Stable Diffusion Generation Dump

Reader mode
Dead ass thread.



All the discord links are ded because discord decided to change their CDN external viewer to temporary. RIP free image hosting.
 
Hmm... sure. Here's something I did recently:


(open the image in a new tab for full resolution)

Trying to do anything with SDXL and weapons is a test of patience.

I wanted to do a crossbow first because it fits the character better (this being Sifra from 'The Demon Inside You' by contrived over on anonkun fiction.live), but I couldn't even find a single lora for crossbows and didn't feel like training one myself.
 
Last edited:
meloa789 said:
Will you threadmark this?
Click to expand...
Click to shrink...
I don't even know what I should threadmark anymore. TBH, at first it was actually pretty hard getting the model to spit out good images so I used to be pretty excited whenever it actually came out well but at this point models have advanced enough where I can throw in a bunch of danbooru tags and get a pretty decent image out of it.



I swapped to posting threadmarks for merges but y'know SDXL came out and I haven't remapped the blocks for that. I might as well just post whatever, and I honestly don't think anyone needs threadmarks for that. Well, I could share model recommendations if you guys are interested.
 
steamrick said:
I'd certainly be interested.
Click to expand...
Click to shrink...
Sure. Recently I've been using this model called Pony Diffusion for Anime. It's a finetune on Pony Diffusion XL6 by the same guy who trained Counterfeit. To some degree it's somewhat overtuned since the Counterfeit guy does a very narrow selection of training data but SDXL has a way huger context window compared to SD1.5 so this is not so much of something hampers is as something that aligns it. As for styles, I've been using two supplements mostly since I'm not a super big fan of the raw generic anime style. Either this one called nai style for pony, which I like because it does a very nice flat texture with good details on eyes and minimalistic linework, instead relying much more on shading to cover the gaps where it can:


Or this lora, which is trained on realism and when invoked on PonyDiffusionforAnime does the opposite and enables a lot of cool bloom and lighting texture but stays away from going full 2.5D by implementing extremely well defined linework and outlines. This one is way more finicky (and is a LoRA rather than a LoCon), since I have to prompt it at a low weight otherwise the realism takes over. It's a lot more demanding but I think the resulting style ends up looking a lot more unique.


OFC this isn't exactly a super great 1:1 comparison because the prompts are very different, due to the second LoCon needing a lot more tags to avoid slipping into its default realism style. This is a better comparison, with way less modifications to the prompt, though I did neglect to normalize the seed (since I rare have to do prompt tuning with SDXL based models I've gotten into the habit of just setting the seed to -1 for easy randomization).

 
Last edited:
I've given in to the dark side.

That is to say.... I've started genning with novelai:







 
Jemnite said:
Sure. Recently I've been using this model called Pony Diffusion for Anime. It's a finetune on Pony Diffusion XL6 by the same guy who trained Counterfeit. To some degree it's somewhat overtuned since the Counterfeit guy does a very narrow selection of training data but SDXL has a way huger context window compared to SD1.5 so this is not so much of something hampers is as something that aligns it. As for styles, I've been using two supplements mostly since I'm not a super big fan of the raw generic anime style. Either this one called nai style for pony, which I like because it does a very nice flat texture with good details on eyes and minimalistic linework, instead relying much more on shading to cover the gaps where it can:
Click to expand...
Click to shrink...
I'll admint that I've kind of steered clear of Pony Diffusion because of the annoying prompting requirements.

I really like the style of a model called Artium (v1.0), but didn't like the way it kept drifting off into photorealism.

In the end I made a remix of it that I called Artium_anima that I've been using as my default model for some time now. The main reason I waited until now to post is that I only ever uploaded the first version, but I did experiment more with it since and linking to an 'old' version didn't seem right.

Anyway, I uploaded my current working model now, so feel free to check it out.

Anyway, here's a not-so-good* example of what it can do:
*not-so-good because between the use of regional prompter, a detail tweaker lora, some upscaling and a judicious inpaint, this is certainly not representative of raw text-to-image.

For more examples of the default output style, feel free to check the sample images I've provided on civitai (see the link above)

Jemnite said:
That is to say.... I've started genning with novelai:
Click to expand...
Click to shrink...
Looks good.

The only thing I know about the NovelAI image generator is that it uses a diffusion based approach and once upon a time, AnythingV3.0 was leaked from them, greatly accelerating SD1.5 getting good.
Do you know if they use something based on SD1.5 / SDXL or something or if they've trained up something completely custom?
 
Reader mode
You must login or register to reply here.
Back
Top