Trained for 28 epochs and 200 steps.

Trained with datasets ['text-embeds', 'guimiv3']
Learning rate 0.0002, batch size 8, and 1 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: /data/yihao/FLUX.1-dev
VAE: None

pytorch_lora_weights.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e796cbc8eecf1afca260ecfcaaecdcf18c18b2370eb3b7699adb92d7e04060be
-size 89841632
+oid sha256:52d7dda838c03a24deaa37b6a47729e8a271fab3bf01d515a4ca753a097b90b1
+size 209263048