PIXLRelight-ONNX / README.md
Reza2kn's picture
Add FP16 WebGPU-oriented ONNX export
07b0476 verified
|
Raw
History Blame Contribute Delete
1.15 kB
# PIXLRelight Renderer ONNX
This repository contains ONNX exports of the feed-forward PIXLRelight renderer core. The renderer takes a source RGB image and a 9-channel target-light intrinsic tensor, then returns relit RGB plus gain and bias maps.
Exports:
- `pixlrelight_renderer.onnx` + `.data`: FP32, CPU/server-safe. Inputs are float32 `source_images` `[B,3,512,512]` and `target_intrinsics` `[B,9,512,512]`.
- `webgpu/pixlrelight_renderer_fp16.onnx` + `.data`: FP16, browser/WebGPU-oriented. Inputs are float16 `source_images` `[B,3,512,512]` and `target_intrinsics` `[B,9,512,512]`.
- `int4/pixlrelight_renderer_int4.onnx` + `.data`: weight-only INT4 MatMulNBits export, currently the smallest browser-facing artifact. Inputs remain float32; activations remain floating point.
Outputs are `rgb`, `rgb_gain`, and `rgb_bias`.
Boundary: this is the PIXLRelight renderer core, not the upstream Marigold RGB target-image-to-intrinsics decomposer. Marigold is a separate diffusion pipeline and is intentionally not embedded in these ONNX files.
Upstream: https://github.com/mlfarinha/pixlrelight
License: CC BY-NC 4.0, following the upstream model.