--- license: apache-2.0 ---

RVR: One-step Generative Streaming Real-time Video Restoration

RVR teaser

> **RVR** is the first generative video restoration model to reach **real-time 1080p streaming on a consumer-grade GPU** (β‰ˆ26 FPS on a single RTX 5090), sustains **31 FPS at QHD (2560Γ—1440)** and **14 FPS at 4K (3840Γ—2160)** on a single H100, and streams at resolutions where every compared diffusion-based VR baseline runs out of memory.

arXiv Project Page HuggingFace License

--- ## Updates - [2026/06] Release the inference code and pretrained weights πŸŽ‰ --- ## ✨ Highlights - **Mask-free shifted-window self-attention (MFSWA).** Each spatial window is **pre-gathered into a dense tensor**, so every attention call reduces to a single standard scaled-dot-product (SDPA) call β€” *no attention mask, cyclic shift, or padding ever enters the graph*. This gives a **1.62Γ— throughput gain over its full-attention teacher** at essentially identical quality, with **no dedicated sparse kernel**. - **Restoration-aware Autoencoder (ReAE).** A lightweight encoder–decoder jointly fine-tuned with the DiT in pixel space removes the heavy-3D-VAE / tiled-decoding bottleneck. - **Causal chunk-wise streaming.** A minimal causal protocol (no rolling KV cache, no overlapped DiT inference) bounds the temporal axis, confining the residual \(\mathcal{O}(N^2)\) cost to the spatial axes. --- ## πŸ“Š Results ### Efficiency at 2560Γ—1440 (single H100, causal streaming, 24 frames) | Metric | SeedVR2-3B (tile)| DOVE (tile)| FlashVSR-Tiny | **RVR (Ours)** | |---|:---:|:---:|:---:|:---:| | Avg. Time (s) ↓ | 17.320 | 27.615 | 2.493 | 0.766 | | FPS ↑ | 1.39 | 0.85 | 9.61 | 31.32 | | Peak Mem. (GB) ↓ | 35.35 | 59.24 | 34.35 | 38.01 | > At **3840Γ—2160**, every compared diffusion-based VR baseline **OOMs** on a single H100; RVR sustains **14 FPS**. ### Qualitative comparison RVR teaser --- ## πŸ›  Installation ```bash git clone https://github.com/Holiday/RVR.git cd RVR conda create -n rvr python=3.10 -y conda activate rvr # Install PyTorch matching your CUDA toolkit first, e.g. CUDA 12.4: pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu124 # Install RVR (editable) and its dependencies: pip install -e . ```
Hardware notes - **Server:** single H100-80G reproduces the QHD/4K numbers above. - **Consumer:** single RTX 5090 reaches β‰ˆ26 FPS at 1080p with the *same checkpoint* (default PyTorch SDPA path, bfloat16, causal chunk protocol). - No hardware-specific retraining or kernel rewrite is required on any platform.
--- ## πŸ—‚ Model Zoo | Model Name | Date | Backbone | Link | |---|---|---|---| | RVR | 2026.06 | Wan2.2-TI2V-5B | [πŸ€— HuggingFace](https://huggingface.co/H-oliday/RVR) | ```bash huggingface-cli download H-oliday/RVR --local-dir checkpoints/ ``` Expected checkpoint layout (the directory passed to `from_pretrained`): ``` checkpoints/ β”œβ”€β”€ reae.safetensors # Restoration-aware Autoencoder weights β”œβ”€β”€ prompt_embedding.safetensors# precomputed empty-prompt text embedding (key: "prompt_emb") └── transformer/ # diffusers-format DiT β”œβ”€β”€ config.json └── diffusion_pytorch_model.safetensors ``` --- ## πŸš€ Quick Start ### Python API ```python from rvr import RVRPipeline pipe = RVRPipeline.from_pretrained("checkpoints/").to("cuda", dtype="bfloat16") pipe.restore_video("low_quality.mp4", "restored.mp4", upscale=4) ``` `restore_video` also accepts an image folder as input and can write a PNG sequence (`png_save=True`). Tunable knobs: `clip_len` (MIDDLE chunk size, multiple of 4), `dit_overlap`, `fps`, `quality` (0–100, mapped to x265 CRF), `queue_size`. ### Streaming (causal, chunk by chunk, no future frames) ```python session = pipe.stream(clip_len=24, resolution=(1920, 1080)) for lq_chunk in read_chunks("low_quality.mp4", n=24): # lq_chunk: [T, H, W, 3] uint8 hq = session.step(lq_chunk) # [1, T', 3, H', W'] in [0, 1], or None if buffered if hq is not None: write(hq) tail = session.flush() # flush the final buffered frames ``` ### Command line ```bash python scripts/inference.py \ --input low_quality.mp4 \ --output restored.mp4 \ --checkpoint checkpoints/ \ --upscale 4 \ --clip-len 24 \ --dtype bfloat16 \ ``` Use `--png` to write a PNG sequence. --- ## πŸ“ Repository Structure ``` RVR/ β”œβ”€β”€ README.md β”œβ”€β”€ requirements.txt β”œβ”€β”€ setup.py β”œβ”€β”€ scripts/ β”‚ └── inference.py # CLI entry point (thin wrapper over RVRPipeline) └── rvr/ β”œβ”€β”€ __init__.py # exports RVRPipeline β”œβ”€β”€ pipeline.py # RVRPipeline: from_pretrained / to / restore_video / stream β”œβ”€β”€ runner.py # four-stage pipelined runner (reader β†’ H2D β†’ GPU β†’ writer) β”œβ”€β”€ io.py # frame reading, GPU preprocessing, mp4 / PNG writing β”œβ”€β”€ models/ β”‚ β”œβ”€β”€ reae.py # β˜… Restoration-aware Autoencoder β”‚ └── transformer.py # β˜… DiT + mask-free shifted-window self-attention └── streaming/ β”œβ”€β”€ chunk.py # fixed-size causal chunk protocol β”œβ”€β”€ tae.py # streaming autoencoder (causal boundary state) └── dit.py # one-step streaming DiT (fixed timestep, RoPE offsets) ``` --- ## 🎬 More Visual Results > Full-length restored clips (low-quality input β†’ RVR, played back to back). --- ## πŸ™ Acknowledgements RVR builds on [Wan2.2-TI2V-5B](https://github.com/Wan-Video), the lightweight autoencoder [TAEHV](https://github.com/madebyollin/taehv), and the [RealBasicVSR](https://github.com/ckkelvinchan/RealBasicVSR) degradation pipeline. We thank the authors of [SeedVR2](https://github.com/ByteDance-Seed/SeedVR), [DOVE](https://github.com/zhengchen1999/DOVE), and [FlashVSR](https://github.com/OpenImagingLab/FlashVSR) for releasing strong baselines, and the [UltraVideo](https://github.com/Tele-AI/UltraVideo) team for the training corpus. --- ## πŸ“œ License Released under the [Apache 2.0 License](LICENSE). The Wan2.2 backbone and any third-party weights remain subject to their original licenses.