upscale
Runs video2x with the processor, model and scale you pick. RealESRGAN, Anime4K or RealCUGAN.
Self-hosted · video2x + FFmpeg
Queue a season,
walk away.
A dashboard for restoring anime in bulk. Point it at a folder, chain upscaling, interpolation and re-encoding into one pipeline, and let your own GPU work through the queue.
Source
1080p · 24 fps · H.264
1
Upscale
realesrgan
realesr-animevideov3
2×
2
Interpolate
rife-v4.6
2×
3
Optimize
libx265
CRF 19
tune animation
Output
1080p ·
24 fps ·
H.264
Operations
Four things it does to a file
Every job — and every step inside a pipeline — is one of these. Chain them in any order and the intermediate files stay on disk.
interpolate
Runs RIFE to multiply the frame rate. Turns 24 fps into 48 or 96 without the soap-opera artefacts of naive blending.
optimize
Re-encodes with the codec, CRF preset and audio settings you choose. Optional hardware encoding on NVIDIA, AMD or Intel.
integrity check
Probes the file with FFmpeg to confirm it decodes end to end, so a truncated download never reaches the queue.
What you get
Built for a media box, not a laptop
It assumes you have a GPU, a lot of episodes, and no interest in paying per minute of video.
- Live log streaming per job, with per-step progress and ETA
- Reusable pipelines saved as named, ordered step chains
- File explorer with natural-sort episode ordering
- GPU health monitor with NVIDIA driver wedge detection
- Configurable GPU streams and FFmpeg worker count per host
- Portainer-friendly stack file for one-click home server deploys
- Single-password login gate in front of the dashboard
- Everything on your machine — no cloud, no telemetry
The dashboard
Dense where it counts
Models
Pick a processor, then a model
Only some model and scale combinations are valid; the UI enforces the rules so you cannot queue a job that will fail an hour in.
| Processor | Scales | Best for |
|---|---|---|
realesrgan |
2× · 3× · 4× | Pure neural upscaler. Best raw quality for anime video, and the highest VRAM cost. |
libplacebo (Anime4K v4) |
2× · 3× · 4× | GPU shaders. Very fast and light enough to run a whole season overnight beside other workloads. |
realcugan |
2× · 3× · 4× | Middle ground. Strong on stylised content, with the most control over denoise behaviour. |
rife |
frame rate |
Every RIFE generation through rife-v4.26.
rife-v4.6 is a sensible default for modern
anime.
|
Codecs: libx265, libx264, libvpx-vp9 or stream copy, with quality presets from CRF 16 to CRF 26 and the standard x264/x265 speed ladder.
Quick start
Running in one command
Generates a strong secret and a random password, creates the media folders, and starts the stack from prebuilt images. Nothing to build.
CPU · prebuilt images
git clone https://github.com/IvanMicai/anime-upscaling.git
cd anime-upscaling
make quickstart
It prints the generated password when it finishes. Open
http://localhost:4750 and log in with it. Re-running
make quickstart never overwrites a password you
already set.
NVIDIA GPU · built from source
cp .env.example .env
# set AUTH_PASSWORD and AUTH_SECRET first
make run-gpu
Needs the NVIDIA Container Toolkit on the host. Drop source
videos in data/input; results land in
data/output, data/optimized or
data/interpolated depending on the job.
Before you expose it: publish the web port only — the API is meant to stay on the internal Compose network. The login is a single shared password, so put the service behind HTTPS, a VPN or a trusted reverse proxy if it is reachable outside a private network.