iDDM trains few-step diffusion from scratch in one stage: 4.48 FID at 4 steps on ImageNet-256
proceduralia · x · 2026-10-02
- Method: iDDM enables few-step diffusion without a teacher model, self-distillation, or JVPs — a single model trained from scratch in one stage.
- Results: From the same checkpoint, it achieves 4.48 FID at 4 sampling steps and 2.38 at 50 steps on ImageNet-256.
- Takeaway: Removing the usual distillation machinery simplifies training pipelines for few-step diffusion; details are laid out in a thread.
Related event: CompVis Releases iDDM for Few-Step Diffusion in Single-Stage Training(2 posts)→
More from Research
- Structure-free AI drug discovery model Ptarmigan-1 opens for public use — ycombinator · 2026-10-02
- Qualcomm's IVD benchmark shows VLMs lag humans badly at real-time face-to-face Q&A — rishit_dagli · 2026-10-02
- Michigan Researchers Distill T5Gemma-2 Embeddings Into a More Diffusible Latent Space, Beating GPT-2-M — umich · 2026-10-02
- GraphForge: Evidence-Graph Workspace Synthesis Lifts GDPVal +65.7 With Only 2,169 Trajectories — ustc-community · 2026-10-02
- CMU's Predictive Credit Protocol Finds No Confirmed Gains From Research-Agent Explanations Across 336 States — CarnegieMellonU · 2026-10-02
- FloWright Co-Evolves Multi-Agent Workflows, Boosting Small Models by up to 7.41% — Xuehang Guo · 2026-10-02