🔥 SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer

💡 The paper introduces SANA-Video, a small diffusion model designed for efficient video generation. The problem addressed is the high cost and slow speed of existing video generation models. To solve this, the authors propose two core designs: Linear DiT, which leverages linear attention as the core operation, and a constant-memory KV cache for block linear attention. This cache provides global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient long video generation.

The method used is a block-wise autoregressive approach for long video generation, which employs a constant-memory state derived from the cumulative properties of linear attention. The authors also explore effective data filters and model training strategies, which narrow the training cost to 12 days on 64 H100 GPUs, a significant reduction compared to other models.

The results show that SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models, while being 16 times faster in measured latency. The model can generate high-resolution, high-quality videos up to 720x1280 resolution and minute-length duration at a remarkably fast speed. Additionally, SANA-Video can be deployed on RTX 5090 GPUs, accelerating the inference speed of generating a 5-second 720p video from 71 seconds to 29 seconds, a 2.4 times speedup. Overall, SANA-Video enables low-cost, high-quality video generation, making it a significant contribution to the field of video generation.

📅 Published on Sep 29, 2025

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2509.24695
• PDF: https://arxiv.org/pdf/2509.24695
• Project Page: https://nvlabs.github.io/Sana/Video

🤖 Models citing this paper:
• https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_720p
• https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_480p
• https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_480p_diffusers

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/helenai/check-optimum-intel-support

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#VideoGenerationModels #DiffusionTransformer #BlockLinearAttention #EfficientVideoProcessing #AutoregressiveVideoGeneration

🔥 Packing Input Frame Context in Next-Frame Prediction Models for Video Generation

💡 The paper introduces FramePack, a neural network designed to improve video generation by enhancing next-frame prediction models. The main problem addressed is the limitation of transformer context length, which restricts the number of frames that can be processed. To overcome this, FramePack compresses input frames, allowing the transformer context length to be fixed regardless of the video length. This enables the processing of a large number of frames and increases the batch size, making it comparable to image diffusion training.

The method proposed by FramePack involves compressing input frames and using an anti-drifting sampling method to generate frames in inverted temporal order. This approach helps to avoid exposure bias, which occurs when errors accumulate over iterations. Additionally, FramePack can be used to fine-tune existing video diffusion models, allowing for more balanced diffusion schedulers with less extreme flow shift timesteps.

The results show that FramePack improves the visual quality of video generation by supporting more balanced diffusion schedulers. The increased batch size and improved frame prediction also enhance the overall performance of video diffusion models. Overall, FramePack provides a novel approach to video generation by addressing the limitations of transformer context length and improving the efficiency of next-frame prediction models.

📅 Published on Apr 17, 2025

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2504.12626
• PDF: https://arxiv.org/pdf/2504.12626
• Project Page: https://lllyasviel.github.io/frame_pack_gitpage/

🤖 Models citing this paper:
• https://huggingface.co/URWAIFU/framepack-eichi-f1

📊 Datasets citing this paper:
• https://huggingface.co/datasets/agreeupon/wrkspace-backup-ttl

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/linoyts/FramePack-F1
• https://huggingface.co/spaces/makululinux/FramePack-F1
• https://huggingface.co/spaces/ObiJuanCodenobi/VidGen-Emilio

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#VideoGenerationModels #NextFramePrediction #TransformerContextLength #FrameCompressionTechniques #NeuralNetworkArchitecture

🔥 LongCat-Video Technical Report

💡 The paper introduces LongCat-Video, a 13.6 billion parameter video generation model based on the Diffusion Transformer framework. The model is designed to generate high-quality long videos efficiently, which is a crucial step towards creating world models. LongCat-Video has a unified architecture that can perform multiple tasks, including text-to-video, image-to-video, and video continuation, using a single model.

The model achieves efficient long video generation through a coarse-to-fine generation strategy and block sparse attention, allowing it to generate 720p, 30fps videos within minutes. The coarse-to-fine generation strategy works by gradually increasing the resolution and detail of the video, both in terms of time and space. Block sparse attention is a technique that reduces the computational cost of the model by only attending to certain parts of the input data.

The model was trained using a multi-reward reinforcement learning from human feedback approach, which enables it to achieve performance comparable to state-of-the-art models. The use of multi-reward reinforcement learning from human feedback allows the model to learn from human evaluators and improve its performance over time.

The results show that LongCat-Video excels in generating high-quality long videos, maintaining temporal coherence and quality even in videos that are several minutes long. The model's efficiency and performance make it a significant contribution to the field of video generation, and the fact that the code and model weights are publicly available will accelerate progress in this area. Overall, LongCat-Video is a foundational model that takes an important step towards creating world models, which are complex models that can simulate and generate realistic videos and other types of data.

