Commit
•
ab51240
1
Parent(s):
fa18cd2
Update README.md
Browse files
README.md
CHANGED
|
@@ -7,29 +7,46 @@ language:
|
|
| 7 |
|
| 8 |
[[Open Grid]](https://files.catbox.moe/otgslr.png) | [[Open CivitAI](https://civitai.com/models/272058)] | [[Show models on HuggingFace](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/tree/main/ds6)]
|
| 9 |
|
| 10 |
-
Here I publish
|
|
|
|
|
|
|
| 11 |
|
| 12 |
Batch Size == 1
|
| 13 |
TE LR == UNet LR
|
| 14 |
I didn't use gradient checkpointing.
|
| 15 |
|
| 16 |
-
|
| 17 |
-
[`<lora:
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
-
|
| 20 |
-
[`<lora:senko_ds6_ponyxl_lr1e-
|
| 21 |
-
[`<lora:senko_ds6_ponyxl_lr1e-
|
| 22 |
-
[`<lora:senko_ds6_ponyxl_lr1e-
|
|
|
|
|
|
|
|
|
|
| 23 |
|
| 24 |
-
|
| 25 |
-
[`<lora:senko_ds6_ponyxl_lr3e-
|
| 26 |
-
[`<lora:senko_ds6_ponyxl_lr3e-
|
|
|
|
| 27 |
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
[`<lora:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
-
|
| 34 |
-
[`<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
|
| 35 |
-
[`<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
|
|
|
|
| 7 |
|
| 8 |
[[Open Grid]](https://files.catbox.moe/otgslr.png) | [[Open CivitAI](https://civitai.com/models/272058)] | [[Show models on HuggingFace](https://huggingface.co/NeuroSenko/senko-lora-ponyxl/tree/main/ds6)]
|
| 9 |
|
| 10 |
+
Here I publish results of my experiments and my subjective opinion about hyperparams for PonyXL while I tried to make Senko-lora for that.
|
| 11 |
+
|
| 12 |
+
Training was done using 120 images, 20 epochs, 6600 steps totally. I checked only the last epochs.
|
| 13 |
|
| 14 |
Batch Size == 1
|
| 15 |
TE LR == UNet LR
|
| 16 |
I didn't use gradient checkpointing.
|
| 17 |
|
| 18 |
+
---
|
| 19 |
+
❓ [`<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
|
| 20 |
+
❓ [`<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
|
| 21 |
+
|
| 22 |
+
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.
|
| 23 |
|
| 24 |
+
---
|
| 25 |
+
❓ [`<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
|
| 26 |
+
❌ [`<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
|
| 27 |
+
✅ [`<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
|
| 28 |
+
✅ [`<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))
|
| 29 |
+
|
| 30 |
+
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.
|
| 31 |
|
| 32 |
+
---
|
| 33 |
+
✅ [`<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
|
| 34 |
+
✅ [`<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
|
| 35 |
+
✅ [`<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))
|
| 36 |
|
| 37 |
+
Adafactor LR == 3e-4 works fine with different dim/alpha params.
|
| 38 |
+
|
| 39 |
+
---
|
| 40 |
+
❓ [`<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
|
| 41 |
+
❓ [`<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
|
| 42 |
+
❓ [`<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
|
| 43 |
+
❓ [`<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
|
| 44 |
+
|
| 45 |
+
I didn't find good hyperparams for locon with prodigy optimizer - it breaks anatomy or doesn't care about prompt at all.
|
| 46 |
+
|
| 47 |
+
---
|
| 48 |
+
❌ [`<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
|
| 49 |
+
❌ [`<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
|
| 50 |
+
❌ [`<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
|
| 51 |
|
| 52 |
+
The lora I made using SDXL checkpoint doesn't work, the same for lora trained on CounterfeitXL.
|
|
|
|
|
|