search

tencent-ailab / IP-Adapter

The image prompt adapter is designed to enable a pretrained text-to-image diffusion model to generate images with image prompt.
Apache License 2.0
5.08k stars 331 forks source link

IP-Adapter-FaceID example training dataset? #249

Open jadechip opened 8 months ago

jadechip commented 8 months ago

Since the IP-Adapter-FaceID models are not available for commercial purposes, is there an example dataset for training a IP-Adapter-FaceID model? Would someone be interested in crowd sourcing a model with more a more permissive license?

whiterose199187 commented 8 months ago

The non-commercial restriction comes from use of https://github.com/deepinsight/insightface. I wonder what are the alternatives available. Would the use of https://github.com/davidsandberg/facenet give comparable results? I am a dev but don't know anything regarding deep learning.

Anyways, I would be interested in contributing to crowd sourcing a model for commercial use in whatever capacity I can.

whiterose199187 commented 8 months ago

hello,

I have a client who is willing to contribute by sponsoring GPU hours for training if anybody is interested.

xiaohu2015 commented 8 months ago

the model in facenet maybe also not allowed to commercial purposes

whiterose199187 commented 8 months ago

hello @xiaohu2015

I saw the license is MIT and did not see any explicit prohibition of using for commercial purposes, can you please point me to where you read that?

thhung commented 8 months ago

@whiterose199187 Yes, I didn't find anything related to the pretrained model in that repo. BTW, how much GPU hours can be offered? We might need to reproduce the dataset first.

whiterose199187 commented 7 months ago

@thhung do you have an estimate of what would be required? I can accordingly make a proposal to the sponsor with that info

thhung commented 7 months ago

@thhung do you have an estimate of what would be required? I can accordingly make a proposal to the sponsor with that info

@whiterose199187 Sorry for the delayed response. I was on vacation. Is it possible to request this gradually? It depends on several factors (such as GPUs, model size, dataset size, image resolutions, code optimization). It will be clearer once we have the first version end-to-end, which should be small enough to implement quickly. I would aim for a resolution of 512x512. If you would like to discuss further, you can contact me at damienlovedata[at]gmail[dot]com.

gabrielhrr commented 4 months ago

@thhung @whiterose199187 any news about the FaceID trained with facenet?