FixAnything: 20 paired videos of LoRA let a video diffusion model repair any 3D rendering
kwangmoo_yi · x · 2026-08-26
CMU's FixAnything (ECCV 2026) uses one generalist video model to repair rendering artifacts from any 3D representation:
- Approach: framed as video-to-video translation. 3DGS/NeRF/mesh/point-cloud renders each have artifacts, but a render still preserves camera trajectory and coarse layout — an effective control signal for a pretrained video diffusion model (Wan2.1-I2V-14B)
- Data efficiency: even a very sparse point cloud (e.g. from COLMAP) gives effective camera control, with minimal LoRA finetuning on 20 paired videos
- Alignment: camera poses recovered by COLMAP from outputs serve as the reward; Flow-DPO steers toward geometrically consistent results
- Anchoring: clean training views along the trajectory act as anchors, propagating appearance, lighting and structure into degraded in-between frames
More from Multimodal
- Halcyon early-fusion image models match late-fusion with 275x less pretraining compute, open-sourced — davidwromero · 2026-10-16
- Running Wan2.1 14B FP16 on a 12GB laptop GPU: VRAM swap stress test finishes in 4m11s — RYuu-1120 · 2026-10-16
- Striking Midjourney generations: black holes and planetary aesthetics — ciguleva · 2026-10-16
- Fable 5.5 may drop today as Andrew Curran hints at early release — kimmonismus · 2026-10-16
- Kling 4.0 Adds Timeline Editing, Endorsed by 'Seoul Spring' Director Kim Seong-su — xiaohu · 2026-10-16
- One prompt, no reference images: Argon builds 3D Gaussian splats from geometry equations in NumPy — fofrAI · 2026-10-16