Real-ESRGAN

Real-ESRGAN is the xinntao team's open-source universal image / video super-resolution project, first released in 2021, licensed under the permissive BSD-3-Clause, and written primarily in Python (PyTorch). The models and algorithms are widely reused; this is the de-facto standard for open-source super-resolution. VisLane's already-listed Upscayl is its GUI desktop application; the two share the technical core. Multiple model variants are provided (realesrgan-x4plus / realesrgan-x4plus-anime / realesr-general-x4v3), each fine-tuned for general photos, anime, 3D renderings, or portraits.

Core capabilities center on multi-degradation scenarios + practical usability. Model family: realesrgan-x4plus (general photos x4, basic), realesrgan-x4plus-anime (anime/illustration x4, texture preservation), realesr-general-x4v3 (Real-ESR General v3, with lightweight face restoration and detail enhancement), plus optional realesr-animevideov3 (video super-resolution with temporal consistency). Degradation modeling simulates real-world degradations during training (Gaussian noise / compression / blur / downsample / JPEG artifacts), rather than pure bicubic downsampling — this is the key improvement over the original ESRGAN, with significantly better recovery on real photos than academic-dataset training. Inference: PyTorch / NCNN / ONNX Runtime; CPU / NVIDIA GPU / AMD GPU / Apple Silicon all supported; x4 upscaling of 1024x1024 takes ~1-2s on GPU. Ecosystem: NCNN port enables embedded deployment (Xiaomi Mi 11 Pro's built-in AI upscale is based on Real-ESRGAN NCNN); ComfyUI node ComfyUI-Real-ESRGAN-Upscaler; built into Stable Diffusion WebUI / A1111 / Forge as the upscale script.

Use cases include: old photo / video quality restoration, SD / Flux generated image secondary upscaling (eliminating AI blur), e-commerce product image enlargement, video post-production 4K upscale, and anime / illustration high-definition conversion. Hardware floor: runs on CPU (slow), NVIDIA GPU recommended (4GB+ VRAM); NCNN port can run on Raspberry Pi-class embedded (but slow); FP16 halves VRAM. Onboarding: pip install realesrgan → python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg -o output.png; ComfyUI users drag in an Upscaler node; Upscayl GUI for non-technical users.

Licensed under BSD-3-Clause (permissive), allows commercial closed-source use, modification, and redistribution with only copyright notice retained; no requirement to release derivative source — this is the license-friendliness delta versus AGPL-3.0 (same tier as GPL-3.0) and GPL-3.0 super-resolution projects. Positioning versus VisLane's already-listed Upscayl (also by xinntao, cross-platform GUI) and Magnific (commercial AI upscale): Real-ESRGAN is the "CLI / API-level core"; Upscayl is its GUI desktop wrapper; Magnific is "creative-grade repaint upscale" (adds details but deviates from the original, commercial closed-source). The three complement each other: use Real-ESRGAN for real scenarios, Magnific for top-quality creative use.