modelscope
Question about optimal hyper parameters for training Qwen-Image-Edit-2509
DiffSynth is great. I am trying to train Qwen-Image-Edit-2509. I have a dataset of 5.4k examples. I'm curious if full parameter fine-tuning could be feasible with this dataset size, and if so at what hyperparameters? I've tried 1e-5 and 5e-6 at a 128 effective batch size and the model degrades when I evaluate it.
Base model:
Step 1000 after training (clearly degraded in quality):
The loss barely budges:
The next thing I'll try is LoRA, but I'm just confused how I couldn't get the training to have any positive impact on the model at any point across a variety of training runs. Curious if anyone, besides the Qwen team, knows how to successfully fine-tune this model. Thanks!
@dylanzonix We have also observed similar issues: the Qwen-Image-Edit and Qwen-Image-Edit-2509 models are difficult to fine-tune, despite following exactly the same mathematical principles as Qwen-Image. If we find a better improvement approach, we will update the code accordingly.
@dylanzonix We have also observed similar issues: the Qwen-Image-Edit and Qwen-Image-Edit-2509 models are difficult to fine-tune, despite following exactly the same mathematical principles as Qwen-Image. If we find a better improvement approach, we will update the code accordingly.
Thanks a lot. I am still actively working on this as well and can give updates from my end as well. I'm also happy to share with you the dataset I am using if you are interested. It's custom made exactly for Qwen-Image-Edit-2509
I also encounter the same problem: the loss intend to oscillation instead of converging when fine tuning qwen-image-edit and qwen-image-edit-2509 with lora.
Training the LoRa model for Qwen Image Edit 2509 using ModelScope significantly reduced the loss. However, training with DiffSyhtn resulted in a persistently oscillating loss.
Training the LoRa model for Qwen Image Edit 2509 using ModelScope significantly reduced the loss. However, training with DiffSyhtn resulted in a persistently oscillating loss.
Interesting. You think the issue could be with the DiffSynth library's implementation?
Training the LoRa model for Qwen Image Edit 2509 using ModelScope significantly reduced the loss. However, training with DiffSyhtn resulted in a persistently oscillating loss.
Interesting. You think the issue could be with the DiffSynth library's implementation?
Has the problem been solved?
