2-Step DMD2 LoRA Distills Qwen-Image-2.1 From 20 Steps for Realtime Editing
linoy_tsaban · x · 2026-10-26
- linoytsaban 发布 Qwen-Image-2.1-DMD2-2steps:一个将 Qwen-Image-2.1 从 20 步采样压缩至仅 2 步前向传递(fixed sigmas [1.0, 0.5], truecfgscale=1.0)的蒸馏 LoRA,实现了接近实时的交互式编辑。
方法与训练
- 使用 DMD2(on-policy distribution matching,no regression/GAN loss)进行训练,采用 teacher CFG guidance
- LoRA 配置:rank 32 / alpha 32,仅作用于注意力层(attention-only)
- 在单张 H100 上训练约 2000 步,分辨率约 1 MP,支持多长宽比(multi-aspect)
质量
- 与官方 20 步演示在相同 seed 下进行对比(sketch-to-photo、发型、表情、照片修复、多参考构图):2 步的输出结果与基础模型非常接近
- 保留了多参考输入功能("Picture 1"/"Picture 2" 样式的 prompt)
使用方法
- 需要 diffusers 开发分支(pip install git+...@qwenimage21-lora-training),并使用 QwenImage21Pipeline
- 在作者占位符基准测试(placeholder benchmark)中提供了实时演示的 Space 和权重链接
More from Multimodal
- ElevenLabs Chimera pairs Eleven v4 audio with Seedance 2.5 video — Aiden_Tech_Ai · 2026-10-26
- Local book-to-audiobook-video pipeline runs on 8GB VRAM with Qwen Image + TTS — top115 · 2026-10-26
- 80s-style Flock Safety parody ad made locally with ComfyUI + MiniMax H3 on a 5090 — Effective_Garbage_34 · 2026-10-26
- One composition, three palettes: AI-generated stylized landscape backdrops impress — AIandDesign · 2026-10-26
- AI series 'The Chiseled and the Beautiful' returns with uncanny face-morph video — Darri3D · 2026-10-26
- Dotey shares full prompt + open-source skill for generating AI history documentary videos — dotey · 2026-10-26