The system adapts Wan2.1-I2V-14B with LoRA and conditions on both degraded video latents and a per-frame trust mask. Clean frames serve as appearance anchors. A second preference-optimization stage uses recovered camera-pose accuracy as a reward, encouraging geometrically consistent rather than merely plausible-looking output.
FixAnything is useful for improving novel-view rendering when source observations are sparse or camera movement exposes reconstruction errors. Its code and weights are released under Apache 2.0. The output remains generated video; downstream users should distinguish visual refinement from a guaranteed correction of the underlying 3D geometry.
