[v3.1 is still in further testing. Updates regarding new findings will be updated in the "About this version" section]
UrangDiffusion v3.1 (oo-raw-ng Diffusion) is the first UrangDiffusion version that utilize Animagine XL 4.0 as the base.
The name “Urang” comes from Sundanese, meaning “We/Our/I.” The history behind the name is to make the model not only suitable for me but also for many people. Another reason is that I use many resources (training scripts, dataset collecting scripts, etc.) from other people. It’s unfair to claim this model as “my sole work.”
Standard Prompting Guidelines
Prompting guide:
Default negative prompt:
lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurryDefault configuration: Euler a with around 25-30 steps, CFG 5-7, and ENSD set to 31337. Sweet spot is around 28 steps and CFG 6.
Training Configurations
Finetuned from: Animagine XL 4.0 Base (NOT 4.0-Zero)
Finetuning:
Dataset size: ~1,600 images
GPU: 1xA100 80GB
Optimizer: AdaFactor
Unet Learning Rate: 1.25e-6
Text Encoder Learning Rate: N/A (Turned off)
Batch Size: 48
Gradient Accumulation: 1
Warmup steps: 5%
Min SNR: 5
Epoch: 15
FAQ
Q: Images are sometimes noisy.
A: This is a common issue with Animagine XL 4.0 models in general. The base model is trained with only 10 epochs, which lead to the model being undertrained. Unlike Initial N or Initial I model that are trained with more resources.
Q: Hires fix model?
A: Check out the cover image metadata, you'll find it there.
Q: Initial N/Initial I is better.
A: Just leave and do not use the model. Simple. No need to announce your departure. Except you're willing to leave a constructive feedback or willing to fund future projects.
Special Thanks
My co-workers(?) at CagliostroLab for the insights and feedback.
Nur Hikari and Vanilla Latte for quality control.
Linaqruf, my tutor and role model in AI-generated images, and also the person behind tag ordering.
License
UrangDiffusion v1.0-v2.5 falls under the Fair AI Public License 1.0-SD license, while v3.x falls under the CreativeML OpenRAIL++-M license.
Description
Initial release.
===
Training Configurations
Pretraining:
- Dataset size: ~17,200 images
- GPU: 1xA100
- Optimizer: AdaFactor
- Unet Learning Rate: 2.5e-6
- Text Encoder Learning Rate: 1.25e-6
- Batch Size: 48
- Gradient Accumulation: 1
- Epoch: 10 (epoch 8 is used)
Finetuning:
- Dataset size: ~1,300 images
- GPU: 1xA100
- Optimizer: AdaFactor
- Unet Learning Rate: 2e-6
- Text Encoder Learning Rate: - (Train TE set to False)
- Batch Size: 48
- Gradient Accumulation: 1
- Epoch: 10 (epoch 8 is used)
Added Series
Wuthering Waves and hololiveEN Justice
FAQ
Comments (10)
This model sucks Ass, bring back Aingdiffusion xl.
Hi, there are a dozen reasons not to bring back AingDiffusion XL:
- LoRA/Lokr merging into the model makes it unstable in generation.
- The learning rate used is too high, causing most of the generations to be burned.
- No new series and no new styles were trained into the model; it's basically Animagine XL 3.1 with extra steps.
- Using AdamW optimizer, which has been proven to contribute to artifact issues in the model.
- And many more reasons not to bring it back.
The dataset used for UrangDiffusion is still very small compared to AingDiffusion XL's 40k dataset due to limited funding, since full training is more expensive than Lokr training. Please give me your reasons for saying the model sucks, so I can fix it in later versions. Saying that the model is "ass" doesn't help me with anything.
Thank you.
@kayfahaarukku Alright, maybe my comment was a bit toxic. So, I'm truely and sincerely sorry about it. This model is not bad. But, the older one managed to create what I could say THE best AI illustrations I ever experienced, among many tools not only it had the best AI art quality, but also from it's simplicity (I'm terrible at adding Lora and Embedding, and let's be honest Civitai tutorials are not helping either). Plus, I'm not gonna pretend to be a programmer, I'm not. downloading the model and it's perpous is still a mystery to me, and I couldn't understand half of the reasons you sent. However, I get it nothing can be done, and I understand. But, hopefully if you could manage to make this version's experience anywhere near as the older one in future updates, it will mean so much to me and others who appreciate your work.
Best regards, and again sorry, just got overwhelmed and it came out wrong.
@Dean7896 It's okay, I know it's not your true intention to say that.
With all the plus points that UrangDiffusion has over AingDiffusion XL, I admit that UrangDiffusion is lacking something, though I don't know what.
I think it's mostly because of the smaller dataset I used. I will try to use a larger dataset for the next version.
If you can give me a little hint about what makes you think AingDiffusion XL is superior compared to UrangDiffusion (e.g., it loses certain styles, some characters don't generate properly, etc.), I will try to address it in the next version. You can DM me if you're uncomfortable saying it in public.
Thank you.
@kayfahaarukku The few issues I had is that I didn't manage to create illustrations with my favourite art style: mix between anime, semi-3d or semi-realism, and the right amount of brightness (best anime art style hands down can't be proofing wrong), It's optional as a LoRA but it didn't give it justice at all. Then, the colours are a bit pale or bland in comparison with the older versions. And finally, I think some prompts are being ignored when generating (specially from the negative prompts bar) either that or it just ruins the overall image, I think it happens when using too many prompt at the same time. But, it didn't used to be a problem before.
-Anyway, that concludes my end, for "what I think is missing in the Urangdiffusion". But, this is only a personal opinion. I'm sure there are others that can give better details, that are better in this domain. I'm not a programmer nor an artist I just like some good art and yours was the only one who could manage to deliver. That's why I'm gonna just put faith in the process.
Thank you for your response, keep up the great work, and have a good day.
@Dean7896 Ah I see, I fine-tuned the model using anime screenshots to achieve a default 2D aesthetic. However, this can be overridden by including specific artist names. To get your desired style, I suggest researching and incorporating artists listed in danbooru who work in that style into your prompts. This is currently the best solution I can offer, at least until either I release the next version, or the "not finetuned" version of the model.
It seems to me that the model is still quite raw. Comparing the pictures with kita ikuyo shows that the model is some kind of blurry and more greedy for details. I wish you good luck in your further training and hope that the next version will be a cut above, but for now, unfortunately, this is a step back, but I understand that AingDiffusion XL was abandoned for a reason. I'm sorry for my English.
Hi, I've been through a busy week and just read your comments. Thank you for the feedback.
I realize that version 1.0 is not perfect yet, since I've interrupted the training several times due to faulty configuration. I also just found out that I forgot to turn on several settings during fine-tuning, which might contribute to the "blurry" issue you're talking about.
What I can say right now is that fixes for these issues are coming soon. Thank you for waiting.
@kayfahaarukku I believe in you, pal
Just did my first image, for a 1.0 it's not bad. Definitely will only get better as time goes on, very much looking forward to it. Especially since your pretty much my go to model author.