@Bumblebee95 No, I mean there is an extension for completing/correcting tags, it just makes things a little quicker. As you're typing your tags, a little box pops up to show you suggestions or corrections. As an example, if you want a camouflage shirt or something, you'd type "camouflage" and the little window pops up to suggest "camouflage > camo" which is the correct tag e6 uses.
The one i'm using is called "a1111-sd-webui-tagcomplete" but it only comes with Danbooru tags, im hoping there is a variant that uses e621 tags. Sorry if that was confusing!

@Bumblebee95 Thank you, I was having trouble finding where the options were (I was expecting a separate file) but I found them in the script and now have working e6 tags. Thanks and also good work on your latest update :)

This rarely happens to me. What sort of workflow are you working with (example prompt and settings)? I rarely use style/artist tags, but when I have that's when I notice them the most, maybe that's why you're seeing so many?

I'm pretty sure textual inversion is a good option for avoiding watermarks and text. I use boring_e621 on pretty much everything I generate, if you're not using it you might find it helpful.

some say the SD doesn't understand what the "artist nametag" is. So it's useless to counter them. I've been trying myself, a lot of bunch of them. like. (artist name, artist tag, signature, watermark:1.3). And still it appeared on every pic

@yiffysan its also usually pretty easy to manually edit them out. Magic fill tool or whatever in photoshop would get rid of it a lot quicker than trying to fight it with negative prompting. You can also just generate slightly higher res in say the vertical hight by 20-50px and simply crop out the section that contains the signature, as its usually close to the bottom or top right.

Worst ones imo is when it tries to add Patreon logos lmao

@awoo it's kinda cute when the artist tag is appeared on sport swimsuit as if it's the name of the brand

Is it simply a matter of reducing noise in hires? Logically it seems that way to me, but when I do it with this model (less than 0.60) I tend to get weird blocky digital artefacts and tiling on the output image.
Hires fix in general confuses me a bit, its not exactly the same as using the Extras tab, is it?

@magacitlll717 what upscaler? I assume you're not using hires at all? Just working with 512px images and manually upscaling?

@awoo Yes, I just set the generation size to large, I didn’t connect upscaler separately. Well, if it doesn’t turn on automatically, I’m not very good at it yet.

I’ve been getting amazing results using the negative embeddings specified by the SnoutMix model plus ADetailer for the faces. I generate at 900x900 (no hiresfix) and then run it through Ultimate SD Upscaler at a low denoise (0.2-0.35) and set it to upscale by 2. ADetailer, ControlNet, and Ultimate SD Upscale are your BEST friends. If you’re not familiar with them you should be 🫡

I’d recommend using the 4xNMKD super scale for upscaling. It rocks.

@AIApprentice thanks very much for the suggestions! I will give them a try

When using hires fix, it's important to remember that the Denoising Strength slider controls how much freedom the upscaler has. A lower setting like 0.4 will result in lower deviation when the changes kick in. It seems like the hires upscaler is just an embedded img2img, so it tends to operate the same way.

r esrgan 4x with a denoise of no more than 0.3

the latent family of upscalers leave behind artifacts in lower denoise strengths afaik

@Talboc thanks, it's been confusing for me understanding exactly how hires fix works, but thinking of it like a streamlined img2img helps it make sense. Can you (or anyone else) explain what "latent" actually means in terms of an upscaler? I understand the basics of how ones like ESRGAN and SWIN IR work, and im pretty familiar with non AI upscalers (like bicubic) but what does "latent" mean within the context of SD? My assumption is instead of using say the ESRGAN model to upscale, "latent" uses the active checkpoint to upscale? Is this what it means? Also, do upscaler models like ESRGAN also use the active checkpoints model to upscale? Is there any difference between using hires.fix opposed to Extras tab upscaling?

Sorry for all the questions, also directed to anyone else who can answer. Im just super confused about what exactly hires fix is...

since bb95 is a merge of fluffyrock, the FluffyRock Autocomplete list should have every artist it can do. artist that has number count of ~150 or above tend to be most reliable

@paperplain Thank you!

@nemo_theoceanborn be aware as this is a photorealistic merge model, it won't do artist styles exactly. It can change the anatomy of subjects a fair bit tho.

@awoo It does indeed affect the result a lot, try using by honovy or by nuzzo and you will get very good, two completely different results, also helps the anatomy a lot in some cases.