Commit
•
05fea99
1
Parent(s):
f33c75c
Add ReadMe description
Browse files
README.md
CHANGED
@@ -1,3 +1,20 @@
|
|
1 |
---
|
2 |
license: creativeml-openrail-m
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: creativeml-openrail-m
|
3 |
+
tags:
|
4 |
+
- stable-diffusion
|
5 |
+
- text-to-image
|
6 |
+
datasets:
|
7 |
+
- ProGamerGov/StableDiffusion-v1-5-Regularization-Images
|
8 |
---
|
9 |
+
|
10 |
+
**Min-Illust-Background-Diffusion**
|
11 |
+
|
12 |
+
This fine-tuned Stable Diffusion v1.5 model was trained for 2000 iterations with a batch size of 4, on a selection of photos of things taped to wall. Training was performed using [ShivamShrirao/diffusers](https://github.com/ShivamShrirao/diffusers) with full precision, prior-preservation loss, the train-text-encoder feature, and the new [1.5 MSE VAE from Stability AI](https://huggingface.co/stabilityai/sd-vae-ft-mse). A total of 2100 regularization / class images were used from [here](https://huggingface.co/datasets/ProGamerGov/StableDiffusion-v1-5-Regularization-Images). Regularization images were generated using the prompt "artwork style" with 50 DPM++ 2S a Karras steps and a CFG of 7, using the MSE VAE. A negative prompt of "text" was also used for this dataset.
|
13 |
+
|
14 |
+
Use the tokens **ttw style** in your prompts for the effect. Note that the effect also appears to occur at a much weaker strength on prompts that steer the output towards specific artistic styles.
|
15 |
+
|
16 |
+
This model will likely not perform well on generating portraits and related tasks, as the training data was primarily composed of landscapes.
|
17 |
+
|
18 |
+
|
19 |
+
|
20 |
+
|