Posts

Lol, Teaspoon is an admin??! 😳

@Cypher
Thanks! Will add them in next pass Soon!

@Teaspoon
1.LORA:Lily Valley https://civitai.com/models/1185223
2.LORA:Mjölna https://civitai.com/models/1186001/my-little-ponymlp-g4-mjolna

I saw a couple of X Y plots ages ago, and wonder what is the best way to generate them, automatically adding image as well as the configuration that changed for them in the plot.
(Things like model, sampler, scheduler, steps, prompt, CFG, size/aspect ratio etc.)

@tyto4tme4l
It’s intended to be NSFW in my case.

@Thoryn
If it’s SFW, then maybe try prompting it in Bing Image Creator and then use the result for img2img in Stable Diffusion? Bing has a much better prompt following.

I’m struggling hard with text to image prompts getting characters to interact with furniture, prompting for them to be on something like a changing table is seemingly impossible with PonyV6 and ZoinksNoob. I’ve tried CFG from 4 to 7, and even kept the prompts to 75 blocks(?) on many of the attempts.

Figured I’d work around it by using a dresser instead, but it refuses to put characters on it.
The dresser gets identified, since if I type e.g. wooden or white dresser it usually changes accordingly, but it still puts the character on a bed, in a crib, ordinary table or the floor instead. Even when I up the strength for dresser, and puts crib and bed in negative prompts, it still includes one of them to put the character there, if not the floor.
Using alternative words for dresser, like chest of drawers, or similar things like commode, has the same outcome.

Later in the year I’ll hopefully have a better setup and can try inpainting/sketching (as if I’m not bad enough at drawing under optimal conditions already, trackpad really sucks for drawing), but for now I’m dumbfounded as to why I can’t get my prompts to work how I want. Any tips would be appreciated.

@derp621
@Cypher
Fixed, thank you!

All pending G4 character I had should be there now.

Will move to adding pending G5 characters, then G1-3.5, then OCs.

@derp621
https://civitai.com/models/1155005/my-little-pony-collection-pack-ponyxl-lyra-heartstrings

Lyras Heartstrings model link leads to Lotus Blossums

All right, I think I’m done with formatting in a way that’s easy to read and easy to add to, so will get back to adding new characters…

@Cypher
What happened D:

I’m surprised

This link redirects me to https://notebooklm.google/?location=unsupported. Although twibooru loads just fine 😳

I think this grouping makes sense. I’d remove the separation to background/minor.

Consider separating the characters by newlines instead of commas, like this:

Maybe with separate groups for every letter of alphabet and maybe spread into several vertical columns.

@Teaspoon
If you still want to sort by category, Major, Minor, Incidental, and Background Character is another degree of separation. Incidental characters would be characters that do play some kind of role in an episode vs background characters which could literally not exist and nothing would change.

@Bendy and Boney
I’d probably still do

@Teaspoon
I think Alphabetical will be the most useful, as everyone has different definitions of minor/background/major. The only actual distinctions should be between objective things like generations. Maybe even make male/female or adult/foal distinctions? idk, but alphabetical is probably enough.

I think alphabetical or maybe also organization by season/generation/other medium of origin (e.g. comics/movie/specials) is what I’d do.

Wanting some opinions…

As you can see, that’s a lot of characters. And I still have some 100 more to add, plus things like alt outfits and the like.

What I’m running in is the fuzzy separation between Background Character - Minor Character - Primary Character. Lyra is a background character, but she did have a few lines in one episode, so does that push her to Minor? Daring Do, not really important to the “overarching plot” of the series, but she’s a fully developed character in more than one episode, so does that push her from Minor to Primary?

That sort of thing.

So basically, does anyone care about these sort of distinctions, how much? Or should I just go fully alphabetical (and split by A-G, H-P, Q-Z or the like)?

@CappyAdams
They do. You just need a custom node set like Dynamic Prompts. The A1111 extension actually split off a bunch of its code into a separate library, and the custom node sets generally use the same library, so they basically work the same.

With Dynamic Prompts, you’d basically go into the wildcards folder inside the folder for dynamic prompts, and put the json files there, IIRC.

I’ve been trying to make images where a character is also seen in a mirror, but don’t get any results that are remotely close to decent.
Is it possible with just text to image prompting, or do I have to use sketching and image to image, external image tools or other tricks?
For the record currently using Pony V6.

@CappyAdams
There are custom nodes for nearly anything imaginable in ComfyUI. Many of the wildcard nodes support the .json format.

@Lord Waite
Silly question (probably), but does the json file work with comfyui? I assume it requires some custom nodes-

Ooh, nice!

Unless the model specifies it needs a specific sampler, I tend to opt for Restart since it’s like a mini hi-res fix.

Do you have a similar resource for Scheduling type? In most of my tests the differences are minimal, as art timing. Except of some Illustrious based models where not using the recommended adds a lot of “noise patterns” (waves, flowers, swirls…) in empty areas.

This doesn’t have all of them, but there’s a fair amount of information on them right here:
https://stable-diffusion-art.com/samplers/

In practice, I tend to like the SDE ones, but the fact that they don’t converge can be an issue, and they don’t work well on some models, IIRC. Euler/Euler a seems to be what people default back to a lot, partially because of A1111 defaulting to it. I do like DPM++ 2M, and UniPC needs less steps then normal and can be useful when you’re generating a bunch of pictures. DPM Fast sucks. It was an early attempt to be fast, and I haven’t really seen it come out with good pictures.

Never a bad idea to try switching around samplers occasionally when generating things to see if another sampler works better. Just remember what you had it on before.