sd-webui-supermerger VS clip-interrogator-ext

Have u tried the supermerger extension ? https://github.com/hako-mikan/sd-webui-supermerger

Have u looked at https://github.com/hako-mikan/sd-webui-supermerger - it may or may not allow you to give comment. Im going on what im experiencing with this LORA experiment, it sort of updates parts of it, and if u put balanced models back in on reversing or offsetting the overcooked stage it gets to over time the model does seem to improve versus base 1.5.

I found this extension: supermerger

I merged this on Colab, so you don't need a powerful graphics card. You'll need the block merge extension,the model toolkit extension and the supermerger extension.

Instead what I'd recommend is a manual loop to hone in on what prompts work well. You'll need two extensions, assuming you're using a1111: AestheticScorer and clip interrogator. The aesthetic scorer will effectively rate generated images and attach a score from 1 to 10 to the image Metadata. There are a few image viewers which can view and sort by aesthetic score, breadboard and diffusion toolkit I believe are popular options. You can filter by the score to effectively throw out lower quality Generations, allowing you to focus on some of the better quality ones for Taste matching. After that you can sort through the remaining and see which ones fit your tastes the best.

I think CLIP Interrogator may be what you're looking for.

clip-interrogator-ext: to get a potential prompt from an image (I don't know if it works as I use a my own fork, because I use a newer version of tranformers)

yeah this was my first thought too, https://github.com/pharmapsychotic/clip-interrogator-ext

I basically started taking all of my best AI generated images as well as any images I see only that I really like and want to capture that asthetic in my works and use https://github.com/pharmapsychotic/clip-interrogator-ext.git extensions. Its Pharmapsychotics clip interrogator ext. Basically this long term process requires you to interrogate images you like often and then drop both positive and negative prompts from those images into your prompt. You will start out with a smaller prompt but it will grow over time if you keep adding tokens from your best images. In particular your negative prompt will start to look insane but its important to trust it just be careful to avoid repeating the same tokens and using any color specific language. I have used this method over the course of a couple of weeks to grow an existing prompt that has produced great images into behemoth prompt that has been minting extremely creative images for me in various angles colors and compositions. Another note of importance is that if you are constantly interrogating and adding prompt tokens from similar imagery as in women it will heavily bias your output towards women even if there are none in the positive prompt. Theoretically I feel what this is doing is narrowing down the models nodes to a very specific aesthetic that you are going for and therefore producing more provoking and top quality images especially in highly tuned models.

Seed Travel and Clip Interrogator extensions are both listen in the extensions tab of a1111, so thats the easiest route. But sure: https://github.com/yownas/seed_travel and https://github.com/pharmapsychotic/clip-interrogator-ext

Another way to do this is using the Clip Interrogator extension. This does a better job of analyzing the image and also does negatives. https://github.com/pharmapsychotic/clip-interrogator-ext.git