Open chongxian opened 5 months ago
Thank you very much for recognizing our work.😄
In fact, I have already been training the SDXL version of MIGC. However, the SDXL model has a larger number of parameters compared to the SD1.x models, which will require more computational resources and necessitate further contemplation on additional training techniques. Additionally, the number of Cross-Attention layers in SDXL has been expanded to 70 based on SD1.x, necessitating an ablation study on the integration points of MIGC into SDXL.
In summary, training on SDXL will demand a considerable amount of effort. Once I have finished training, I will share it as soon as possible. Stay tuned.
the SD14 is not great in depicting even simple objects SDXL needed but the positional control is pretty good
@fritol You can try changing the base model to get better generation quality.
hello,thanks for your excellent work, will you release the pretrained weights of sdxl in the future?