FixAnything uses video AI to clean up broken 3D renders

FixAnything is an open-source tool designed to salvage broken 3D rendering walkthroughs. Whether your scene comes from 3D Gaussian Splatting (3DGS), NeRF, standard meshes, or sparse point clouds, it takes renders plagued by holes, floaters, and tearing artifacts and cleans them into coherent, artifact-free video.

Anyone working with novel view synthesis knows the frustration: the scene looks flawless from the training camera angles, but the moment you pan into unobserved areas, the geometry falls apart. Instead of trying to fix messy point clouds directly, FixAnything treats the entire repair pipeline as a video-to-video translation task.

Using video priors with geometry constraints

Even a heavily degraded render still preserves two vital signals: the camera trajectory and the rough spatial layout of the scene. Those signals turn out to be the perfect steering wheel for a pretrained video model.

FixAnything builds on top of the open-source Wan2.1-I2V-14B video diffusion model. By fine-tuning lightweight LoRA adapters on only about 20 paired video clips, the model learns to interpret coarse point clouds or broken splats as strict motion guidance.

To keep the model from hallucinating structures that drift between frames, the researchers used Flow-DPO reinforcement learning. They run COLMAP on the generated outputs to recover camera poses; outputs with higher geometric consistency receive higher rewards. Additionally, original training frames are locked as anchors, ensuring that real-world details remain untouched while artifacts are hallucinated away.

Running it locally and hardware demands

The entire inference pipeline and LoRA weights are publicly available on GitHub and Hugging Face under the Apache 2.0 license. Pretrained base weights can be pulled from either Hugging Face or ModelScope.

By default, the script takes a 61-frame rendered video along a camera path at 832x480 resolution, and automatically chunks longer camera paths into overlapping windows. If you only have a handful of casual photos, the repository also bundles a MapAnything workflow that creates a coarse reconstruction first and passes the camera pass directly to FixAnything for finishing.

The environment requires PyTorch 2.6 and CUDA 12, with FlashAttention-2 recommended. Keep in mind that Wan2.1 is a 14B model whose checkpoint weighs around 60 GB, so you will need a GPU with substantial VRAM to run inference comfortably.

Why it matters

For creators and technical artists working with Gaussian Splatting and NeRF flythroughs, this is a massive quality-of-life upgrade. Cleaning up stray floaters and avoiding uncaptured blind spots usually requires hours of manual pruning or carefully restrained camera paths. FixAnything bypasses that friction entirely by turning cheap rough drafts into polished footage.

In our view, this approach represents a very smart application of generative video priors to 3D vision problems. Rather than wrestling with imperfect 3D representations, it leans into what video models already do best—generating photorealistic temporal consistency—while using classical geometry to keep them grounded. While it outputs video rather than updating the underlying 3D splats, it is a remarkably practical fix for anyone whose final delivery format is motion video.