ArtiFixer is a few-step causal auto-regressive model that enhances and extends 3D reconstruction. The related source code provides implementations for training, evaluation, and inference, supporting various stages including bidirectional training, diffusion forcing, and Self-Forcing-style DMD distillation.
ArtiFixer was developed by NVIDIA (Spatial Intelligence Lab) and based on Wan2.1's 14B model.
This model is for research and development only.

GOVERNING DOWNLOAD TERMS: Use of the model is governed by the NVIDIA License.
ADDITIONAL INFORMATION: The Wan 2.1 14B base model is governed by the Apache License, Version 2.0.

Global

Developers and researchers working on 3D reconstruction, diffusion models, and auto-regressive techniques for enhancing and extending 3D reconstruction capabilities.

Other: Hugging Face: 06/04/2026 via https://research.nvidia.com/labs/sil/projects/artifixer/

Architecture Type: Transformer
Network Architecture: ArtifixerTransformer (built on Wan2.1's WanTransformer3DModel)
This model was developed based on Wan-AI/Wan2.1-T2V-14B-Diffusers.
Number of model parameters: ~16.9B trainable (16,910,955,584)

Input Type(s): Image, text
Input Format(s): RGB (Red, Green, Blue), opacity maps, camera ray maps, and text
Input Parameters: Rendered RGB and opacity maps from the underlying 3D representation, camera ray maps, and text prompt
Other Properties Related to Input: Model refines and extends renderings from imperfect 3D reconstruction, requiring camera intrinsics and camera poses.

Output Type(s): Image
Output Format: RGB (Red, Green, Blue)
Output Parameters: Two-Dimensional (2D) image frames
Other Properties Related to Output: Generates enhanced images via a few-step causal auto-regressive diffusion model.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g., GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Runtime Engine(s): PyTorch, Hugging Face Diffusers, Hugging Face Transformers, FlashAttention (FA3 on Hopper, FA4 on Blackwell)
Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere, NVIDIA Hopper, NVIDIA Blackwell
Supported Operating System(s): Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

ArtiFixer v1.0
ArtiFixer integrates with PyTorch and requires CUDA environments. It uses Dockerfiles for CUDA 12 and 13, supporting both x86_64 and aarch64 architectures. The model can be run using torchrun (or accelerate) with multi-GPU setups and requires dependencies such as flash-attn, accelerate, diffusers, and transformers.

Data Modality: Text, Image
Image Training Data Size: Less than a Million Images
Text Training Data Size: Less than a Billion Tokens
Data Collection Method by dataset: Hybrid: Automated/Synthetic
Labeling Method by dataset: Automated
Properties (Quantity, Dataset Descriptions, Sensor(s)): Multimodal dataset combining 3D reconstruction data from DL3DV-10K (the DL3DV-ALL-960P release) with text captions. Includes sparse 3D point clouds, camera parameters, and RGB images. Camera poses are estimated with COLMAP, reconstructions are produced with 3DGUT (MCMC densification, via the 3DGRUT library), text captions are generated with a vision-language model (Qwen3-VL-30B-A3B-Instruct), and metric scale is estimated with MoGe. The dataset supports training for 3D reconstruction tasks with both real and generated data.

Data Collection Method by dataset: Automated
Labeling Method by dataset: Automated
Properties (Quantity, Dataset Descriptions, Sensor(s)): We evaluate our model on 4 benchmarks (DL3DV, Nerfbusters, M360, TandT). We follow standard procedures for sparse reconstruction evaluation: only use a subset of frames, i.e., 3, 6 or 9, to evaluate on the remaining held-out frames. We follow the protocol proposed by Cat3D for M360 evaluation, Difix3D+ for Nerfbusters and DL3DV evaluation, ReconX for TandT, and finally propose our own evaluation set for DL3DV.

Artifact removal on the Nerfbusters and DL3DV benchmarks (Difix3D+ protocol; NB = Nerfbusters):

Additional results — Mip-NeRF 360 sparse-view (3/6/9-view), DL3DV novel-content generation, and Tanks & Temples (supplement) — are reported in the paper.

Data Collection Method by dataset: Hybrid: Automated/Human
Labeling Method by dataset: Hybrid: Automated/Human
Properties (Quantity, Dataset Descriptions, Sensor(s)): Evaluated on diverse 3D reconstruction benchmarks including DL3DV and Nerfbusters, assessing performance on held-out validation frames, full source trajectories, and prepared trajectories. Metrics include reconstruction quality, multi-view consistency, and inference speed.

Acceleration Engine: FlashAttention (FA3/FA4); PyTorch SDPA (cuDNN) fallback on Ampere
Test Hardware: NVIDIA A100 80GB, NVIDIA H100, NVIDIA GB200 (Blackwell)

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.