VRGDG-SeedVR2-TensorRT-Studio README

The project's README from github.com/vrgamegirl19/VRGDG-SeedVR2-TensorRT-Studio. As one markdown file: README.md.

VRGDG SeedVR2 TensorRT Studio

Local, GPU-accelerated video restoration and upscaling with SeedVR2, TensorRT, and a purpose-built browser interface.

VRGDG SeedVR2 TensorRT Studio turns the SeedVR2 pipeline into a practical Windows workflow: load a video, test a short preview, compare the result frame by frame, and complete long renders with resumable checkpoints. Processing stays on your machine.

VRGDG SeedVR2 TensorRT Studio full interface

Highlights

Before and after examples

Each recording demonstrates the Studio's interactive comparison view, with the original footage on one side of the wipe and the restored result on the other.

Requirements

Recommended setup
Operating system Windows 11
GPU NVIDIA RTX with a current driver
Storage At least 35 GB free
Network Internet connection for the first installation
Python Installed automatically when needed; the project requires Python 3.12+

TensorRT engines are built locally for the installed GPU and runtime. Do not copy .rtxplan files between different GPU models or TensorRT versions.

Quick start

  1. Download or clone this repository.
  2. Double-click Install SeedVR Studio.bat and leave the window open until setup completes.
  3. Double-click Launch SeedVR Studio Pro.bat.
  4. Drop a source video into the Media panel.
  5. Render a short preview, inspect it, and then render the full video.

The installer prepares the private Python environment, FFmpeg, CUDA PyTorch, SeedVR2, SageAttention 2, TensorRT RTX, model weights, and GPU-specific engines. Downloads and engine preparation are resumable. The Pro launcher also detects an incomplete setup and opens the installer automatically.

For repair options and troubleshooting, see the installation guide.

How to update

Update from inside Studio

If the update button (↻) is visible beside Exit Studio:

  1. Finish or stop any active render.
  2. Click ↻ to check the main update channel.
  3. If an update is available, click Update and restart.

Studio closes, downloads only the changed application files, refreshes Python dependencies only when required, and then restarts automatically. Models, GPU-specific TensorRT engines, .venv, outputs, manifests, logs, and saved settings remain in place.

The safe updater requires a clean Git-backed installation on the main branch. It refuses to overwrite tracked local edits or diverged Git history. If dependency repair fails after an update, it restores the previous application revision automatically.

One-time update for existing installations

Users who do not yet see the ↻ button can install the update system with one final manual update:

  1. Download Update SeedVR Studio.bat.
  2. Place the downloaded BAT beside Launch SeedVR Studio Pro.bat in the existing Studio folder.
  3. Completely close SeedVR Studio.
  4. Double-click Update SeedVR Studio.bat.

The one-time updater downloads the current application code, creates an application-code backup under outputs\, and restarts Studio. It preserves the existing .venv, downloaded models, TensorRT engines, rendered videos, browser-saved settings, and other untracked user files.

For installations originally downloaded as a ZIP, this step also converts the existing folder into a clean Git-backed main installation. After it finishes, future updates use the ↻ button inside Studio.

Workflow

1. Load and inspect media

Drop a video onto the Media panel or browse for a file. After a preview or full render completes, use the viewer tabs to switch between the source and restored result.

Load a source Choose a viewer mode
Media panel with the original-video drop zone Original, Restored, Compare, and Side by side viewer tabs

The shared viewer also supports timeline scrubbing, frame stepping, synchronized playback, fullscreen, zoom, and pan.

2. Choose the render path and target

Start with SeedVR2 + TensorRT for the accelerated path, then select a hardware preset and output target. Presets provide sensible starting values; Custom / manual leaves every setting under your control.

Render engine Hardware preset Output and aspect
SeedVR2 and TensorRT render-engine selector Hardware preset selector Output-size and crop-policy selectors

Output presets include original-size enhancement, 1K/1080p, 2K/1440p, and 4K/2160p. Preserve the source aspect ratio or center-crop it to 16:9.

3. Tune restoration

Choose the model, temporal batch, seed, and color-matching strategy. FP8 models use less VRAM; FP16 models prioritize quality and require more memory. Temporal batches follow SeedVR2's 4n+1 rule, while the TensorRT path currently supports batches 5 and 21.

Model, batch, and seed Color correction Resumable rendering
Model, temporal-batch, and seed controls Color-correction selector and available modes Resumable long-video render and chunk-length controls

Color correction is applied during the main restoration. Use none to retain source colors; lab is a useful general-purpose matching mode. The other available modes are wavelet, wavelet_adaptive, hsv, and adain.

For long videos, enable resumable rendering. Auto selects a chunk length from the duration, frame count, frame rate, and temporal batch; manual choices range from 30 seconds to 30 minutes. Completed chunks remain available if a job stops, and Retry from last chunk continues at the first unfinished section.

4. Balance speed, memory, and finishing

TensorRT uses Optimized Fast decoding automatically. If it fails, the Studio retries the saved latents with its internal Stable decoder, avoiding another restoration pass.

Performance Post-processing Skin finishing
Attention, block swapping, and VAE-tiling controls Sharpening, film grain, and seam-smoothing controls Optional skin-finishing controls

Post-processing can be changed later with Reprocess post only, without rerunning SeedVR2.

