SeedVR2 VAE TensorRT Engines for NVIDIA Blackwell
High-performance, pre-compiled NVIDIA TensorRT engine plans (.rtxplan) built for the official SeedVR2 VAE (seedvr2_ema_vae_fp16), specifically compiled and optimized for the NVIDIA Blackwell GPU architecture (RTX 5090, RTX 5080, RTX 5070, RTX 5060, RTX 5050, and Blackwell workstation / data center GPUs, Compute Capability sm_100 / sm_120).
🌟 Overview
SeedVR2 is a state-of-the-art diffusion-based video restoration and super-resolution model developed by ByteDance. The video VAE component (seedvr2_ema_vae_fp16) handles spatial-temporal latent encoding and decoding, which represents a primary computational and memory bottleneck during video upscaling and restoration workflows.
This repository provides pre-compiled TensorRT execution engine plans (.rtxplan) optimized for high-throughput, low-latency tiled spatial-temporal inference on NVIDIA Blackwell hardware:
- Hardware Target: NVIDIA Blackwell architecture (
sm_100/sm_120, RTX 50-Series: RTX 5090, RTX 5080, RTX 5070, RTX 5060, RTX 5050). - Spatial Configuration: Optimized for deterministic 256×256 spatial tiling (
tile_256) to maximize L2 cache utilization, minimize VRAM spikes, and prevent out-of-memory errors on 4K/8K video upscaling. - Temporal Configuration: Dedicated engine profiles for fine-grained frame batches across both encoder (53f–89f) and decoder (21f–61f) pipelines.
- Base Checkpoint: Comfy-Org/SeedVR2 (seedvr2_ema_vae_fp16) / ByteDance SeedVR2.
📦 Available TensorRT Engine Plans
🔄 Decoder Engines (vae_decoder_tile_256_*f.rtxplan)
Optimized for decoding compressed spatial-temporal latents back to RGB video frame sequences.
| Engine Filename | Module | Tile Size | Temporal Frames | File Size | Target Architecture |
|---|---|---|---|---|---|
vae_decoder_tile_256_21f.rtxplan |
VAE Decoder | 256×256 | 21 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_25f.rtxplan |
VAE Decoder | 256×256 | 25 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_29f.rtxplan |
VAE Decoder | 256×256 | 29 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_33f.rtxplan |
VAE Decoder | 256×256 | 33 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_41f.rtxplan |
VAE Decoder | 256×256 | 41 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_45f.rtxplan |
VAE Decoder | 256×256 | 45 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_49f.rtxplan |
VAE Decoder | 256×256 | 49 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_53f.rtxplan |
VAE Decoder | 256×256 | 53 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_57f.rtxplan |
VAE Decoder | 256×256 | 57 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_decoder_tile_256_61f.rtxplan |
VAE Decoder | 256×256 | 61 frames | ~308 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
📥 Encoder Engines (vae_encoder_*f_tile256.rtxplan)
Optimized for encoding input RGB video frame sequences into SeedVR2 spatial-temporal latents.
| Engine Filename | Module | Tile Size | Temporal Frames | File Size | Target Architecture |
|---|---|---|---|---|---|
vae_encoder_53f_tile256.rtxplan |
VAE Encoder | 256×256 | 53 frames | ~214 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_57f_tile256.rtxplan |
VAE Encoder | 256×256 | 57 frames | ~215 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_61f_tile256.rtxplan |
VAE Encoder | 256×256 | 61 frames | ~215 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_65f_tile256.rtxplan |
VAE Encoder | 256×256 | 65 frames | ~216 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_69f_tile256.rtxplan |
VAE Encoder | 256×256 | 69 frames | ~216 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_73f_tile256.rtxplan |
VAE Encoder | 256×256 | 73 frames | ~217 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_77f_tile256.rtxplan |
VAE Encoder | 256×256 | 77 frames | ~217 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_81f_tile256.rtxplan |
VAE Encoder | 256×256 | 81 frames | ~218 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_85f_tile256.rtxplan |
VAE Encoder | 256×256 | 85 frames | ~218 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
vae_encoder_89f_tile256.rtxplan |
VAE Encoder | 256×256 | 89 frames | ~218 MB | NVIDIA Blackwell (SM100/SM120 / RTX 5090/5080/5070/5060/5050) |
🛠️ Performance & Architectural Advantages
Native Blackwell Kernel Tuning: Built and tuned specifically with the TensorRT Blackwell compilation pipeline, fully utilizing Blackwell 5th-Generation Tensor Cores, enhanced memory bandwidth, and fused convolution-activation operators across the entire RTX 50-Series lineup (RTX 5090, RTX 5080, RTX 5070, RTX 5060, RTX 5050).
Deterministic Tiled Processing (Tile 256): Fixed spatial tiling at 256×256 ensures bounded VRAM consumption regardless of full video resolution (1080p, 4K, 8K), eliminating Out-of-Memory (OOM) failures and enabling continuous batching.
Substantial Latency Reduction: Provides significant speedups over standard PyTorch FP16 eager and
torch.compileexecution, removing VAE encode/decode bottlenecks in iterative video generation and upscaling pipelines.
📜 Credits & References
- SeedVR / SeedVR2 Foundation: ByteDance Seed Team
- Official VAE Repository: Comfy-Org/SeedVR2 (
seedvr2_ema_vae_fp16.safetensors) - Acceleration Framework: NVIDIA TensorRT
- License: Apache-2.0 License
- Downloads last month
- -