|
--- |
|
tags: |
|
- stable diffusion |
|
--- |
|
 |
|
|
|
[[Open Grid]](https://files.catbox.moe/otgslr.png) | [[Open CivitAI](https://civitai.com/models/272058)] |
|
|
|
Here I publish results of my experiments and my subjective opinion about hyperparams for PonyXL while I tried to make Senko-lora for that. |
|
|
|
Training was done using 120 images, 20 epochs, 6600 steps totally. I checked only the last epochs. |
|
|
|
Batch Size == 1 |
|
TE LR == UNet LR |
|
I didn't use gradient checkpointing. |
|
|
|
--- |
|
❓ [`<lora:senko_ds6_ponyxl_lr1_linear_prodigy_dim16_alpha8:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr1_linear_prodigy_dim16_alpha8.safetensors) overfit |
|
❓ [`<lora:senko_ds6_ponyxl_lr1_linear_prodigy_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr1_linear_prodigy_dim32_alpha16.safetensors) overfit |
|
|
|
Prodigy works a bit worse with artists tags and style loras, but still look good. Probably it can be useful if you are fine with the default style lora produces. |
|
|
|
--- |
|
❓ [`<lora:senko_ds6_ponyxl_lr1e-4_constant_adamw8_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr1e-4_constant_adamw8_dim32_alpha16.safetensors) overfit |
|
❌ [`<lora:senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim16_alpha1:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim16_alpha1.safetensors) doesn't work |
|
✅ [`<lora:senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim16_alpha8:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim16_alpha8.safetensors) OK |
|
✅ [`<lora:senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr1e-5_constant_adamw8_dim32_alpha16.safetensors) OK (published as [senko-ponyxl-v2](https://civitai.com/models/272058/model-versions/306696/wizard?step=3)) |
|
|
|
Adam 1e-4 bakes the style from dataset what is noticable on some grids. While I tried to make lora using LR == 1e-3, loss become equal to 1 at second epoch so I stropped training. But afair I used dim32/alpha16 which won't work with such high LR at all, so it was my mistake. |
|
|
|
--- |
|
✅ [`<lora:senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim16_alpha1:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim16_alpha1.safetensors) OK |
|
✅ [`<lora:senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim16_alpha8:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim16_alpha8.safetensors) OK |
|
✅ [`<lora:senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_lr3e-4_constant_adafactor_dim32_alpha16.safetensors) OK (published as [senko-ponyxl-v1](https://civitai.com/models/272058?modelVersionId=306653)) |
|
|
|
Adafactor LR == 3e-4 works fine with different dim/alpha params. |
|
|
|
--- |
|
❓ [`<lora:senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim16_alpha8_conv16_convalpha_8:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim16_alpha8_conv16_convalpha_8.safetensors) breaks anatomy on complex concepts |
|
❓ [`<lora:senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim16_alpha8_conv32_convalpha_16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim16_alpha8_conv32_convalpha_16.safetensors) TE overfit |
|
❓ [`<lora:senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim32_alpha16_conv16_convalpha_8:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim32_alpha16_conv16_convalpha_8.safetensors) TE overfit |
|
❓ [`<lora:senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim32_alpha16_conv32_convalpha_16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_ponyxl_locon_lr1_linear_prodigy_dim32_alpha16_conv32_convalpha_16.safetensors) TE overfit |
|
|
|
I didn't find good hyperparams for locon with prodigy optimizer - it breaks anatomy or doesn't care about prompt at all. |
|
|
|
--- |
|
❌ [`<lora:senko_ds6_sdxl_lr1e-5_constant_adamw8_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_sdxl_lr1e-5_constant_adamw8_dim32_alpha16.safetensors) doesn't work |
|
❌ [`<lora:senko_ds6_sdxl_lr3e-4_constant_adafactor_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_sdxl_lr3e-4_constant_adafactor_dim32_alpha16.safetensors) doesn't work |
|
❌ [`<lora:senko_ds6_counterfeitxl_lr1e-5_constant_adamw8_dim32_alpha16:1>`](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/blob/main/ds6/senko_ds6_counterfeitxl_lr1e-5_constant_adamw8_dim32_alpha16.safetensors) doesn't work |
|
|
|
The lora I made using SDXL checkpoint doesn't work, the same for lora trained on CounterfeitXL. |
