A team announced a distillation approach enabling instant-response, visually consistent live video streaming. They distilled a large 40-step diffusion teacher using 3-way CFG (120 evaluations per video chunk) into an unguided 2-step causal student model.
Yuesong, a researcher at Video Rebirth, open-sourced HyperFlow, a LoRA that distills MiniMax-H3's 49-step diffusion trajectory into 8 steps using data-free flow self-distillation, with the base model as its only teacher. It cuts generation time roughly 3x on H200 GPUs and reaches real-time speed on an 8-GPU B200 cluster when combined with other optimizations. Code, weights, and demo comparisons are public, under the MiniMax H3 Community License.
Yuesong, a researcher at Video Rebirth, open-sourced HyperFlow, a LoRA that distills MiniMax-H3's 49-step diffusion trajectory into 8 steps using data-free flow self-distillation, with the base model as its only teacher. It cuts generation time roughly 3x on H200 GPUs and reaches real-time speed on an 8-GPU B200 cluster when combined with other optimizations. Code, weights, and demo comparisons are public, under the MiniMax H3 Community License.
An arXiv paper presents LiveVVT, a rolling streaming diffusion framework for Video Virtual Try-On that runs in real time. It preserves bounded bidirectional modeling inside causal recurrent generation, avoiding the quality loss of naive causal enforcement while removing the latency of complete-clip diffusion. This matters for practitioners deploying try-on systems in continuous, live settings.
An arXiv paper presents LiveVVT, a rolling streaming diffusion framework for Video Virtual Try-On that runs in real time. It preserves bounded bidirectional modeling inside causal recurrent generation, avoiding the quality loss of naive causal enforcement while removing the latency of complete-clip diffusion. This matters for practitioners deploying try-on systems in continuous, live settings.
Researchers introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based method for distilling diffusion and flow models into few-step generators. PDD avoids the mode collapse and optimization difficulties of variational score distillation and adversarial losses, enabling faster image and video generation.
Researchers introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based method for distilling diffusion and flow models into few-step generators. PDD avoids the mode collapse and optimization difficulties of variational score distillation and adversarial losses, enabling faster image and video generation.
An arXiv paper introduces Video DeltaNet (VDN), a video-native hybrid attention architecture combining local Softmax attention with a linear-memory branch using Video Delta Attention. Instantiated on MiniMax H3, it denoises a 14.3-second 768p video in 6.70 seconds on eight B200 GPUs — a 14.5x speedup over the 50-step dense baseline — targeting livestream video generation.
An arXiv paper introduces Video DeltaNet (VDN), a video-native hybrid attention architecture combining local Softmax attention with a linear-memory branch using Video Delta Attention. Instantiated on MiniMax H3, it denoises a 14.3-second 768p video in 6.70 seconds on eight B200 GPUs — a 14.5x speedup over the 50-step dense baseline — targeting livestream video generation.
Researchers propose FlashDecoder, a pure-Transformer video decoder that replaces slow 3D convolutional decoders in latent video diffusion models. It decodes latents to pixels frame by frame using a rolling KV cache with a fixed temporal window, enabling constant-latency streaming regardless of video length.
Researchers propose FlashDecoder, a pure-Transformer video decoder that replaces slow 3D convolutional decoders in latent video diffusion models. It decodes latents to pixels frame by frame using a rolling KV cache with a fixed temporal window, enabling constant-latency streaming regardless of video length.