NevermindNilas/TheAnimeScripter

The Anime Scripter (TAS) – Quick Overview

What it is – An open‑source toolkit that lets you upscale, interpolate, restore, deduplicate, and otherwise enhance video, with a focus on anime. It can be run from a command‑line interface, a native Windows desktop app, or as an After Effects panel, and it supports a huge range of AI models for each step.

Key capabilities

Step	What it does	Example models (AI)	Compute back‑ends
Upscale	Increases resolution while preserving line‑art detail	ShuffleCUGAN, Adore, AniScale 2, SPAN, etc.	CUDA, TensorRT, DirectML, OpenVINO, NCNN/Vulkan, Apple MPS, ROCm
Interpolate	Generates intermediate frames for smoother motion	RIFE 4.6‑4.26, RIFE‑Elexor, GMFSS Fortuna, NVIDIA Maxine Frame Generation	Same as above
Restore	Denoises, de‑blocks, de‑JPEG/H.264, de‑halo, etc.	SCUNet, NAFNet, DPIR, DeJPEG, GaterV3, Maxine denoise	Same as above
Dedup / Smooth dedup	Detects duplicate frames (e.g., anime on‑twos) and optionally interpolates across them	SSIM, MSE, VMAF, FlowNetS	CPU, CUDA, ROCm
Depth & Segmentation	Produces per‑pixel depth or masks for anime scenes	Depth Anything V2/V3, Limbo, Anime Segmentation, BiRefNet	Same as above
Object / Scene detection, Stabilisation, Motion blur	Adds classic or deep‑learning based effects	YOLOv9, PySceneDetect, DUT, RIFE‑driven shutter	Various

All steps are chained in a single GPU pass – decode → dedup → restore → interpolate → upscale → encode – so no intermediate files are written to disk. The pipeline can use hardware‑accelerated decoding (NVDEC) and encoding (NVENC, QSV, AMF, etc.) and falls back to CPU where needed.

How you use it

• CLI – tas --input episode.mp4 --upscale --interpolate (default anime‑friendly models). Flags let you pick specific methods, output resolution, presets, batch lists, or YouTube URLs.
• Standalone Windows app – a graphical UI that bundles its own backend (no Python needed). You can toggle steps, preview live, and save presets.
• After Effects panel – integrates directly into Adobe’s workflow for editors who want to run TAS inside AE.

Hardware support

• NVIDIA RTX/GTX (CUDA + TensorRT for best speed)
• Older NVIDIA cards (DirectML fallback)
• AMD / Intel on Windows (DirectML or OpenVINO)
• AMD on Linux (experimental ROCm)
• Apple Silicon (Metal Performance Shaders)
• Any Vulkan‑capable GPU (NCNN backend for some models)

Installation

• One‑liner Windows installer: iwr -useb https://tas.nevermindnilas.dev/install.ps1 | iex
• From source: Python 3.14 + requirements.txt + platform‑specific extra‑requirements (CUDA, DirectML, ROCm, etc.)
• First run auto‑downloads the required model files.

Why it matters

• Brings together dozens of state‑of‑the‑art anime‑focused AI models under a unified, scriptable interface.
• Handles the whole video pipeline in memory, saving disk I/O and time.
• Provides both power‑user CLI access and user‑friendly GUI/AE integrations, covering a wide range of creators.

License – AGPL‑3.0 for the code; model weights retain their original licences (see THIRD_PARTY_NOTICES.md).

---

All details above are taken directly from the repository’s README.