Toufikben/productchat-models

ProductChat Studio model artifacts

This repository contains model artifacts used by ProductChat Studio. Each artifact keeps its upstream attribution and license notice. The repository license field is Apache-2.0 for the LaMa artifact; see the per-model notices below.

Included artifacts

File	Function	Upstream	License	SHA-256
lama_fp32.onnx	512×512 masked inpainting	sapienkit/LaMa-ONNX, based on advimman/lama	Apache-2.0; Places2 attribution required	1faef5301d78db7dda502fe59966957ec4b79dd64e16f03ed96913c7a4eb68d6
RealESRGAN_x4plus.pth	General-purpose 4× super-resolution	xinntao/Real-ESRGAN	BSD-3-Clause	4fa0d38905f75ac06eb49a7951b426670021be3018265fd191d2125df9d682f1

LaMa input contract

lama_fp32.onnx expects image with shape [1,3,512,512] and mask with shape [1,1,512,512], both float32. The mask uses 1 for the region to erase and 0 for the region to keep. The output is RGB float32 in the [0,255] range.

MI-GAN status

MI-GAN is not included yet. The official implementation is MIT-licensed, but the released pretrained weights do not currently provide a clear model-card or weight-license statement. The weights will only be added after the upstream authors confirm that commercial redistribution is permitted.

Notices

• LaMa: see LICENSE-LAMA.txt. The model card attributes Places2, which is licensed CC-BY 4.0.
• Real-ESRGAN: see LICENSE-REALESRGAN.txt.
• This repository does not include DreamLite.

Sources

• LaMa: https://github.com/advimman/lama
• LaMa ONNX export: https://huggingface.co/sapienkit/LaMa-ONNX
• Real-ESRGAN: https://github.com/xinntao/Real-ESRGAN