VisionSR

Paid APILinuxSelf-hostedWebWindows

What VisionSR is

VisionSR is a free, Apache-2.0 open-source engine for local image enlargement and restoration. It supports upscaling up to 4x, face recovery, detail reconstruction, background removal, image analysis, and batch folder processing. Automatic routing examines image content and damage, then selects a model and tile size using available VRAM and inspectable heuristic rules. The project lists nine models covering photographs, portraits, scans, line art, anime and illustrations, face restoration, background removal, and face detection. Interfaces include a CLI, Python library, HTTP API, web UI, and Electron desktop app. The API handles analysis, enhancement jobs, live progress over SSE, and results. Supported backends include PyTorch, ONNX Runtime, DirectML, and OpenVINO. Processing stays local without uploads, per-image charges, or queues. Python 3.11 or later is required. GPU installation requires an NVIDIA card, while the packaged Windows ONNX and DirectML runtime supports DX12 GPUs from NVIDIA, AMD, or Intel.

Who it suits

VisionSR suits users processing photographs, portraits, scans, illustrations, or line art locally. It offers several interfaces for developers and desktop users, with hardware requirements centered on Python and GPU capability.

What it does well

Upscaling up to 4x
Local processing without uploads or queues
CLI, API, web, and desktop interfaces
Batch folder processing
Nine listed specialized models

What to know first

Python 3.11 or later is required
GPU installation requires an NVIDIA card
Training pipeline is not built
Cloud or multi-tenant layer is not built

Verdict

VisionSR provides broad local image-restoration workflows with automatic model routing and multiple interfaces. Its hardware, Python, and unfinished cloud-related components define the main constraints.

Facts from VisionSR's own pages

API: The HTTP API includes image analysis, enhancement jobs, live progress events over SSE, and job results. · github.com, 12 Oct 2026
API access: Yes · github.com, 22 Sept 2026
Automatic routing: It analyzes image content and damage, then automatically selects a suitable model and tile size based on available VRAM. · github.com, 12 Oct 2026
Batch processing: Yes · github.com, 22 Sept 2026
Desktop app: Yes · github.com, 22 Sept 2026
Desktop features: The desktop app includes a watched hot folder and native open and save, and starts its own engine process. · github.com, 12 Oct 2026
Hardware: The README describes tiled inference designed for a 4 GB GPU; GPU installation requires an NVIDIA card, while the packaged ONNX Runtime and DirectML app is described as supporting DX12 GPUs from NVIDIA, AMD, or Intel. · github.com, 12 Oct 2026
Hardware requirements: The project requires Python 3.11 or later, and an NVIDIA card for GPU use; tiled inference is designed to work from 4 GB VRAM. · github.com, 12 Oct 2026
Image tools: Listed capabilities include super resolution, face restoration, and background removal. · github.com, 12 Oct 2026
Image workflows: Features include batch folder processing, face restoration, background removal, image analysis, and model inspection. · github.com, 12 Oct 2026
Integrations: The README documents PyTorch, ONNX Runtime, DirectML, and OpenVINO as supported inference backends. · github.com, 12 Oct 2026
Intended users: The project describes a 4 GB laptop GPU as its design target and includes model paths for photographs, portraits, scans, and line art. · github.com, 12 Oct 2026
Interfaces: It provides a CLI, Python interface, HTTP API, web UI, and Electron desktop app. · github.com, 12 Oct 2026
License: The project is licensed under Apache-2.0, and the README says model weights have their own licenses and are downloaded rather than redistributed. · github.com, 12 Oct 2026
Limits: The README says the training pipeline and cloud or multi-tenant layer are not built. · github.com, 12 Oct 2026
Local processing: The project says images are processed locally, with no upload, per-image cost, or queue. · github.com, 12 Oct 2026
Maximum scale: 4x · github.com, 22 Sept 2026
Models: The README lists nine models for super resolution, anime and illustration upscaling, face restoration, background removal, and face detection. · github.com, 12 Oct 2026
Model selection: Its content classification and model selection use inspectable heuristic rules and image statistics rather than a learned classifier. · github.com, 12 Oct 2026
Notable limits: The README says the training pipeline and cloud or multi-tenant layer are not built. · github.com, 12 Oct 2026
Purpose: VisionSR is an AI super-resolution engine for image upscaling, restoration, face recovery, and detail reconstruction. · github.com, 12 Oct 2026
Runtime options: The documented backends include PyTorch, ONNX Runtime, DirectML, and OpenVINO; the README says the shipped desktop runtime uses ONNX and can run through DirectML on DX12 GPUs. · github.com, 12 Oct 2026
Security: The repository README lists CI security workflows for CVEs, CodeQL, secrets, and filesystem checks. · github.com, 12 Oct 2026
Security checks: The repository README lists CI security workflows for CVEs, CodeQL, secrets, and filesystem checks. · github.com, 12 Oct 2026

Where VisionSR stands on Freedom251

Image Upscaling Software#85 of 99