5. Preview, render, and reopen results

Set a short preview range, render it, and compare the result before starting the complete video. The status bar reports readiness, elapsed time, and ETA while a job runs.

Preview range Render actions
Preview start and length fields Reprocess, preview, and full-render buttons
Render status Previous results
Ready state, elapsed time, ETA, power, and fullscreen controls Previous-results selector and load button

Every job receives its own folder under outputs\ with rendered media, manifests, and logs. Use Load selected output to reopen a result from the current workspace.

Render engines

Engine Best for Temporal batches Notes
SeedVR2 + TensorRT Recommended accelerated workflow 5, 21 Uses TensorRT VAE decoding and automatic fallback from saved latents.
SeedVR2 (Legacy) Compatibility and additional batch choices SeedVR2-compatible values Standard PyTorch path; can be substantially slower on long videos.

The temporal-batch menu updates automatically when the render engine changes.

Settings that matter most

Files and local data

Model weights live in models\ and are not committed to GitHub. Outputs, virtual environments, TensorRT engines and caches, temporary ONNX exports, and optional runtime files also remain local.

Path Purpose
web/ JavaScript Studio interface
api_server.py FastAPI service
seedvr_studio/ Rendering pipeline and legacy Gradio integration
tools/ TensorRT, assembly, post-processing, and diagnostic utilities
tensorrt_backend/ Optional native TensorRT sources
scripts/ Setup, launcher, download, verification, and engine-preparation scripts
vendor/seedvr2/ Compatible Apache-2.0 SeedVR2 integration required by the Studio

Documentation

License

The Studio source is released under the MIT License. SeedVR2 and bundled runtime components retain their upstream licenses; see the third-party notices.

VRGDG-SeedVR2-TensorRT-Studio (vrgamegirl19/VRGDG-SeedVR2-TensorRT-Studio)

Searching Database…

167 ·GitHub
VRGDG-SeedVR2-TensorRT-Studio
Что это?Локальная студия восстановления и апскейла видео с ускорением на GPU, построенная вокруг модели SeedVR2 с ускорением TensorRT и управляемая через браузерный интерфейс на JavaScript/FastAPI для предпросмотра, сравнения и рендеринга видео.
Это настольное приложение или веб-сервис, к которому я подключаюсь?Это локальное приложение, которое вы запускаете на своём ПК с Windows: бэкенд на FastAPI (apiserver.py) раздаёт JavaScript-фронтенд в вашем браузере, а вся обработка происходит на вашей машине, а не на удалённом сервере.
Как это запустить?Скачайте или склонируйте репозиторий, дважды щёлкните 'Install SeedVR Studio.bat' и дождитесь окончания установки (она ставит Python, FFmpeg, CUDA PyTorch, SeedVR2, SageAttention 2 и движки TensorRT), затем дважды щёлкните 'Launch SeedVR Studio Pro.bat' и перетащите видео в панель Media.
Какое железо нужно, прежде чем пробовать?GPU NVIDIA RTX с актуальным драйвером, Windows 11, не менее 35 ГБ свободного места и интернет-соединение для первой установки; Python 3.12+ установочный скрипт ставит автоматически.
Есть ли подводные камни с движками TensorRT?Движки TensorRT компилируются под конкретно установленную GPU и рантайм, так что файлы .rtxplan нельзя копировать между разными моделями GPU или версиями TensorRT; финишная обработка кожи тоже не генеративна и не может воссоздать черты лица, потерянные при восстановлении.
Это работает и на GPU AMD или только на NVIDIA?Для пути с ускорением TensorRT поддерживаются только GPU NVIDIA RTX; рекомендованная конфигурация требует Windows 11 и карту NVIDIA RTX, о поддержке GPU AMD ничего не говорится.
Можно ли обновиться, не потеряв модели и готовые рендеры?Да — и встроенный обновлятор, и одноразовый ручной обновляют только код приложения и зависимости, не трогая модели, движки TensorRT, .venv, результаты, манифесты и настройки.
Что будет, если долгий рендер прервётся?Возобновляемый рендеринг сохраняет готовые фрагменты, а «Повторить с последнего фрагмента» продолжает с первой незавершённой части вместо перезапуска всей задачи.
Насколько быстрый апскейл на практике?Для ориентира: 8-секундный клип апскейлился до 2K примерно 8 минут на самой большой модели 7B Sharp FP16 на RTX 5090; реальное время зависит от разрешения, частоты кадров и доступной VRAM.
Что на самом деле меняет финишная обработка кожи?Это негенеративный постпроцесс, использующий стабилизированную маску кожи для уточнения существующих пикселей, поэтому он сглаживает кожу, но не может придумать черты лица, потерянные при восстановлении.
What are similar projects?dlss5-visual-enhancer, ALH-Pro, video2dlssnr, DLSS5-Autopilot, NeuralScreen, Waifu2x-Extension-GUI, MoE4All, palette-from-image-ai
Which topics is it in?video-upscaling, resumable, rtx, upscaling, windows-11, fastapi, video-processing, gpu-acceleration, windows-app, offline, nvidia, cuda
Creator
vrgamegirl19
Size
40.3 MB
Stars
167
Updated
CommitSep 05
Age
1mo
Commits · 90 Days