| library_name: onnxruntime | |
| tags: | |
| - onnx | |
| - onnxruntime | |
| - image-restoration | |
| - image-upscaling | |
| - face-restoration | |
| - laria | |
| # Image postprocess ONNX models | |
| ONNX post-processing models used by Laria/image generation tooling. | |
| ## Files | |
| - `RestoreFormer_PP.onnx`: face restoration model. | |
| - `face_detection_yunet_2023mar.onnx`: YuNet face detector used before restoration/alignment. | |
| - `RealESR_Gx4_fp16.onnx`: lightweight Real-ESRGAN x4 upscaler variant used by the current scripts. | |
| - `real_esrgan_general_x4v3.onnx` + `real_esrgan_general_x4v3.data`: quantized Real-ESRGAN general x4v3 package. | |
| - `real_esrgan_x4plus.onnx` + `real_esrgan_x4plus.data`: Real-ESRGAN x4plus package. | |
| - `metadata.json`: source/runtime metadata for `real_esrgan_general_x4v3`. | |
| ## Expected pipeline | |
| For generated images, run face restoration first at generation resolution, then upscale the restored image. This keeps the face restoration localized and lets the upscaler harmonize the restored face with the rest of the image. | |