Two Suggestions: Adding more to metadata info, VAE tags

(pre-statement: both of these suggestions just came out of me wanting to list more info without having to put it in the description of my posts, this could easily reduce the amount of info a poster has to provide manually in the description if they want to give a more clear step-by-step image creation process)

Firstly going over metadata, considering there's been a few additions to PNG info, I think it's worth revisiting metadata info and to allow more information to be stored in it. Basically, anything that PNG info can pick up on that we currently do not have fields to fill out for:
-Clip skip value
-LORA hashes
-TI Hashes
-VAE hashes
-Version number (this is different for each SD UI, so maybe also allow a field to specify which UI was used?)
Now I do understand that there are older images without these metadata fields avaliable, as looking at my gens from each date, it was really only around July and August that at least A1111 started adding more to the PNG info, but I still feel like these are valuable fields to add, especially now that the model name and VAE name, not just the hashes, are now included in PNG info. Posters could still add these fields to older posts of course, as I'll plan to do so if these are added.

Secondly, I'm wondering if it would be worth creating tags for VAEs? It could be valuable info for anyone who wants to do recreations for research purposes, just like listing the model tag and metadata. If so, how could we go about creating tags for them?

There is a reason why it will be quite difficult to add more information to the metadata, we've already talked about this a couple times.

Also, i don't think the VAE name/hashes is important.

And imo tags for vae sounds like a bad idea, for example because half of ai artists dont even know what vae they are using / use a model with baked vae

...and why would you need a version number?

(this is different for each SD UI, so maybe also allow a field to specify which UI was used?)

there is comfyui, so mb we can just use tags for this. BUT, anyway we will have a problem if we want to specify which ui was used, because there are a lot of different online services for generating images

This has been suggested numerous times on the discord server for ~6-8 months unfortunately, mainly clip skip value, model name, and generation size. Some of this stuff can be overwritten with an img2img upscale or otherwise, and like iodoff mentioned, the metadata can be written differently depending on ui/service used etc. However still feels like there can be a baseline or default that is accounted for as a start just as there is now, ie a1111 webUI formats.

I've also suggested adding various tag types at one point for things such as LoRa/Lycoris, as some get pretty popular and can alter styles heavily. Checkpoint tags/other model tag types as a differentiation instead of just having an image that used a handful of LoRa models be tagged with lora having the actual information hidden behind metadata you'd have to download the original size to see if it's not in commentary. I had suggested more relating to this on the discord before, but kind of given up bringing a lot of this stuff up since it was months ago with little interest given.

I also don't think vae tags are too necessary or needed imo - too many models have them baked in and they're all relatively similar.
I'm still all for adding in the main missing fields for metadata though, would be helpful in a lot of scenarios.

topic #191 brings up the topics you mentioned in a more generalized way. Basically, having the ability to add arbitrary fields and search by those would be the most future-proof. Similar to the ability you can search for the existence of EXIF keys like EXIF:PNG:Parameters.

Secondly, I'm wondering if it would be worth creating tags for VAEs?

I heavily disagree because every VAE is based on kl-f8, which came from CompVis. Merges inherently make things worse which users unfortunately don't understand and sites like Civitai have proliferated, so I wouldn't honor them by creating tags. And even if they were tagged it would be difficult as mentioned by others where it's often arbitrarily baked into the model.

I would personally like to have all the settings that aren't nondeterministic listed, like initial generation size, clip skip, sigma skip, Eta noise seed delta, etc, etc. You can't reliably replicate things like upscalers, especially if they're latent, so those parameters aren't really useful for image replication.

Ocean3 said:

I also don't think vae tags are too necessary or needed imo - too many models have them baked in and they're all relatively similar.
I'm still all for adding in the main missing fields for metadata though, would be helpful in a lot of scenarios.

VAEs can heavily alter the image.

kl-f8-anime
https://aibooru.online/media_assets/206250
blessed2
https://aibooru.online/media_assets/206249

It's probably hard to track but they do have a large effect on the resulting output.

Jemnite said:

I would personally like to have all the settings that aren't nondeterministic listed, like initial generation size, clip skip, sigma skip, Eta noise seed delta, etc, etc. You can't reliably replicate things like upscalers, especially if they're latent, so those parameters aren't really useful for image replication.

VAEs can heavily alter the image.

kl-f8-anime
https://aibooru.online/media_assets/206250
blessed2
https://aibooru.online/media_assets/206249

That's essentially what I was referring to as being similar, color/saturation/minor details aside - in my experience they rarely change the composition of images much, factors depending. Just a different flavor/filter to put things simply.

I second eta noise values and such, that can be a large component that I didn't mention. I see this info on images not just being valuable for reproduction (which I don't care for personally), but also just for referencing to see what is doing what etc.