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

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!

AP_T2 is an interactive video generation model!. Users control the movements for real-time interactive world exploration.

APT2 autoregressively generate a single latent frame at a time in 1NFE. The model fully exploits KV cache for nonstop streaming up to a minute long (1440 frames). Our 8B model runs 640x480 24fps real-time on 1xH100, or 1280x720 24fps real-time on 8xH800!

RT @Iceclearwjy: 🔥Introducing #SeedVR2, the latest one-step diffusion transformer version of #SeedVR for real-world image and video restora….

RT @op7418: 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 https://t.c….

RT @CeyuanY: Glad to share Seaweed-7B, a cost-effective foundation model for video generation. Our tech report highlights the key designs t….

Excited to announce our latest research Seaweed-APT. Unlike diffusion and autoregressive models that require iterative sampling, our model generates 1280x720 24fps videos using only a single step (1NFE)!. For more information, visit our website:.

AnimateDiff-Lightning Gradio demo is out!

AnimateDiff-Lightning is incredible for video-to-video generation. Our model is released. Excited to see what the community will create with it.

Excited to present AnimateDiff-Lightning for lightning-fast video generation. Our model is more than ten times faster than the original AnimateDiff, achieving new SOTA on few-step video generation.

RT @gorkemyurt: is currently in the front page of HN and we are handing 50 new websocket connections per second, th….

RT @modal_labs: 🥔⚡️ - SDXL Lightning as fast as you can type

Excited to announce SDXL-Lightning, a lightning fast 1024px text-to-image generation model! Our model achieves new SOTA on one-step/few-steps generation! It is significantly better than prior methods. Try it out and check out our technical report!

Excited to announce our new paper: Diffusion Model with Perceptual Loss Our work is the first to train diffusion models with perceptual loss. It improves sample quality without CFG. Works for both conditional and unconditional models.

Common diffusion noise schedules and sample steps are flawed! Stable Diffusion has a flawed schedule that limits the images to be plain. We fixed it so it can generate more cinematic results! The findings are applicable to all diffusion models.

Announcing Background Matting V2. A real-time high-resolution background matting method. Our method achieves HD@60FPS and 4K@30FPS. Check out our project website for detail!: