Trained for 6 epochs and 200 steps.

Trained with datasets ['4o-training-embeds-thinned', '4o-training-images-thinned']
Learning rate 0.0001, batch size 2, and 40 gradient accumulation steps.
Used DDPM noise scheduler for training with epsilon prediction type and rescaled_betas_zero_snr=False
Using 'trailing' timestep spacing.
Base model: PixArt-alpha/PixArt-Sigma-XL-2-1024-MS
VAE: madebyollin/sdxl-vae-fp16-fix

README.md

@@ -56,8 +56,8 @@ You may reuse the base model text encoder for inference.
 
 ## Training settings
 
-- Training epochs: 3
-- Training steps: 100
+- Training epochs: 6
+- Training steps: 200
 - Learning rate: 0.0001
 - Effective batch size: 160
   - Micro-batch size: 2

optimizer.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91b34e3ecb83bd412a30a6ce6b4d6abc7d4d644672b50cecd71c2bf385255f01
+oid sha256:2978bc167b75923e55aefe4aebd87d6519a850f629a655f1a2d09dbc9bb88007
 size 3665677155

random_states_0.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:89eff11307725d964676a11b97b677c0871dcb58d799618b6e4b5d6bae53b92f
-size 14604
+oid sha256:0f37c4c9bcca7b5a99eeeb7fcc0eddfdffc750ee230871a5fa321b2e133fb19d
+size 14668

scheduler.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ba77877730b79876fc3e1695338ca017ac9048f044e8311f771f5ace8bdd67fb
+oid sha256:5d77007765f1aa3b545d69cebf036aa1404b5b490fa4e3443586d3b52269cbf3
 size 1000

training_state.json

@@ -1 +1 @@
-{"global_step": 100, "epoch_step": 100, "epoch": 4, "exhausted_backends": [], "repeats": {"4o-training-images-thinned": 0}}
+{"global_step": 200, "epoch_step": 200, "epoch": 7, "exhausted_backends": [], "repeats": {"4o-training-images-thinned": 0}}

transformer/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7f3c20a20ad6785eb582affec1cec5f06bdb9cfc1bac236b06d1fc49b4ad8930
+oid sha256:d28d1a92bce803ce8cbab5afbfc0d129f4464f2bccf50473fae21557b6f52179
 size 1221780352