NextGAN paper hits FID 1.36 one-step on ImageNet 256, scales to 14B video models
serrjoa · x · 2026-11-06
arXiv paper 2611.06270: conventional adversarial losses supervise critic values without directly specifying generator update gradients. NextGAN trains a scalar critic by regressing its input gradient onto feature-space displacements between real and generated samples at interpolated points in frozen representation spaces — a Flow Matching-like approach. Results: one-step FID 2.87 (pixel JiT-B) and 1.54 (latent DiT-XL/2) on ImageNet 256×256, 1.36 with the adversarial form; adapts pre-trained image/video generators up to 14B parameters without auxiliary losses or teacher scores; a single Wan2.1-1.3B model supports 1/2/4-step sampling, scoring 83.13 on VBench at two steps, beating rCM by 1.25. Code released.
More from Research
- University of Michigan's 26-lecture robotics math course (ROB 501) is free on YouTube — techNmak · 2026-11-06
- CDT+GT and indexical uncertainty in multi-agent clone settings: one policy network or many matters — jessi_cata · 2026-11-06
- Epoch AI researchers debate: inability to find slow-improving tasks is evidence of acceleration — Jsevillamol · 2026-11-06
- Mathematicians debate whether LLMs are filling the 'convex hull' of provable theorems — notmisha · 2026-11-06
- DeepMind paper fixes Softmax and cross-entropy gradient loss from fp rounding — jiqizhixin · 2026-11-06
- Santa Fe Institute chronicles the rise of agent-based modeling in social science — sebkrier · 2026-11-06