FixAnything: Unified 3D Rendering Artifact Repair via Video Generative Priors
gabriel1 · x · 2026-08-26
FixAnything is a generalist video model designed to repair rendering artifacts from any 3D representation—whether 3DGS, NeRF, meshes, or sparse point clouds.
Core Principles & Method:
- Video-to-Video Translation: The fix is posed as a video-to-video translation task, achieved by repurposing a pretrained video diffusion model (Wan2.1-I2V-14B) with minimal modification and finetuning.
- Sparse Point Cloud Control: Even a very sparse point cloud (e.g., from COLMAP) provides effective camera control for the video model, requiring only minimal LoRA finetuning on 20 paired videos.
- Geometric Consistency Reward: Camera poses recovered by COLMAP from each output serve as the reward; Flow-DPO then steers the model toward geometrically consistent outputs.
- Anchoring Mechanism: The model treats clean training views along the trajectory as anchors, propagating appearance, lighting, and scene structure into the degraded frames in between.
The paper has been accepted to ECCV 2026.
Related event: FixAnything Unifies 3D Rendering Artifact Repair with Video Priors(2 posts)→
More from Multimodal
- Officially calling it: the post-training era for video models — gorkem · 2026-08-26
- AI relighting turns ordinary portraits into commercial-grade shots without changing the face — sven_ai · 2026-08-26
- Hands-on With NoSpoon H3 Music Video Agent: Three Takes Cut Into a Multicam — Kyrannio · 2026-08-26
- sanoTTS: 1.4M-Param Model Runs Real-Time on $3 Chip — kastnerkyle · 2026-08-26
- New Models to Know: MoE-ViE, τ0-VLA, 4DAnyone, and More — TheTuringPost · 2026-08-26
- HYPERREALITY Releases Second Trailer — Listen_Expert · 2026-08-26