API inconsistency: ProjectGaussians uses log_scales while ProjectGaussiansJagged uses scales

[harrism]

Summary

There is an API inconsistency between the two Gaussian projection paths:

• ProjectGaussians (non-jagged): Takes logScales as input and returns dL/d(log_scale) gradients in the backward pass
• ProjectGaussiansJagged (jagged, via gaussian_render_jagged): Takes raw scales as input and returns dL/d(scale) gradients in the backward pass

Background

This divergence was discovered while fixing PR #433 (chain rule for log_scale gradient). The GaussianSplat3d class internally stores mLogScales, and the scales() accessor returns exp(mLogScales). However:

• The non-jagged ProjectGaussians::forward() directly uses mLogScales
• The jagged gaussian_render_jagged() API accepts raw scales (what the user passes), which then flows through to ProjectGaussiansJagged

Current Workaround

To handle this divergence, PR #433 added a template parameter ApplyLogScaleChainRule to quaternionAndScaleToCovarianceVectorJacobianProduct:

template <typename T, bool ApplyLogScaleChainRule = true>
quaternionAndScaleToCovarianceVectorJacobianProduct(...) {
    // ... compute dL/d(scale) ...
    if constexpr (ApplyLogScaleChainRule) {
        // Return dL/d(log_scale) = dL/d(scale) * scale
        return {dLossDQuat, scale * dLossDScale};
    } else {
        // Return raw dL/d(scale)
        return {dLossDQuat, dLossDScale};
    }
}

The regular backward uses ApplyLogScaleChainRule=true, while the jagged backward uses ApplyLogScaleChainRule=false.

Recommendation

Consider unifying the APIs in a future release by changing gaussian_render_jagged to accept log_scales instead of scales. This would:

1. Make the API consistent with the internal representation (mLogScales)
2. Simplify the gradient computation code by removing the template parameter workaround
3. Match the pattern used by the non-jagged projection path

Note: This would be a breaking API change for users of gaussian_render_jagged.

Affected Files

• src/fvdb/detail/ops/gsplat/GaussianUtils.cuh - Contains the workaround
• src/fvdb/detail/ops/gsplat/GaussianProjectionBackward.cu - Uses ApplyLogScaleChainRule=true
• src/fvdb/detail/ops/gsplat/GaussianProjectionJaggedBackward.cu - Uses ApplyLogScaleChainRule=false
• src/fvdb/detail/autograd/GaussianProjection.cpp - Autograd definitions
• src/fvdb/GaussianSplat3d.cpp - gaussianRenderJagged function
• fvdb/__init__.py - Python wrapper for gaussian_render_jagged

[fwilliams]

Let's remove this API for now instead. I think the "Jagged" version of Gaussian splats should be encapsulated in a GaussianSplatBatch class akin to Grid and GridBatch.