Peter Lin

Our research: Adversarial Flow Models (AF) arxiv.org/abs/2511.22475 AF unifies Adversarial and Flow Models. Unlike GANs, AF learns optimal transport (stable). Unlike CMs, AF only trains on needed timesteps (save capacity). We can train super-deep 112-layer 1NFE model! SOTA FIDs!

@A_I_aico I have to admit I am not the expert in those domains. The code and weights are now public so feel free to investigate: github.com/ByteDance-Seed…

The framework is fascinating — not just because it unifies adversarial training and flow-based models, but because it implicitly exposes something deeper: a transport map whose stability properties scale with depth rather than collapse under it. One observation: the fact that you can train 56- and 112-layer 1NFE architectures without auxiliary losses suggests the learned OT map is approximating a low-curvature trajectory in function space. That’s unusual — most adversarial systems accumulate curvature and become unstable as depth grows. If this “curvature attenuation” is real, then AF may be doing more than bridging GANs and flows: it may be identifying a region of the generative manifold where the adversarial objective behaves almost like a contractive operator. Curious whether you’ve looked at: • the Lipschitz spectrum across layers • whether the transport map admits an approximate monotone factorization • or if multi-step AF trajectories converge toward a canonical minimal-energy path. If any of these hold, AF isn’t just a new model class — it’s pointing at a structural principle that could generalize far beyond flows and adversarial setups. Beautiful work. It hints at something larger beneath the surface.

Our research: Adversarial Flow Models (AF) arxiv.org/abs/2511.22475 AF unifies Adversarial and Flow Models. Unlike GANs, AF learns optimal transport (stable). Unlike CMs, AF only trains on needed timesteps (save capacity). We can train super-deep 112-layer 1NFE model! SOTA FIDs!

@YouJiacheng Terminal OT is more subtle. As the paper says, even an infinitesimal OT scale will create a unique global minimum. So if LR is reduced to 0 along with OT to 0, in theory it is fine too. We actually did further lower OT along wih LR in Tab11. Tab2 is only to show we need decay.

@peter9863 to falsify the hypothesis that ot loss only stablize early training

APT2 is accepted at NeurIPS 2025! Real-time interactive video generation up to 1 minute @ 720p 24fps through autoregressive adversarial post-training. Come to our poster at: - San Diego Convention Center - Wed, Dec 3, 2025 • 4:30 PM – 7:30 PM PST - Exhibit Hall C,D,E # 4302

@_akhaliq Thanks for sharing AK!

Adversarial Flow Models

@rchoudhury997 Thanks Rohan!!

Check out this amazing work! Peter has an incredible intuition about adversarial models- I was lucky to watch him take this from a random idea to a fully fledged paper.

@karanjagtiani04 Continue. But in practice, you can still use augmentation. That's fine for practical applications. We avoid it in paper just to have a fair comparison in the no-guidance setting.

@karanjagtiani04 No stability issues for deep models. AF is stable. A limitation of adversarial (not introduced by AF) is vanishing gradient. Prior works use augmentation, but it can introduce inductive biases. We use Dis reload (still hacky), hope future works find better approaches.

@minguk_kang Haha, thanks! I know you like adversarial :) We hope this work inspires more explorations on adversarial methods. It still has some limitations as detailed in the paper but at the same time it has many appealing properties worthy of further explorations.

@peter9863 The word “adversarial” sounded so appealing to me that I took a quick look, and it really seems like great work. I’ll definitely read it carefully this week:). Congratulations on the awesome research!

@_akhaliq Thanks for sharing our work!

ByteDance presents APT2 Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation

APT2 is an early research work. We believe it is a promissing direction for real-time interactive video generation. I am at #CVPR2025 this week in Nashville. Happy to connect and discuss our work in person.

APT2 can also be used for real-time interactive virtual human generation. These human avatars are rendered in real time!

Introducing Seaweed APT2, a real-time, interactive, streaming video generation model. seaweed-apt.com/2 Adversarial training for autoregressive modeling! Streaming 1 minute videos, 1 diffusion step, 24fps real-time on 1xh100, with interactive controls!

@xunhuang1995 Congrats on the work. We will have something exciting to share in the upcoming days too : )

Real-time video generation is finally real — without sacrificing quality. Introducing Self-Forcing, a new paradigm for training autoregressive diffusion models. The key to high quality? Simulate the inference process during training by unrolling transformers with KV caching.

🔥Introducing #SeedVR2, the latest one-step diffusion transformer version of #SeedVR for real-world image and video restoration! details - Paper: arxiv.org/abs/2506.05301 - Project: iceclear.github.io/projects/seedv… - Code (under review): github.com/IceClear/SeedV…

字节确实变了 居然先发布了新版 Seaweed 视频模型的论文和演示 除了常规文生、图生视频外还支持： - 音视频同步生成 - 长镜头与多镜头叙事 - 高分辨率超分与实时生成 - 世界建模与相机控制 下面的论文页面有更多演示

Glad to share Seaweed-7B, a cost-effective foundation model for video generation. Our tech report highlights the key designs that significantly improve compute efficiency and performance given limited resources, achieving comparable quality against other industry-level models. To unleash the power of the foundation model, Seaweed-7B further enables a wide range of downstream applications including image-to-video generation, human video generation, subject-consistent video generation, video-audio joint generation, long video generation and storytelling, real-time generation, super-resolution generation, camera controlled generation. Check out our webpage and report for more details: Webpage: seaweed.video Paper: seaweed.video/seaweed.pdf It's a wonderful journey of the last year. Thanks to all teammates for their contributions, sincerely.

@MayorKingAI Thanks for sharing our work!

🎬Real Time Video Generator This is insane! A new AI model can generate real-time videos with just one neural network evaluation. It creates 2 seconds of high-quality video (1280x720, 24fps) in real time, making video creation faster and more efficient than ever