Aiko Umesawa ( Danganronpa ) [SD 1.5]版本V1 (ID: 921975)

Aiko Umesawa from the Danganronpa 3 anime. AKA Pikachu girl, AKA the student council member from the Danganronpa 3 anime with the yellow hoodie and frying pan. This began as an experiment to see how well I could do with a severely limited amount of training data. Then I moved it from flux to sdxl, which suffers even more from lower data sources. Then to 1.5 when has an even more difficult time? It's not overly surprising that this lora has issues. But even with the problems I think it's acceptable. Just keep the experimental nature of this lora in mind when using it. The fact that it's usable in the first place is more surprising than it being a little inconsistent. She was in one scene of one episode of the Danganronpa 3 anime - that's not a lot of high-quality training data!

IMPORTANT: This stable diffusion 1.5 lora is trained on a conversion my original flux Aiko Umesawa training data. In short, this sd 1.5 lora suffers a bit from being designed for a very, very, different framework. Still, the testing seems to be giving acceptable results. That said, there are issues. The biggest is that you need to really do some negative prompt wrangling to keep GUI elements from bleeding into your images. You'll find more information on that later in this description.

It's probably going to be a bit unpredictable for you. For models, I've had the best results for this lora with lustermix 2D and ghostmix. Juggernaut reborn does a good job converting her into a photorealistic, or at least close to photorealistic, style by just adding "photo, RAW photo" tags to the prompt. Additionally, with all the models I typically used adetailer for the faces. Finally, the tagging is messy and there's a lot of redundancy. The price paid for semi-automated conversion. I'm planning on rewriting my code soon to make the output a bit cleaner.

Unique features: The main point I try to stress in my Danganronpa loras is authenticity to the source material. It's fairly easy to get the basics of a character's style or clothing down. Characterization is a lot harder. Models have their own concepts of what a scared, or happy, or sad, or angry expression or stance would be. But a character in something like Dananronpa has a lot of careful attention put into ensuring their designs make all of those states unique. I go through every sprite, every CG, every official image I can find to caption and include them in the dataset in hope of best representing that. Whether it works out is another question, but it's what I'm aiming for and why these might be a little odd. And the unpredictability goes double for the SD 1.5 conversions. However, I think they provide a worthwhile take on the characters. With Aiko Umesawa, one thing I was really happy to see was that she sometimes clutches the strings of her hoodie when she has a stressed expression. That's an association from the training data, and it's exactly the kind of connection I'm trying to preserve with my methodology. The characters should preserve characterization, you know?

I trained on a few specific image types which you can replicate, to varying levels of success. This tends to work quite well in flux, OK with SDXL, and generally fairly poorly in SD 1.5. I tried to include a range of the prompt options in the images so you can easily compare and contrast. But these are the major options.

NEGATIVE PROMPT - To push the system away from game text and gui generation, add this to your negative prompt: (((danganronpa S GUI style speech bubble))), (((name bubble))), (((speech bubble))), blue speech bubble, text box:danganronpa S GUI style, GUI style, speech bubble, danganronpa S GUI style, ((no people)), (((text))), (((gui)))

Danganronpa the Animation: To try steering images into the style used in the Danganronpa the Animation, use the tags 'Danganronpa The Animation, screencap from the anime Danganronpa The Animation, anime, dr1anime' in the prompt to emulate the art style from the anime series Danganronpa the animation.

Danganronpa 3 (anime) style: Use "Danganronpa 3 screencap style, danganronpa 3, anime"

Sprites from Danganronpa: Use the tags 'danganronpa sprite style' in the prompt to try emulating the art style of Danganronpa. Despite the name, I trained on full screenshots, with and without the GUI, with this description along with character sprite rips. So in theory it should be able to 'danganronpafy' a full image. This was complicated a bit by the fact that Aiko Umesawa isn't in any of the games. But I did what I could.

Danganronpa S GUI: You can give it a try with this. "danganronpa s, danganronpa sprite style, danganronpa S sprite style, danganronpa s gui style dialog box, Danganronpa S GUI status bar, black text, Danganronpa S GUI title bar"

Clothing: You can try different clothing options with prompts like the following:

Standard outfit with hood up: "Aiko Umesawa, yellow bunny hoodie, white button up shirt, red ribbon, light brown pleated skirt, black ankle socks, white slippers, blushing, hood up, green left eye, blue right eye"

Standard outfit with hood down: "Aiko Umesawa, yellow bunny hoodie, white button up shirt, red ribbon, light brown pleated skirt, black ankle socks, white slippers, blushing, hood down, green left eye, blue right eye"

Locations: I didn't try to train on any specific locations, but some of them appeared enough in the training data that you might be able to improve your results by specifying areas and tags from danganronpa and danganronpa 2/danganronpa S. In particular with the tags 'jabberwock island' and 'Hope's Peak'. I use the location names from the maps at danganronpa-gaming.proboards.

Note on her eyes: Aiko has heterochromia but it can be a little inconsistent in the generations. If the eyes aren't coming out correctly you can try giving it more of a push with tags along the lines of " (((heterochromia))), (((green left eye))), (((blue right eye))) ". Using an additional lora, like JujoHotaru's Heterochromia Helper, might help as well. However, the larger issue just seems to be that some models are more flexible with heterochromia than others. If you want to try Heterochromia Helper, find the hetechro_BG_v100.safetensors lora in JujoHotaru's zip and call it with "<lora:hetechro_BG_v100:1>, heterochromia with blue and green". The downside to the heterochromia lora is that the blue/green tends to push 'everything' into blue or green. So tan skirt? Good chance of turning into a green skirt. And generally, everything gets hit by that stylistic coin flip.

Training material: Saying that there's not a lot of official material on Aiko is an understatement. She's in one short scene of one episode of Danganronpa 3. And there's not a huge amount of fan material to fill in the gaps either. I took screencaps of every mostly-unique frame she was in. Then heavily supplemented that with sprites made by madara120. I topped all that off with selections of material from my other danganronpa loras to lend the styles and location information.

Final thoughts: I made the original flux lora as something of an experiment. Just how well would this process hold up under limited data? Such as a character who was visually unique, but only in a very small number of samples. The big point I've taken away from the SD 1.5 side of things is that it requires shifting material around in the training data to compensate. In particular if I was doing this again I'd get rid of most training data that has text. The gain in possible sprite style just isn't worth how easily 1.5 decides that because it's in the data so much we must always want that. Hence why the negative prompts are needed in order to avoid GUI elements. As with most of the loras I've made, I've noted those issues and will have an updated lora eventually. Though there's a whole lot of cast still waiting to be done before I get around to revisions!

描述:

训练词语: