🎬 The Anime Scripter

Free, open-source AI video enhancement for anime and beyond.
Upscale, interpolate, restore and dedup in a single GPU pass from the CLI, a Windows desktop app, or directly inside After Effects.

720p master → Adore 2x. More demos

✨ Why TAS

• Anime-first models. Line-art-tuned upscalers (ShuffleCUGAN, Adore, Fallin, AniScale 2, Cyte), RIFE up to 4.26, anime depth (Limbo) and anime segmentation, plus general-purpose models for live action and games.
• One pass, in memory. Decode → dedup → restore → interpolate ↔ upscale → encode runs as one pipeline with no intermediate files. NVDEC decoding, TensorRT engines and CUDA graphs keep the GPU busy.
• Built for animation. --smooth_dedup detects frames held on twos and threes and interpolates across them while keeping the original length and audio. --scenechange holds the frame at hard cuts so nothing morphs between shots. --mask keeps HUDs and hardsubs from warping.
• Runs on most GPUs. CUDA, TensorRT, DirectML, OpenVINO, NCNN/Vulkan, Apple Silicon (MPS) and AMD ROCm on Linux.
• Works where you edit. A native After Effects panel, a standalone Windows app and a scriptable CLI with presets, batch input and YouTube URLs.

🧰 What it does

| Feature | Flag | Models | Backends | |---|---|---|---| | Upscale | --upscale | ShuffleCUGAN, Adore, Fallin Soft/Strong, Cyte, SPAN, Open-Proteus, AniScale 2, RTMoSR, Saryn, Gauss, SmoSR, FigSR, AnimeSR, ArtCNN, NVIDIA Maxine VSR, your own Spandrel/ONNX model | CUDA · TRT · DML · OpenVINO · NCNN · MPS · ROCm | | Interpolate | --interpolate | RIFE 4.6 – 4.26 (lite / heavy), RIFE Elexor, DistilDRBA, GMFSS Fortuna, NVIDIA Maxine Frame Generation | CUDA · TRT · DML · OpenVINO · NCNN · MPS · ROCm | | Restore (chainable) | --restore | Anime1080Fixer, SCUNet, NAFNet, DPIR, DeJPEG, DeH264, GaterV3, HurrDeblur, DeHalo, deepDeband, FastLineDarken, LineThinner, AutoCAS, Maxine denoise/deblur | CUDA · TRT · DML · OpenVINO · MPS · ROCm | | Dedup | --dedup / --smooth_dedup | SSIM, MSE, VMAF, FlowNetS | CPU · CUDA · ROCm | | Depth maps | --depth | Depth Anything V2 & 3, Video Depth Anything, Limbo V1/V2 (anime), DA3 streaming | CUDA · TRT · DML · OpenVINO · MPS · ROCm | | Segmentation | --segment | Anime Segmentation, BiRefNet (anime) | CUDA · TRT · DML · ROCm | | Object detection | --obj_detect | YOLOv9 (MIT) S / M / L | TRT · DML · OpenVINO | | Scene detection | --autoclip / --scenechange | PySceneDetect, TransNetV2, MaxxViT, SSIM/MSE | CPU · CUDA · TRT · DML · ROCm | | Stabilize | --stabilize | Classic (feature tracking), DUT (deep mesh warp) | CPU · CUDA · ROCm | | Motion blur | --moblur | RIFE-driven shutter simulation | CUDA · TRT · DML · MPS · ROCm |

Output goes through FFmpeg or TAS's in-process Nelux encoders: x264/x265/AV1/VP9, NVENC, QSV, AMF, ProRes (with alpha), lossless, GIF and PNG/JPEG sequences.

PARAMETERS.MD has every flag, every model with its backends, and recommendations by content type.

📦 Get TAS

| Edition | Platform | What it is | |---|---|---| | TAS-Standalone | Windows | Desktop app with its own backend: no Python, no terminal. Toggle steps into a chain, watch the live preview and save presets. | | TAS-AdobeEdition | Windows, macOS | After Effects 2022+ panel. Follow the installation guide or the video tutorial. | | CLI | Windows, macOS (Apple Silicon), Linux (source) | Every option, scriptable. Nightly builds carry the newest features. |

TAS-Standalone

After Effects panel

Models download on first use, so the first run with a new model takes longer.

Which backend fits your GPU

| Hardware | Use | |---|---| | NVIDIA RTX 20 – 50 / GTX 16 | CUDA (default) or -tensorrt for maximum speed | | NVIDIA GTX 10 and older | -directml | | AMD / Intel on Windows | -directml, or -openvino on Intel | | AMD on Linux | -rocm (experimental) | | Apple Silicon (M1+) | -mps | | Anything with Vulkan | -ncnn (RIFE, ShuffleCUGAN, SPAN) |

NVIDIA Maxine methods need an RTX card (Frame Generation needs an RTX 40 or 50).

🚀 Quick start

Windows CLI one-liner. It installs into a TheAnimeScripter folder in the current directory and asks whether to add that folder to PATH:

iwr -useb https://tas.nevermindnilas.dev/install.ps1 | iex

Then run it. Without the installer, use python main.py in place of tas:

# 2x upscale + 2x interpolation, anime defaults
tas --input episode.mp4 --upscale --interpolate
Picking a method enables its step: TensorRT upscale + RIFE 4.25, scaled to 4K
tas --input episode.mp4 --upscale_method shufflecugan-tensorrt --interpolate_method rife4.25-tensorrt --output_scale 3840x2160
Anime on twos → smooth motion, same length, cuts left untouched
tas --input episode.mp4 --smooth_dedup --scenechange
Chain restorers, encode 10-bit
tas --input old_show.mkv --restore_method deh264_real anime1080fixer --encode_method x265_10bit
Temporally stable depth map
tas --input clip.mp4 --depth_method video_small_v2

--input also takes folders, ;-separated lists, .txt batch files and URLs. Run tas --list_presets and tas --list_methods to explore, or tas -h for every option.

python -m pip install -r requirements.txt -r extra-requirements-windows.txt
python main.py -h

🙏 Credits

TAS builds on the work of many model authors and open-source projects. Third-party license notices are in THIRD_PARTY_NOTICES.md.

Collaborators: Trentonom0r3 (TAS Adobe Edition) · Adegerard (architecture & optimization suggestions)

Missing someone? Email [email protected] or open an issue.