multiple images editing

[Enderfga]

Hi, thanks for the great work!

I was wondering if the image editing module supports editing multiple images at once, or is it currently designed for single-image editing only?

If not supported yet, do you have plans to support batch editing in the future?

Thanks again!

[mi804]

The architecture of Nexus-Gen supports editing with an arbitrary number of input images. However, the current training dataset does not include samples for multi-image editing tasks. In the future, we will incorporate multi-image editing datasets and refine the model through additional training.

[mi804]

Hi, thanks for the great work!

I was wondering if the image editing module supports editing multiple images at once, or is it currently designed for single-image editing only?

If not supported yet, do you have plans to support batch editing in the future?

Thanks again!

Apologies for the misunderstanding earlier. Currently, Nexus-Gen does not yet support batch editing or generation, but this functionality is theoretically feasible from an engineering perspective. We will prioritize implementing it as soon as possible.

[Enderfga]

Thanks for the clarification!

As a quick follow-up:
I’m also curious whether multi-turn dialogue for a single image is supported — for example, asking a follow-up question based on the same image, or giving multiple step-by-step editing instructions.

From what I’ve seen, it seems this might not be currently supported.
Is this limitation also due to the lack of such training data in the dataset, or just not yet implemented?

Would love to hear your thoughts!

[mi804]

From training perspective, the multi-turn datasets in our training corpus exclusively cover image understanding tasks, while generation and editing tasks only contain single-turn dialogue samples.
From the inference standpoint, multi-turn dialogue and editing functionalities can be easily implemented, we have tested your proposed two cases:

1. Generation-Understanding: We observed that the model maintains coherent dialogue interactions with generated images despite lacking such samples in training data.
2. Multi-Turn Editing​: Though the model successfully outputs images in this scenario, it fails to complete editing. Consequently, for multi-turn editing workflows, we recommend employing iterative single-turn edits rather than attempting true multi-turn editing.

[itisfree]

thanks for your work
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
diffsynth 1.1.7 requires transformers==4.46.2, but you have transformers 4.49.0 which is incompatible.
how to solve this?