Shyamgopal Karthik

Results 10 comments of Shyamgopal Karthik

Hey, The gpt-3.5-turbo model call is being made here: https://github.com/ExplainableML/Vision_by_Language/blob/master/src/utils.py#L180. You would need to replace the API call with your call to a LLaMa model, the prompt (with the in-context...

Hi, while we had discussed similar experiments, I don't think we have ever tried this out, so I am quite curious to see what you get with this. Coming to...

Sure, I think most recent versions of diffusers (even 0.24 worked a few months ago I think) would work smoothly. If you are interested, I can provide the pip freeze...

Sure, you can remove the memsave argument (and its application in the model directly if you prefer). There could be issues that it's causing (especially if you're also applying it...

Hello, thanks for your interest in our work! I just went through regenerating the results on different benchmarks and realized that for Fashion-IQ they do seem to be based on...

Hey, sorry for the delayed response. We were using the LAION Aesthetic Predictor (https://github.com/christophschuhmann/improved-aesthetic-predictor). You can find the class definition here (https://github.com/ExplainableML/ReNO/blob/main/rewards/aesthetic.py). The only difference is we return `10.0 -...

Thanks for your interest! If you set the model to `sd-turbo`, you would be using the distilled 1-step version of the SD2.1 model, and as you might observe, it works...

This is quite surprising. Are you facing these issues with one-step models (e.g. SD-Turbo, SDXL-Turbo)? The regularization objective is mostly just to ensure that the norm stays around the original...

Thanks for the suggestion! All the models we used (e.g. CLIP, PickScore, HPSv2.1) were mostly based on the CLIP architecture where it's meaningful to compute scores/similarities from taking a dot...

Thanks for your interest! We added support for Flux-Schnell which supports ControlNets (https://huggingface.co/docs/diffusers/main/en/api/pipelines/controlnet_flux), even if it's not explicitly designed for it. Our codebase does not directly support the FluxControlNetPipeline, but...