📅 Published on Oct 25, 2025

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2510.22200
• PDF: https://arxiv.org/pdf/2510.22200

🤖 Models citing this paper:
• https://huggingface.co/meituan-longcat/LongCat-Video
• https://huggingface.co/Nishant2414/LongCat-Video
• https://huggingface.co/fjkane/LongCat-Video-bf16

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/cpuai/LongCat-Video-Avatar
• https://huggingface.co/spaces/multimodalart/LongCat-Video
• https://huggingface.co/spaces/armaishere/meituan-longcat-LongCat-Video

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#VideoGenerationModels #DiffusionTransformer #LongVideoSynthesis #TextToVideoSynthesis #ImageToVideoGeneration

🔥 OSP-Next: Efficient High-Quality Video Generation with Sparse Sequence Parallelism, HiF8 Quantization, and Reinforcement Learning

💡 The paper introduces OSP-Next, an efficient text-to-video generation model that addresses the high computational costs of existing models. The problem with current models, such as Diffusion Transformers, is that they achieve strong video generation quality but have quadratic costs due to full attention. To solve this, OSP-Next combines sparse attention, parallelism, quantization, and reinforcement learning.

The method used in OSP-Next is a hybrid full-sparse attention architecture, where the sparse component is implemented with Skiparse-2D Attention. This mechanism applies token-wise and group-wise sparse attention along spatial dimensions, leveraging locality while maintaining compatibility with FlashAttention kernels. The authors also propose Sparse Sequence Parallelism, which partitions subsequences across ranks and switches sparse patterns through a single All-to-All communication. This approach reduces communication volume by 75% compared to Ulysses Sequence Parallelism.

Additionally, OSP-Next incorporates HiF8 quantization to enable stable joint training with 8-bit quantization and sparse fine-tuning. The model also applies Mix-GRPO post-training to improve the performance of the sparse model. The authors evaluate OSP-Next on various settings, including 5-second 720P and 5-second 768P, and achieve significant speedups on NVIDIA H200 GPUs and Ascend 950PR hardware.

The results show that OSP-Next achieves a VBench total score of 83.73%, surpassing the Wan2.1 baseline. The model achieves up to 1.64 times single-GPU speedup and over 1.52 times eight-GPU speedup on NVIDIA H200 GPUs. Furthermore, with only a 0.4% drop in VBench total score, OSP-Next-HiF8 achieves 1.69 times and 2.27 times speedups under the two settings on a single Ascend 950PR, demonstrating the efficiency and performance of OSP-Next across hardware platforms. Overall, the paper contributes to the development of efficient text-to-video generation models with high-quality video synthesis and reduced computational costs.

📅 Published on May 27

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2605.28691
• PDF: https://arxiv.org/pdf/2605.28691

🤖 Models citing this paper:
• https://huggingface.co/yunyangge/OSP-Next

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#VideoGenerationModels #SparseSequenceParallelism #HiF8Quantization #ReinforcementLearningForVideo #TextToVideoSynthesis

🔥 LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generation

💡 The paper LongLive-RAG addresses the challenge of generating long videos using autoregressive video diffusion models. The problem with existing methods is that they use sliding-window attention, which can lead to error accumulation and identity drift over time. This is because once the active window accumulates appearance errors, subsequent generations can only condition on this degraded trajectory and drift further away. To overcome this limitation, the authors propose a retrieval-augmented generation framework called LongLive-RAG.

In this framework, previously generated latents are treated as a dynamic and searchable history. At each new block, LongLive-RAG uses a query embedding to retrieve relevant historical latents, allowing the generator to condition on non-local context instead of only the recent window. This retrieval step adds only a small overhead relative to generation and helps reduce error accumulation.

To make retrieval more discriminative, the authors introduce the Window Temporal Delta Loss, which suppresses redundant local similarity and encourages embeddings to capture meaningful temporal changes. The LongLive-RAG framework is general and can be used with multiple autoregressive backbones and generation lengths.

The experiments show that LongLive-RAG improves long video quality and achieves the best average VBench-Long rank. The authors claim that LongLive-RAG is the first method to formulate self-generated latent history as content-addressable retrieval memory, making it a significant contribution to the field of long video generation. The code for LongLive-RAG is available, making it possible for others to build upon and extend this work. Overall, the paper presents a novel approach to long video generation that addresses the limitations of existing methods and achieves state-of-the-art results.

📅 Published on Jun 1

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.02553
• PDF: https://arxiv.org/pdf/2606.02553
• Project Page: http://longlive-rag.github.io/

🤖 Models citing this paper:
• https://huggingface.co/qixinhu11/LongLive-RAG

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#VideoGenerationModels #RetrievalAugmentedGeneration #LongVideoSynthesis #AutoregressiveVideoDiffusion #RetrievalAugmentedFrameworks