Video-diffusion LoRA that repairs artifacts in 3DGS, NeRF, mesh and point-cloud renders while keeping the camera path.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
FixAnything is a research method from Carnegie Mellon University (Khiem Vuong, Deva Ramanan, Srinivasa Narasimhan), accepted at ECCV 2026. It treats cleanup of degraded 3D reconstructions as video-to-video translation. A degraded render along a camera trajectory is fed to a pretrained video diffusion model (Wan2.1-I2V-14B), which fills holes and removes artifacts while keeping the camera path and scene layout.
Training has two stages. Stage I is supervised finetuning of a rank-64 LoRA (under 1% of parameters) on paired videos from DL3DV-10K, rendered from four representations and trained jointly. The degraded render's latent is concatenated with the noisy latent, along with a per-frame mask marking which frames to trust. As few as 20 paired videos are reported to suffice, using 3000 iterations on one H100. Stage II applies Flow-DPO, with COLMAP camera-pose accuracy of each output as the reward, raising pose accuracy from 61.1 to 68.3 AUC@5° with no added inference cost.
The project page reports results on DL3DV, DL3DV-Drone, MipNeRF-360 and LLFF, and a preliminary uncertainty estimate from multiple seeds that correlates with reconstruction error. It is a research release aimed at 3D vision practitioners. A YouTube creator states code is released, but no repository link was in the provided material. Inference requires the large Wan 2.1 14B base model.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Cleanup system for degraded 3D reconstructions (Gaussian splats, NeRFs, meshes) that uses a pretrained video model to fix holes and artifacts while keeping the camera path. It is a LoRA on Wan 2.1 and the code is released.
Closest neighbours by meaning, not just by category.
Research
Research method that rigs animatable Gaussian characters with an inner kinematic skeleton and outer free-form bones.
Model
Feed-forward 3D scene reconstruction model that outputs triangle primitives directly usable as meshes for simulation.
Model
Feed-forward 3D Gaussian Splatting model that reconstructs scenes from arbitrary input views, with or without camera poses.
Model · NVIDIA
Open few-step diffusion refiner from NVIDIA that regenerates low-resolution H3 video drafts at 768P to 4K, 5–15 seconds.
Research
Multi-view video diffusion model that turns monocular or sparse videos of dynamic humans into dense multi-view videos for 4D Gaussian splats.
Model
Research method that reconstructs a 4D Gaussian Splatting human from a single casual monocular video.