Thanks! Nice to hear you like it. What I did is basically an updated berrymix with r_34 replaced by trinart, here's a recipe: 1. novelai * 0.95 + sd-1.5 * 0.05 (weighted sum) 2. set method to "add difference" - model A: mix_from_step_1 - model B: f222 - model C: sd-1.5 - multiplier M: 1 3. mix_from_step_2 * 0.8 + trinart2 * 0.2 (weighted sum)
I also used float16 if you want to get the exact same hash.
Thanks! Nice to hear you like it. What I did is basically an updated berrymix with r_34 replaced by trinart, here's a recipe: 1. novelai * 0.95 + sd-1.5 * 0.05 (weighted sum) 2. set method to "add difference" - model A: mix_from_step_1 - model B: f222 - model C: sd-1.5 - multiplier M: 1 3. mix_from_step_2 * 0.8 + trinart2 * 0.2 (weighted sum)
I also used float16 if you want to get the exact same hash.
I seem to be getting a slightly different hash, despite going float16. I also get totally different unrelated results... I suspect something is up with my A1111 or something. Thank you for your help though anyway!~
I seem to be getting a slightly different hash, despite going float16. I also get totally different unrelated results... I suspect something is up with my A1111 or something. Thank you for your help though anyway!~
Weird, is your webui updated? Do you get a completely different style or mangled results? More realistic or flat textures? Or is it just different poses but same style?
I went through and made my picks from the different tabs on that. Of course, all the examples included with that are generated from just SD1.5, and doesnt really show fully how it would impact a more specialized model (such as NovelAI). That's why i made this.
I went through and made my picks from the different tabs on that. Of course, all the examples included with that are generated from just SD1.5, and doesnt really show fully how it would impact a more specialized model (such as NovelAI). That's why i made this.
Thanks. All this stuff is so new. Learning the basics is a lot, and now there's extensions on top of it... Overwhelming in a fun way.
Very neat. I'll have to try a bunch of these on various models. The stuff getting generated on Yiffy-e18 is often very samey since it seems like half the prompts anyone uses invoke Pino Daeni, and I'm guilty of relying on him too much myself in that model.
To explain step 6 in more detail: The part where you type 0.1-1.0 (+0.1) is a fancy way of telling webui to create a series of images, starting with one at 0.1 denoising (almost identical to the masked section, paying very little respect to the new text prompt) all the way to 1.0 (respecting only the new text prompt and ignoring the masked section entirely), in +0.1 increments (0.1, 0.2, 0.3, etc., up to 1.0).
You could change that +0.1 to +0.05 and get twice as many images, or to +0.01 to get 10 times as many. It's up to an agreement between you and your graphics card.
Also, this is a BE example because that's the most obvious thing to do with this sequence, but it's pretty obvious how creative use of prompts can modify other things instead. I'll leave that up to your imagination.