Research map / Video / Video generation
Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Ultra Flash cascades a streaming low-res autoregressive video generator with a causal memory latent upsampler, an architecture-preserving T2V-to-TV2V super-resolution model, and a high-resolution decoder, reaching ~30 FPS at 1K and ~18 FPS at 2K on one GPU.
Problem
Autoregressive streaming video diffusion models are limited to low resolutions (e.g., 480P), and existing cascaded high-resolution methods are not causal, are slow, or sacrifice the base model's generative capability.
Approach
Repurpose a pretrained T2V DiT as a TV2V super-resolution model by zero-initialized channel extension, use a lightweight causal memory network for latent upsampling and HR decoding, and then distill the SR model to a single-step sparse causal generator with reward signals and cascaded DPO-based self-forcing.
Key results
- Ultra Flash reaches ~30 FPS at 960x1664 (1K) and ~18 FPS at 1440x2496 (2K) on a single B200 GPU, versus 32.0 FPS for Self Forcing at 480x832.
- On VBench total score, Ultra Flash achieves 84.17 versus 83.74 for Self Forcing and 84.04 for Causal Forcing (4 steps).
- Causal memory upsampler (8 blocks) improves CLIP-IQA+ from 0.625 (bicubic) to 0.671 with ~8 ms overhead.
Surprising findings
- The paper reports Ultra Flash at 1K outperforming a 50-step Wan2.1 teacher on temporal flickering, though this is only in the appendix.
Limitations
- Block-sparse kernel is optimized for H200/B200 GPUs; consumer GPUs less optimized.
- DPO positive samples depend on a stronger Wan2.2-5B SR model.
- Real-time generation is limited to 2K; 4K is out of reach.
Experimental setup
Datasets: HR video dataset (unspecified)
Benchmarks: VBench CLIP-IQA+ MUSIQ NIQE LPIPS
Metrics: VBench total score subject consistency background consistency motion smoothness imaging quality aesthetic quality FPS latency peak GPU memory NIQE
Baselines: Wan2.1 CausVid Self Forcing Causal Forcing DummyForcing FlashVSR
Builds on: Self Forcing CausVid Decoupled DMD Wan2.1 FlashVSR Direct Preference Optimization
Closest papers by meaning
- FlashVSR: Towards Real-time Diffusion-Based Streaming Video Super Resolution (CVPR 2026 · similarity 0.91)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.85)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.85)
- SkipSR: Faster Super-Resolution with Token Skipping (NeurIPS 2026 · similarity 0.84)
- Efficient Video Super-Resolution for Real-time Rendering with Decoupled G-buffer Guidance (CVPR 2025 · similarity 0.84)
- VideoGigaGAN: Towards Detail-rich Video Super-Resolution (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.