I'll be using this readme to document stuff basically. Maybe something like my thinking around this project, what i need to get done, what i should focus on, etc etc..
I pushed the changes earlier but haven't gotten time to add new things to project or edit the readme. So, what's the current state of progress and project? Firstly, the first commit i did was to just display an image using a stable-diffusion-v1-5 model by giving it a prompt and various other parameters such as height, width, guidance_scale, negative_prompt and num_inference_steps. Let me explain a bit of how i did this (actually took help of diffusers starter colab) and a bit about parameters also. So, diffusers lib present on hugging face provides various pipelines to load pretrained models which can then be used for image generation. There are many other tasks as well such as image to image, inpainting, etc. (i don't know about most of them) but firstly we'll be looking into text to generation only. The pipeline that i'm using for this is a StableDiffusionPipeline which will load the above mentioned model. Now, coming to parameters -
One thing to keep in mind (struggled a lot with it) - on macbooks, you can use the device 'mps' instead of 'cuda' for faster generation of images. Next up, trying to use a ddpm/ddim scheduler and model for image generation which generates an image from noise. Will use a UNet2D model for this which generates an image of the same shape as input. The way this works is -
Some things i had thought of - https://app.eraser.io/workspace/m5aXyZfI4WYHwhaJghHi?origin=share
I'll be using this readme to document stuff basically. Maybe something like my thinking around this project, what i need to get done, what i should focus on, etc etc..
I pushed the changes earlier but haven't gotten time to add new things to project or edit the readme. So, what's the current state of progress and project? Firstly, the first commit i did was to just display an image using a stable-diffusion-v1-5 model by giving it a prompt and various other parameters such as height, width, guidance_scale, negative_prompt and num_inference_steps. Let me explain a bit of how i did this (actually took help of diffusers starter colab) and a bit about parameters also. So, diffusers lib present on hugging face provides various pipelines to load pretrained models which can then be used for image generation. There are many other tasks as well such as image to image, inpainting, etc. (i don't know about most of them) but firstly we'll be looking into text to generation only. The pipeline that i'm using for this is a StableDiffusionPipeline which will load the above mentioned model. Now, coming to parameters -
One thing to keep in mind (struggled a lot with it) - on macbooks, you can use the device 'mps' instead of 'cuda' for faster generation of images. Next up, trying to use a ddpm/ddim scheduler and model for image generation which generates an image from noise. Will use a UNet2D model for this which generates an image of the same shape as input. The way this works is -
Some things i had thought of - https://app.eraser.io/workspace/m5aXyZfI4WYHwhaJghHi?origin=share
