Thank you for your excellent work. Do you support image editing tasks? How much VRAM is required for operation

[libai-lab]

Thank you for your excellent work. Do you support image editing tasks? How much VRAM is required for operation

[PengDa02]

Thanks for your interest in our work!

An interesting observation is that during the pretraining stage we did not specifically train the model on editing tasks, yet it still exhibits emergent editing capabilities. This suggests that unifying representations into the LLM feature space has the potential to support zero-shot generalization.

In addition, this work is a preliminary and somewhat bold attempt, and for now we are not focusing specifically on editing tasks. You can look forward to future versions. In the meantime, you may refer to the image input and output formats provided to try image editing tasks, or continue post-training to explore further possibilities.

[guozonghao96]

Thanks for paying attention to our works.
Our training framework supports image editing training. The training data format is
{"id": "image_editing_01", "image": "image_src_path, image_target_path", "conversations": [{"from": "human", "value": "<im_start><im_end>\nRemove the apple?"}, {"from": "gpt", "value": "<im_start><image_gen><im_end>"}]}