Tsinghua's MGFlow hits one-step SOTA: FDr6 cut 23-38% on ImageNet, extends to streaming video
jiqizhixin · x · 2026-10-20
Tsinghua University presents MGFlow and Elastic Forcing, two complementary works on sample-distribution training that let generative models match the data distribution at every denoising step — making one-step generation sharp enough for real-time interaction, with the same method extending to streaming video.
- MGFlow: post-trained pMF-H and JiT-H achieve one-step FDr⁶ of 1.45 and 1.64 on ImageNet 256×256, roughly 23% and 38% lower than FD-Loss; one-step generation against FLUX.2 [klein] 4B reaches GenEval 0.900 and PickScore 21.98.
- Elastic Forcing: scores 84.64 on VBench, significantly surpassing Self Forcing, and adapts to streaming real-time generation.
The project page ships a live demo, samples, code, and Hugging Face checkpoints.
More from Research
- Luma's MARBLE: exact sequence parallelism for packed multimodal masks, wins 47/48 vs FlashMask+Ulysses — mrjonfinger · 2026-10-20
- New theory paper argues SFT should reward any good answer, not imitate demonstrator distribution — burkov · 2026-10-20
- Steering both the "God" and "Owl" neurons makes the model proclaim "Lord of the Covenant" — Sauers_ · 2026-10-20
- Qubit estimates to break Bitcoin fall below 900 logical qubits, new Pareto frontier shows — markjeffrey · 2026-10-20
- Anthropic researcher corrects circuit tracing paper: completeness score isn't faithfulness — JacobHHilton · 2026-10-20
- KAIST unveils next-gen humanoid: 12km/h walking, moonwalk, eyes-closed balance — CyberRobooo · 2026-10-20