Available in FLUX (NEW!!!), Illustrious(NEW!!!), Pony XL V6, SDXL AND SD 1.5 versions!!!!
I looked far and wide for Rugrats LoRa's. While I found some, they were not what I was looking for. So, I went ahead and said "Fine, I'll do it myself" and made my own.
All these LoRas were trained with just 20 images. The XL and Pony XL V6 versions were trained with 1024x1024 images using their respective base models. The "AngOne" SD 1.5 version was trained with 512x512 images downscaled from 1024x1024 (the same images used in my "AngXL" Lora) and the Revanimated SD 1.5 model. Every image used is based on official art of the character. Most available images of Angelica online are in poor quality, so I redrew every single image I could find to 300 dpi quality in Procreate by hand. Let's say, the results speak for themselves.
This LoRa is based on the original version of Angelica that debuted in the animated series "Rugrats" in 1991. No fan artwork was used, nor were images based on the character in the spin offs such as "Preschool Daze", "All Grown Up" or the 2021 reboot. I hope to eventually do all of the major Rugrats character in LoRa's one day, so I'm starting out with Angelica as she is my favorite.
For best results
FLUX
Prompt should have angflx in the form of natural language. Example "A angflx little girl with blonde hair..."
1024x1024, 832x1152 works well.
Natural language is key getting good results. Danbooru-style tags won't cut it.
Prompt should follow like this structure: "Description of style, angflx a little girl with blonde hair and blue eyes wearing a purple dress, orange sleeves, blue leggings with green dots, orange socks and purple sneakers. Her hair is tied up in two pigtails with purple bows. Action and description of scene." An actual prompt example would be "A 3d, cgi, detailed rendering in the style of Disney Animation. angflx a little girl with blonde hair and black eyes wearing a purple dress, orange sleeves, blue leggings with green dots, orange socks and purple sneakers. Her hair is tied up in two pigtails with purple bows. She is sitting on the ground while eating chocolate. Chocolate is on her face and clothing. "
I find describing the style at first tends to yield good results.
Strength model from .60 and up and strength clip 2 in ComfyUI will tend to create a more show-accurate render.
Tested with the base FLUX.1 Dev model and fp8 variant. It should work with quantized models, but I haven't done too much testing. Not much testing with Schnell.
Illustrious
Prompt should have an3x1
Tested with ComfyUI exclusively. I have not tested this in Forge/A1111 and I don't think it will work there anyway.
1024x1024 and 832x1216 works well.
For best results, prompt should have "an3x1, 1girl, solo, open mouth, pinafore dress, hair bow, full body, blue leggings, polka dot, long sleeves, purple footwear, orange socks, short twintails, blonde hair, black eyes," I find this will adhere mostly to the style of the checkpoint at around .50 strength. At .70 it will kinda get near show-accurate, but not like other models such as Pony or FLUX.
Adding "masterpiece, best quality, detailed" at the start of the positive prompt tends to yield good results.
Upscaling using the "Ultimate SD Upscale" node works great with 4xUltraSharp. I haven't tested with many of the other upscale models.
Tested with the base Illustrious and NoobAI vPred10 models. It should work with other models, but I haven't tested those out.
AngXL (Pony XL V6)
Prompt should have AngXL.
Tested with ComfyUI exclusively, so your results may vary with Automatic 1111. 1024x1024 and 896x1152 resolution works best.
Adding "score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_cartoon, BREAK, rating_safe, BREAK", is crucial to getting good results due to the nature of how the models were trained. Natural prompting appears to work well.
If you want to keep the generations close to show accurate, DO NOT add "1girl, solo, human, female, child" to the prompt. This prompt example below will help keep it accurate as possible. Adding it right BEFORE "AngXL" may give more detailed results.
Example prompt: "score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_cartoon, BREAK, rating_safe, BREAK, AngXL, a little girl with blonde hair styled in pigtails with bows, wearing a purple dress with orange sleeves, blue leggings with green dots and purple sneakers"
In "Load LoRA" (ComfyUI) for the Angelica Pony XL lora, I recommend a strength model of 0.70-1 and strength clip of 2. You can go as low as 1 for the clip to get more interesting results. If you're looking to make Angelica look older, I recommend pairing with another style LoRA such as Smooth Anime 2 Style and add "child, little girl" in the negative prompt.
Stick to Euler Karras with a cfg of 7-7.5 and 20 steps.
Tested with models Pony Diffusion V6 XL (V6 Start with this one) and AutismMix SDXL (AutismMix_confetti).
Use Confetti Pony autism mix for better results. While I trained with Pony Diffusion V6 XL, but I find better results come from the "autism mix" Checkpoint.
AngXL (SDXL)
Prompt should have AngXL
Use natural language in the prompt. (no danbooru tags).
A weight of 0.69-75 looks more show accurate. (Example, AngXL:0.75 in Automatic1111)
A weight around .60 and lower will fit closer to the model style.
