A.I. Question Training a Keds LoRA (1 Viewer)

[Arcolyte]

The latest round of AI images all show a lot of promise, but if we want to generate real-looking Champions instead of random AI slop we need to put in some real effort, and I think we can pull it off as a team.

What we need to do is train what's called a LoRA. LoRA stands for Low Ranked Adaptation and is a type of training method for fine-tuning Stable Diffusion models. Basically, we need to make an effort to truly "teach" the AI what Keds look like.

In order to do this we need two things:

1. A boat-load of Keds images
   At least 250 images. These need to be really good, ideally high resolution (minimum 1024x1024, square) and clean (no motion blur, no out-of-focus, fully visible with nothing in the way) and being worn. We need as wide as possible variety of different positions, viewpoints, states (bent, flat, etc), lighting. The better the dataset we can put in, the better the training result we will get. We should also make sure the dataset has no knock-offs, Vans, or anything else that could confuse the training.
   
   
2. Processing power
   We need to run the processing to generate the LoRA based on the images. I have very high-end hardware I can contribute for this part.

You can train LoRAs for many of the open-source Stable Diffusion models, like SDXL, etc. In my experience I think we should train one for Flux, which seems to generate the best images right now.

@Seeker do you have any ideas on the logistics for how we could all contribute images? I imagine some of us will need to volunteer as contributors (gathering images from the vast galleries here), and others to help as moderators, culling bad images.

Once we work together to build this LoRA, we can all use it to generate some amazing stuff. We can also use the dataset to train new LoRAs for video synthesis and beyond. It will be valuable to us for years to come.

 

[Seeker]

The latest round of AI images all show a lot of promise, but if we want to generate real-looking Champions instead of random AI slop we need to put in some real effort, and I think we can pull it off as a team.

What we need to do is train what's called a LoRA. LoRA stands for Low Ranked Adaptation and is a type of training method for fine-tuning Stable Diffusion models. Basically, we need to make an effort to truly "teach" the AI what Keds look like.

In order to do this we need two things:

1. A boat-load of Keds images
   At least 250 images. These need to be really good, ideally high resolution (minimum 1024x1024, square) and clean (no motion blur, no out-of-focus, fully visible with nothing in the way) and being worn. We need as wide as possible variety of different positions, viewpoints, states (bent, flat, etc), lighting. The better the dataset we can put in, the better the training result we will get. We should also make sure the dataset has no knock-offs, Vans, or anything else that could confuse the training.
   
   
2. Processing power
   We need to run the processing to generate the LoRA based on the images. I have very high-end hardware I can contribute for this part.

You can train LoRAs for many of the open-source Stable Diffusion models, like SDXL, etc. In my experience I think we should train one for Flux, which seems to generate the best images right now.

@Seeker do you have any ideas on the logistics for how we could all contribute images? I imagine some of us will need to volunteer as contributors (gathering images from the vast galleries here), and others to help as moderators, culling bad images.

Once we work together to build this LoRA, we can all use it to generate some amazing stuff. We can also use the dataset to train new LoRAs for video synthesis and beyond. It will be valuable to us for years to come.

For my part I have originals for all the images in the SeekersBasement sites section. Quality over the years improved greatly from around 2005 and on. Going through the Keds collection and picking candidate photos will be a chore.

 

[tidybowler]

I have a large collection, too (a portion of it overlaps Seeker's posts in the Basement). Some of mine have already been uploaded to "Create Your Own Gallery". A tag search for "Keds" or the few variations of that tag that exist might save some time, although we know there are many images that lack tags.