Inspired by this Don't starve lora for SD1.5 by @1to3fall5 and Don't starve style by @dimondexe. Special thanks to @xLordDonax for giving me detailed feedback about v1.2 and v1.3, @SulfurSniffer for feedback about v1.4. Also, a person who gave 1000 buzz (I don't see the name) some time ago. And everyone else who gave me a motivation.
The lore doesn't include names of original characters, so you won't get them in generations. (v1.5 has names of some characters, but using them on prompt didn't notice real effect.)
I recommend checking the LORA by @NeuralGL for Flux, instead of mine, which is based on game assets, but still looks pretty nice.
Description for V1.5 Pony : Put attention to the original characters have 4 fingers. To get 5, you may need to do inpainting. Strength 1.0-0.85 with additional character LORA, in some cases use '2D, CGI, photo, render in negative, 0.9-0.75 and lower for other cases if you want to keep the original style. Around 0.5 is fine for more detailed images but still maintain some influence of the style. 'chibi' I recommend using no stronger than 0.8. Comparing v1.5 and v1.4 for Pony, weight examples, 'chibi' all in one in the technical post. @SulfurSniffer is creator of Fate Spinner Lora.
Description for V1.4 Illustrious: 'dontstarvestyle' and 'blank eyes' are trigger words. The results I like the most with weights 0.7-0.9. Weights can be seen on the last picture. Lips might have incorrect shapes, especially at higher weights, so use addetailer or inpainting.
Description for V1.4 Pony: Strength '1.0' is usually fine for everything. Use 'dontstarvestyle', 'blank eyes', and 'muted colors' to make images closer to the original palette. The images in examples and in technical post used almost identical settings and prompts like in v1.3, so you could easily compare version v1.3 and v1.4. If you want the character more chibi, use 'solo' and 'chibi:0.7' and lower. Also, it works better with a strength of the Lora '1.0' and lower. Link on technical post. Usually, you don't need any specific negative, except when using LORAs with realistic characters and persons. Also, 'Liara t'soni' is trying to stay semi-realistic; my usual negative for it is '3d, photo, cgi, render'.
Description for V1.3 Pony: In general, everything similar like in v1.2. but you could use a bit lower wieghts. 'Muted colors' is not really a keyword, but use it if you want to get colors closer to original style, maybe even increase the weight of it. 'chibi' by part works like a keyword, but anyway just add similarity to style and proportion. By reason most of the dataset contain nature outside, it sometimes tries to put pot plants and stuff like that, so you could put it in negative if you want to avoid it. As usual, link on technical post to see examples of weights.
Description for V1.2 Pony: For usual usage, the weight ranges from 1.2-0.8. Without any description of the background, it will usually be simple or blank, so I recommend putting even a simple description like 'forest' or 'bedroom'. With the token 'chibi', it will look like a playable character. If you use it with other Lora of characters, I recommend increasing the weight between 1.2 to 1.6. Also, I recommend putting 'pupil' in negative if you want to get clear blank eyes. Some examples can be seen in my main post. The below is also an image with examples of weights. Link to post of weight examples.
Description for V1 Pony: ( I don't recommend using this version now; it's not stable enough.)
By default, with trigger words 'blank eyes, dontstarvestyle' generate 2.5D images. If you want to do it more flat and/or use it with another LoRA or character or person, add 'photo, render, 3d, cgi' in negative. Examples you could see in the technical post. The trigger 'blank eyes' works, but not like you expected; by fact, it works as a name for the style itself. It's possible to do another type of eyes, as the dataset has variants for it, but it's hard to get it, and actually, I like characters with blank eyes like Wendy. The LoRA isn't including names of game characters, so it won't generate specific ones from the Don't Starve series.
Description
FAQ
Comments (5)
You are so productive last days xD Last version you said... haha, just kidding, i don't mind awesome LORA being even better. I did some Morrigan sets and decided to left DS 1.3 bit for later but seeing new version released i remixed a few pictures and did a lot of new ones. Thanks for inspiration :D
Didn't do a lot of deep testing, just enjoyed the process. All were smooth and results are beatiful. The only things i noticed are some hands problem. But i guess it's just custom character's LORA conflict and generating in a nutshell. And sometimes despite eyes being mostly blank you can notice pupil hiding xD But it's not always happening so you can try more or inpaint (or paint lol).
It's very cool you sharing your own tests, explaining keywords and advice some useful negatives. Some people barely do any descripritions where they really NEEDED.
So anyways, thanks for awesome LORA (for the third time now haha). It's very unique, well made and enjoyable. Will definitely tinker with it more and share my experience. Will publish later today.
Thank you very much for your kind words and detailed feedback! I spent much time trying to make an archived version of Kylie Griffin, after many dozens of tries, I figured out why I couldn't make it good enough. Even though I had a failure in it, I think I found better settings for training or at least made it 10 times faster. About pupil it is interesting because I didn't get it even once in my tests. Maybe the reason is a mixing and/or long prompt. I don't like using mixed-style loras and it is rare when I use prompts more than one chunk. About hands, I agree with you, one reason is the style itself being more 'sketchy' and should be even more messy, so some images in the dataset have very simple hands. Another reason is my mistake, I think, even with my realistic loras with good datasets, I have that problem, but I'm not sure if it's so bad compared to loras from other authors.
@DemonAlone yeah, i think it's all about character LORAs conflicts. Shouldn't be prompt, it is not that overbloated but who knows what tokens are problematic. Probably also my low-step Euler tests doing the job xD
It is very cool you didn't just give up with Kylie. Who knows if we had this awesome LORA if you wouldn't be more persistent :)
Yeah, i usually use no more than 1 style. Rarely 2, like artist style + something stylish for shadows/lightning/specific vibe without losing artist style (Sinozick and BSS are usually my two favorite of these kind). DS with its aesthetics doesn't need any additional mixes so it's just character LORA (if needed) + DS.
Yeah, i've seen MUCH worse LORAs someone did just to fill some niches. Like the ones with very specific styles with zero explanation how to reach it without 1/100 ratio if not worse. Yours just have very little "AI moment" based problems and specific source material like you said.
@SulfurSniffer , Well, a few days ago I found a 'cheap' version on my Lora. Even the preview is with a pink-haired girl hugging a red panda, like in v1.3, but the results... So, this is success, I think). About Kylie, I keep working, but try another way to get a dataset. Now, I am understanding why some AI artists make photo sessions. About lower steps and lower CFG, it's interesting to explore again, thank you for the tip. I did research, but maybe I will find a better combination.
@DemonAlone yeah, sometimes i wondering how 19 step Euler A can produce better results than 32 step DPM+++ 2M karras :D Talking about CivitAI, i have feeling that externals have very different results with same settings so not sure about them. And have no idea what scheduler Civit's Euler A using (upd: it's automatic)
It's RNG anyways :) But AI generating and LORA training is all about trial and error, experiments, tests. So why not.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.