SplatViz 1.1 Released

The latest update for Splatviz, version 1.1.0, introduces a range of new features, usability improvements, and code architecture changes aimed at enhancing the user experience and expanding the tool's functionality.

You may recall Splatviz as the Python based 3DGS Viewer and Editor.

Key New Features and Enhancements

One of the most notable additions in Splatviz v1.1.0 is the ability to launch the tool in three different ways.

1. Default Editor Mode: Users can now load the default editor, which allows for rendering and editing of PLY files directly. This mode provides a straightforward entry point for those working with 3D data files.

2. 3DGS Training Attachment: For users involved in 3DGS training, Splatviz can now be attached to a running training session. This functionality integrates seamlessly with current 3DGS implementations, requiring no modifications to existing setups. However, for live scene editing during training, users will need to follow a specific tutorial available on the Splatviz GitHub page. It should also be possible to bring this into gsplat as well, but you will need to insert the following code into the training loop:

   network.render(pipe, gaussians, ema_loss_for_log, render, background, iteration, opt)

   network.render(pipe, gaussians, ema_loss_for_log, render, background, iteration, opt)

   network.render(pipe, gaussians, ema_loss_for_log, render, background, iteration, opt)

   network.render(pipe, gaussians, ema_loss_for_log, render, background, iteration, opt)

3. 3DGAN Face Rendering: This update also brings the ability to render 3DGAN faces, leveraging the models described in a research paper, also authored by Barthel. This feature opens new avenues for researchers and developers working with GANs and 3D face generation.

Improved Performance and Training Monitoring

The performance monitoring capabilities of Splatviz have also been enhanced in this release. A reworked performance widget now provides detailed GPU information, including current memory usage, temperature, and CUDA version. Being able to see memory usage plainly has always been a huge benefit for me personally.

Additionally, a new training widget has been introduced, offering real-time insights into key statistics of the ongoing training process. This includes metrics such as the number of Gaussians, loss values, spherical harmonics degree (sh_degree), and hyperparameters. These insights are crucial for users looking to fine-tune their training parameters and achieve optimal results.

Usability Improvements

Splatviz v1.1.0 also focuses on enhancing usability, with several updates aimed at streamlining the user experience:

• Training Controls: A new feature allows users to pause and resume training with a single button, offering greater control over long-running processes.

• Editing Presets: The process of saving and loading editing presets has been improved. Now, all associated sliders are saved with a preset, eliminating the need to recreate them each time a preset is loaded.

• Slider Customization: Users can now edit slider names and ranges without the need to delete and recreate them, providing more flexibility in customizing the editing interface.

• Session History: Splatviz now supports saving and restoring the last editing session. Users can close and reopen sessions while retaining access to a history of their edits, improving workflow continuity.

• Camera Controls: The update introduces more intuitive camera controls, including the ability to translate the camera position using the middle mouse button and adjust camera movement speed. These changes make navigation within 3D scenes more fluid and customizable.

• Code Highlighting: The editing code cell now features syntax highlighting for known variables such as Gaussian objects and sliders, making the coding process more user-friendly and less error-prone.

Code Architecture and Future Developments

On the backend, the code architecture has been refined to facilitate the addition of new renderers and widgets. This structural improvement not only simplifies future development but also makes the tool more extensible for advanced users looking to customize their Splatviz experience.

While the update includes support for attaching Splatviz to a training session on a server, this feature is still in its early stages and has not yet been fully tested. However, it represents a promising direction for future updates, particularly for users who manage large-scale training sessions across multiple systems.

With new features for 3D visualization, improved performance monitoring, and a host of usability tweaks, this release is poised to benefit. As Splatviz continues to evolve, these updates lay a solid foundation for future innovations in 3D editing and neural network training visualization. It continues to be MIT Licensed and can be accessed here.