Pedro Cuenca
Pedro Cuenca
Addresses #12. I also double-checked that the input for LPIPS are images in the range `[-1, 1]`, everything looks correct in that regard.
I did some changes to make it easier to jit the safety checker: - Do not transpose (assumes a certain number of dimensions). The array should be prepared externally. -...
This is the biggest change derived from #5. In that branch I was previously using a standard Python dataclass and a `dict` to manage the scheduler state, and I converted...
The scheduler actually provides 51 `timesteps` when setting `num_inference_steps` to 50. PyTorch reference: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py#L136
Some small changes to the notebook: - Normalize images so their values lie in the [0, 1] range. - Use all the generated words if no period is part of...
Testing procedure: - Compile and run on device. Wait until the model is downloaded and loaded. - Kill the app, detach from Xcode. - Shutdown the device, wait until it's...
Like in Stable Diffusion, no attention mask appears to be used for input tokens: https://github.com/huggingface/open-muse/blob/2a0365707c0c3d91b0b91bc456161db693652ff1/muse/pipeline_muse.py#L93-L101 But [according to third party analysis](https://github.com/huggingface/diffusers/pull/2634) this appears to have been a mistake all along....
Some initial tests for now.
Entering a new user message: https://github.com/huggingface/huggingface.js/assets/1177582/19bdad52-b8ea-45f0-affc-c2fffb4e7f05 After generation is done it scrolls down to the end of the response.