To match how the character looks, it's best to start out a prompt with " a cartoon little girl with blonde hair, pigtails and dress"
Keep in mind, details such as her hair bows and polka dots on her tights may always not come out right and running several generations may be needed to get the desired result. Inpainting also helps.
AngOne (SD 1.5)
Prompt should have AngOne
Use Danbooru tags. (see examples)
Use a weight between 0.65 to 0.80. Higher the weight, the more closer it will look show accurate. A lower weight may be desired to reflect the model you're using. I would experiment.
I recommend generating images in 512x512 or 512x768. If you decide to use Hirez fix, I recommend lowering the denosing to around 0.40-60 and set upscaler to 4xAnimeSharp.
Adding to the prompt "bangs, blonde hair, pinafore dress, purple dress, orange long sleeves, bow, short twintails, hair bow, blue and green polka dot leggings, purple footwear" should get you an image that will have much of the core elements of the character.
Keep in mind, details such as her hair bows and polka dots on her tights may always not come out right and running several generations may be needed to get the desired result. Inpainting also helps.
Have fun and remember, please do not use this LoRa for unsavory purposes, you dumb babies!
Unfortunately, as I do not own the rights to this character, this LoRA will NOT be available on the Civitai generation service.
Description
The initial version as a bit too big, so I updated the Lora with a network rank of 128. Previous version used 256, taking nearly 1.7 GB! This one is half the size and should give the same results.
FAQ
Comments (10)
Did you make a mistake or something? Why or how is this lora 1.7GB...
Yeah...it's big. The type of training in koyha probably makes it so.
@ronaldhennessy It's probably your network rank. 256 is overkill imo. Higher rank means more parameters, but at the cost of file size and high VRAM usage. I also read you get diminishing returns the higher you go. Majority of people do 32/1, 64/1 and the max I've seen recommended is 128/1. Your lora is the same size as some full checkpoints.
Not arguing if it's a good or bad model, I'm sure it's very good and the images look great, but I feel like when a lora is closer to the size of a checkpoint, then something isn't right. The goal should be to aim for the lowest possible lora size (in the MB, not GB) whilst still retaining detail of whatever you're training.
Yes this IS WAY too big
You're right. The network rank was a BIT too high at 256. Going to work on an updated version with a lower network rank. Thanks for the feedback.
@sp00ns for a character Lora I usually do 16/8.
Hate to butt in but what's a good manual or guide for the settings? I'm glad I read this comment but I don't want to have to randomly stumble into basic information. I don't think any of the SDXL guides I read ever mentioned network rank, and I was always curious why so many XL loras were 200 megs (including the result of my own dabbling) when others had more reasonable sizes. I guess the default network rank of 32 results in that size.
@dudemansir There isn't really a one-size-fits-all guide. You kind of just pick bits of information up from everywhere. The rank essentially dictates the complexity of the adaptations your model can make, with a higher rank allowing it to learn more detailed modifications at the cost of added parameters, which are great, but the downside is that it bloats the model by a considerable amount, and the vram usage goes up with diminishing returns the higher you go.
Finding the "sweet spot" where the model retains the detail at the lowest size possible is the goal(imo). It's a balancing act, and experimentation is usually the only way. From my own observations, as mentioned above, I see 32, 64, and 128 mostly, but as FatsIncognito pointed out, they use 16/8 for characters, which I've never tried personally but clearly works for them and would result in a lower file size than the ranks mentioned above. You just have to find something that works for you, depending on your specific situation and needs. I'd argue dataset and tagging are the most important to training because the AI will destroy all your epochs if tagging is bad regardless if you have the best or most optimized settings possible.
The smaller the network size, the less details LoRA will remember
My opinion is that LoRA with smaller size than Dim 8 Alpha 4 (50Mb) is crazy, well SDXL LoRA cannot be smaller than SD v1.5 LoRA.
ㅤ
SDXL Checkpoint 6.46 Gb and LoRA size 1.2-14Mb is very funny, Yes saving space on HDD/SSD is good but when LoRA quality becomes bad it is not good.
ㅤ
For Characters
Dim 32 Alpha 16 (200Mb) or Dim 16 Alpha 8 (100Mb) You can of course use (50Mb) but I like the result with (100Mb) better, it's just my opinion, it's up to you to decide.
For Styles
Dim 32 Alpha 16 (200Mb) or Dim 16 Alpha 8 (100Mb) and Dim 8 Alpha 4 (50Mb)
ㅤ
With styles a little easier you can choose between 200Mb, 100Mb, 50Mb, you can even train with different settings and choose for yourself the best option
ㅤ
And Сaptions are very important the better the Сaptions the better the LoRA will be but there should not be too many of them
ㅤ
Example
charactername, 1boy/1girl, solo, male focus (for 1boy), eye color, hair color, character outfit/jacket/pants/boots/and other, looking at viewer, full body/upper body/cowboy shot, standing/sitting/lying, outdoors/indoors, and other
I use WD14 Captioning in Kohya_ss and edit captions if anything is not accurate.
Black screen in comfui