Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,27 @@
|
|
1 |
---
|
2 |
license: creativeml-openrail-m
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: creativeml-openrail-m
|
3 |
---
|
4 |
+
|
5 |
+
SD1.5 experiments with Huber and MSE loss. All models trained for 4 epochs on approximately 250k images from a variety of sources. Approximately half from LAION Aesthetics, and a few thousand 4K video rips with COG-VLM captions.
|
6 |
+
|
7 |
+
Trained using Everydream2 Trainer (https://github.com/victorchall/EveryDream2trainer) on an RTX 6000 Ada 48gb. Each epoch takes approximately 10 hours for a total of about 40 hours per model.
|
8 |
+
|
9 |
+
- Multi-aspect ratio trained with nominal size of <=768^2 pixels for each bucket
|
10 |
+
- Batch size 12 with grad accum 10.
|
11 |
+
- AdamW 8bit optimizer with standard betas of (0.9,0.999) and weight decay of 0.010.
|
12 |
+
- Trained with automatic mixed precision FP16 and TF32 matmul
|
13 |
+
- 3.0e-6 LR cosine schedule with a ~12 epoch target to decay, ending around 2.3e-6 at end of training
|
14 |
+
- Pyramid noise using discount 0.03
|
15 |
+
- Zero offset noise of 0.02
|
16 |
+
- Min SNR gamma of 5.0
|
17 |
+
- Unet only training, text encoder left frozen.
|
18 |
+
- Conditional dropout of 10%
|
19 |
+
|
20 |
+
The following models were produced:
|
21 |
+
- 768_huber.safetensors - Huber loss only
|
22 |
+
- 768_mse.safetensors - MSE loss only
|
23 |
+
- 768_ts0huber_ts999mse.safetensors - Huber loss at timestep 0 interpolated to MSE loss at timestep 999
|
24 |
+
- 768_ts0mse_ts999huber.safetensors - MSE loss at timestep 0 interpolated to Huber loss at timestep 999
|
25 |
+
|
26 |
+
Worth noting timestep 0 is the lowest-noise-added step and 999 is most noised timestep.
|
27 |
+
|