Train on images, and use on videos

Hi !
First of all, thank you for the amazing work :)
I have a question regarding the architecture of SAM3 and its training.
When I look at the diagram you provided, I have the intuition that the new Detector module is responsible of an independent task, which is detecting masks from (text) prompts and image embedding. If it is really the case, would it possible, and make sense, to fine-tune SAM3 on images only, to update weights of the text encoder and detector only, and use/plug these updated modules for video inference, along with the tracker and memory bank ? Or should the finetuning of tracker and memory bank also be considered ?
In my use case, we do not have labeled videos yet, only images, so it would make sense to train SAM3 on images only, and be able to still infer efficiently on videos.
I hope my question is clear enough, and I thank you in advance for your answer.

Start with the SAM3 architecture diagram and the repository’s fine-tuning and video-inference documentation or example notebooks. Determine whether image-only fine-tuning of the text encoder and detector can be reused with the tracker and memory bank, and document the supported workflow or any required video training; no file or test is named.

Written by the indexing model from the issue text.