Continues Editing with Sequential Text Prompts

[JianXimple]

Summary:

I've introduced a new feature to the instruct-nerf2nerf project that enables continuesly editing 3D NeRF scenes using text prompts. This function allows for the loading of previously edited NeRF scenes, re-rendering them from specific locations and poses to match the original dataset. And then use the next prompt to edit the new scene. The key here is the ability to perform continuous and sequential text prompt editing.

Detailed Explanation:

Feature Overview:

• Scene Loading: This function can load an edited 3D NeRF scene.
• Re-rendering Capability: It re-renders images corresponding to the locations and poses from the original dataset.
• Sequential Text Prompt Editing: Users can now edit scenes continuously with text prompts. This is particularly useful for complex long prompts, where edits can be applied in a sequential manner.

Use Cases:

• Sequential Editing: Ideal for complex scene modifications where a single text prompt may not suffice.
• Iterative Creative Process: Artists and creators can now iteratively refine their scenes, adding depth and detail with each edit.

Implementation:

• For simplicity, I changed the train validation split ratio from the nerf library to 1.0. The original value is 0.9. Using 0.9 will cause the index of the images to shuffle, which can be solve by using a dictionary to store the index and the original file name.
• For naming the new images, I set it to be current_index:02d, which needs to change if the total number of images exceeds 100.

Results of Continues Editing:

I believe this feature will be a valuable addition to the instruct-nerf2nerf project, offering users more control and creative freedom in their 3D scene generation and editing endeavors.

[ayaanzhaque]

Thanks for these contributions! I recommend that for any additional changes to Instruct-NeRF2NeRF, you create a new repo with your changes. Here's an example: https://github.com/cvachha/instruct-gs2gs. I'll go ahead and close this PR for now.