[OLD] Picking Between chaiNNer and enhancr - Sirosky/Upscale-Hub GitHub Wiki
enhancr or chaiNNer?
Both enhancr and chaiNNer support video inference. However, chaiNNer is more of a multi-purpose image and video utility rather than a dedicated video upscaling application. enhancr is dedicated to video upscaling, interpolation and restoration. As a result, it is substantially faster than chaiNNer-- even in its free tier. The paid tier, unlocked at $7.50 per month on Patreon, supports tensorrt inference. I won't bore you with the technical mumbo jumbo of tensorrt-- it is effectively just a massive speedboost (think 5x faster or more) for nvidia GPUs. So if you have a lot of episodes of the Simpsons to get through, and have a nvidia GPU and don't mind shelling out some cash, then enhancr is hands-down the far better solution. This guide won't cover setting up enhancr, as fulsome documentation is available here.
That being said, despite being slower, chaiNNer is still a great option because of its broad utility. It also supports a wider array of models than enhancr. If you want to upscale individual images, manga/comics, Stable Diffusion or Midjourney output etc., chaiNNer is the way to go.
Comparison Table
chaiNNer | enhancr (free) | enhancr (paid) | |
Price | Free | Free | $7.50/month |
Upscale Videos | ✅ | ✅ | ✅ |
Upscale Images/Image Sequences | ✅ | ❌ | ❌ |
Video Interpolation | ❌ | ✅ | ✅ |
Video Inference Speed | 🔥 (via Pytorch) | 🔥🔥🔥 (via DirectML or NCNN) | 🔥🔥🔥🔥🔥 (via tensorrt) |
Inference Frameworks Supported | Pytorch, ONNX, NCNN | ONNX, NCNN, DirectML | Pytorch, NCNN, DirectML, tensorrt |
Notable Architectures NOT Supported* | DITN, CUGAN/Shuffle Cugan** | DITN, DAT, OmniSR, SRFormer, SwinIR (partial support) | DITN, DAT, OmniSR, SRFormer, SwinIR (partial support) |
Output Containers | mkv, mp4, mov, avi, image sequence | mkv, mp4, webm, mov, image sequence | mkv, mp4, webm, mov, image sequence |
Codecs Supported for Export | 264, 265, VP9, FFV1 | 264, 265, VP9, FFV1, AV1, ProRes | 264, 265, VP9, FFV1, AV1, ProRes |
UI Style | Node-based | Tab-based (should be familiar to most people) | Tab-based (should be familiar to most people) |
* It should be noted that the vast majority of the public models are trained on ESRGAN and Compact anyway. The architectures not supported by enhancr are newer / more novel archs that aren't widely trained on at this time.
** chaiNNer supports these architectures via ONNX runtime, though ONNX is typically slower.