Ana de Armas SVG illustration — lyraxana · 2026-10-24
- NVIDIA releases a video anomaly detection model on Hugging Face — cocktailpeanut · 2026-10-24
- FLUX.3 Klein (5.6B) tested across 20 anime styles: still behind MOSS-Image, Krea 2 and Qwen-Image 2.1 — SkirtSwimming8950 · 2026-10-24
- Dreamina's $0.29 first month comes with 1,575 credits — but how far do they go? — eyishazyer · 2026-10-24