ExplorerAI-v0.-1.3b is a research checkpoint developed by ExplorerAI through further fine-tuning of ArtiFixer 1.3B, for research on reconstruction-conditioned image refinement and novel-view synthesis.
This release documents the model's initialization and provides measurements of how its weights changed during further training. Our checkpoint analysis identifies concentrated updates in camera/opacity conditioning and neighbor-view projections, alongside more distributed updates in the transformer backbone. These findings describe parameter changes; they are not measurements of image quality or geometric accuracy.
The model weights are available for non-commercial research and evaluation under the NVIDIA License. Availability of the weights does not grant unrestricted commercial use.
What changed during further training?
We compared the candidate checkpoint against the specified initialization using a full CPU FP64 scan.
| Weight-space measurement | Result | Interpretation / parameter role |
|---|---|---|
Global relative Frobenius update, ‖W_new − W_base‖F / ‖W_base‖F |
1.6094% | Measures the overall update magnitude relative to the original weight norm. |
| Global cosine similarity | 0.9998705 | Measures alignment between the flattened original and fine-tuned weight vectors, rather than similarity of model capabilities. |
| Neighbor-view V projection matrices: relative update | 47.5230% | Transform neighboring-view content into value features for cross-attention, influencing how texture and appearance information contributes to target-view generation. |
| Camera-conditioning matrices: relative update | 21.6898% | Map camera-ray geometry into internal features, influencing how the model uses viewpoint and viewing-direction information to condition generation. |
| Opacity-conditioning matrices: relative update | 14.3097% | Encode opacity from reconstruction renderings, providing cues about regions with sufficient or limited geometric coverage that can guide refinement and completion. |
The module measurements are aggregated over all 30 transformer blocks and refer to the projection weight matrices, excluding their biases. Relative updates can be large when the original matrix norm is small. These percentages do not represent an originality ratio or a performance improvement.
References
@inproceedings{delutio2026artifixer,
title={ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models},
author={de Lutio, Riccardo and Fischer, Tobias and Chang, Yen-Yu and Zhang, Yuxuan and
Wu, Jay Zhangjie and Ren, Xuanchi and Shen, Tianchang and Tothova, Katarina and
Gojcic, Zan and Turki, Haithem},
booktitle={SIGGRAPH},
year={2026}
}