This DoRA was trained on 600k images of hyper sized anime characters. It focus mainly on breasts/ass/belly/thighs/fat. This dataset is a subset of the larger hyperfusion dataset, but filtered down to body shape/size related images only. The full dataset would have taken more than a year to train on SDXL, lol.
Recomendations:
Dora/LoRA strength: 1.0 (DoRA's work in most WebUI's by now)
Resolution: ~1024px
samplers: any sampler PonyXL supports
in v10 you can push the lora weight more than in v9, so do that if your concepts are not working as well as you like.
Uploaded 1.4 million custom tags used in hyperfusion here for integrating into your own datasets
v10 Noob_vpred Release 2025/07/29:
Did you guys think I disappeared? Nope, just hopelessly training a model with a frozen text encoder for 7 months.
This new DoRA has the same concepts you are used to by now, but with a few new concepts as usual. Also 200k more images than v9.
This version is trained on NoobAI_Vpred, so there is no guarantee it will work with anything else. Especially not on non-v_pred models.
Wanted to try training with the Text Encoder frozen one last time. Also decided to stick to it no matter how long it took. And now I can definitively say I will be including TE in future models just for the sake of time. It works, but its way too slow for my setup.
Use the tag list in v9 for now, until I get around to building the new one with the small number of new concepts.
This one should handle concepts a little better than v9_sdxl, and is less prone to exploding gradients as well.
v9 Pony Release:
This model has been training for over 2 months now, but since Flux dropped, I decided to release what I have so far to free up a GPU. Technically it should have trained for longer, but I'm impatient, and some of you are probably tired of waiting anyway.
The tags are mostly the same as the last v8 release for SD1, with a few new additions like blob content for example. See the tag.csv for more in "Training Data".
Pony is a little tricky to train on, so I was experimenting a lot with this model. Because of this you should try to keep the DoRA strength near 1.0. Anything above 1.1 tends to explode. (weight regularization like scale_weight_norms is critical for training on pony, fyi)
To keep training time reasonable I trained at 768x768 resolution initially, and had planned on finishing up training with 1024px resolution, but then Flux happened. The results still seem reasonable.
I put plans and progress here every now and then.
Changelog Article Link
Description
Trained on 400k (hyper focused) images extracted from the larger hyperfusion dataset.
The big takeaway from training this version is that scale_weight_norms is critical to keep the model stable on Pony. If you remove it the model will quickly become unusable over the course of a few epochs. Results may differ for smaller datasets.
Training Notes:
~400k images
LR 3e-4
Unet only training (It might be slower but SDXL really doesn't need TE training in my experience)
batch 4
GA 16
dim 16
alpha 8
c_dim 8 (technically a DoRA LoCon)
c_alpha 4
optimizer: started with Adafactor, switched to Adamw8Bit half way.
scheduler: cosine
base model Pony
flip aug
375 token length
bucketing at 768 max 1024
tag drop chance 0.1
network dropout 0.25
tag shuffling
--min_snr_gamma 4
--ip_noise_gamma 0.02
--scale_weight_norms 1
custom code to drop out 75% of tags 5% of the time to hopefully improve short tag length results
15% CogVLM captions inserted 30% of the time.
over 60 days training time