Research map / Video / Video generation
FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers
Minguk Kang, Suha Kwak
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal refinement (TR) (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | rFVD: 121.87 -> 86.94 | Improved |
| add sliding-window causal attention (SW-CA) (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | rFVD: 117.77 -> 136.08 | Hurt |
| add grouped-query attention (GQA) (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | rFVD: 117.77 -> 121.87 | Hurt |
| add spatial upsampling (SU) (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | PSNR / rFVD: 31.05 -> 31.49 PSNR; 86.94 -> 96.19 rFVD | Mixed |
| scale model from 56.8M to 769.3M parameters (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | rFVD: 96.19 -> 44.74 | Improved |
| add streaming training with rolling KV cache (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | PSNR: 32.56 -> 37.52 | Improved |
| add adversarial training (ablation) | FlashDecoder · Component ablation (480p, 17 frames) | PSNR / rFVD: 37.52 -> 37.08 PSNR; 12.29 -> 10.77 rFVD | Mixed |
| increase rolling window from 2 to 4 frames (ablation) | FlashDecoder · Window size ablation (720p) | PSNR / FPS: 38.38 -> 38.49 PSNR; 76.3 -> 60.6 FPS | No clear effect |
| scale model from S to XL (ablation) | FlashDecoder · Model scaling (480p, Stage 1 only) | rFVD: 89.23 -> 31.00 | Improved |
| add inference optimizations (compile, CUDA graph, FP8) to FlashDecoder-XL | FlashDecoder · UltraVideo 720p (4x16x16 latent) | throughput (FPS): 76.3 -> 151.0 | Improved |
| FP8 quantization of MLP layers | FlashDecoder · UltraVideo 1080p (4x16x16 latent) | PSNR (dB): 41.55 -> 40.84 | Hurt |
| replace Wan2.2 conv decoder with FlashDecoder-XL | Wan2.2 decoder · UltraVideo 1080p (4x16x16 latent) | peak memory (GB): 41.0 -> 3.7 | Improved |
Closest papers by meaning
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.89)
- FlashFFTConv: Efficient Convolutions for Long Sequences with Tensor Cores (ICLR 2024 · similarity 0.86)
- Gradient-free Decoder Inversion in Latent Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.86)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.86)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.86)
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.
