rockeycoss / spo Goto Github PK

I trained the model following your training code and found that the recorded loss curve was very weird. The train_loss always fluctuates violently around 0.693. Additionally, the performance of the trained model was very poor. Could you share your training records? Is there anything specific I need to pay attention to during training?

(venv) C:\AI\SPO>py inference_spo_sdxl.py
C:\AI\SPO\venv\Lib\site-packages\diffusers\utils\outputs.py:63: UserWarning: torch.utils._pytree._register_pytree_node is deprecated. Please use torch.utils._pytree.register_pytree_node instead.
torch.utils._pytree._register_pytree_node(
C:\AI\SPO\venv\Lib\site-packages\huggingface_hub\file_download.py:1132: FutureWarning: resume_download is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use force_download=True.
warnings.warn(
Loading pipeline components...: 100%|████████████████████████████████████████████████| 7/7 [00:00<00:00, 11.60it/s]
0%| | 0/50 [00:00<?, ?it/s]C:\AI\SPO\venv\Lib\site-packages\diffusers\models\attention_processor.py:1244: UserWarning: 1Torch was not compiled with flash attention. (Triggered internally at ..\aten\src\ATen\native\transformers\cuda\sdp_utils.cpp:455.)
hidden_states = F.scaled_dot_product_attention(

I have tried to launch it previously and it downloaded a ton of models.
Now it's stuck here seemingly forever with GPU spiking.

I'm on a 3070 with 8gb vram, maybe this is why?

Hi, thanks for your great work. But when I try to train SPO_SDXL with the default config file spo_sdxl_4k-prompts_num-sam-2_3-is_10ep_bs2_gradacc2.py using 4 80G A800 GPUs, there is an error showing CUDA out of memory. Can you help me confirm your training configuration?

Interesting, thanks for the XL model. Will you release a 1.5 model ?
Our experiments with Stable Diffusion v1.5 and SDXL demonstrate that SPO significantly outperforms the latest Diffusion-DPO in aligning generated images with complex, detailed prompts and enhancing aesthetics, while also achieving more than 20× times faster in training efficiency.

Hi, great work!
I see you will release training code. How about reward model dataset or reward model?

I found the webui version in the end, but still, really interested in the difference between these 2 versions, and what should we do to transfer one into another one?