Research map / Video / Video generation

Generating the Wild: Individual-Consistent Image-to-Video Generation for Wildlife

NeurIPS 2026 · Video › Video generation · approach: Autoregressive long video generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes WildIcon, a frequency-guided identity-conditioning branch on frozen Wan2.2-I2V that injects foreground and high-frequency identity tokens into mid/late cross-attention, plus the WildlifeVid wildlife video dataset, and shows gains in identity consistency and downstream animal ReID.

Problem

Image-to-video models fail to preserve fine-grained, identity-discriminative local cues (fur texture, stripes, spots, contours) of individual wildlife animals, limiting their use for synthesizing training data for animal re-identification under data scarcity.

Approach

Extract foreground-isolated high-frequency residuals from the reference image, encode them with a trainable CNN alongside frozen DINO foreground tokens, and inject the fused identity tokens into mid-to-late DiT cross-attention blocks of a frozen Wan2.2 backbone, with an auxiliary ReID-feature identity loss.

Key results

• WildIcon raises Subject Consistency to 92.94% vs 91.02% for the best baseline (Phantom) and 90.79% for Wan2.2-A14B on the 200-case benchmark.
• WildIcon achieves FVD 139 vs 152 for Wan2.2-A14B.
• Adding WildIcon-filtered synthetic frames improves CLIP-ReID mAP on Tiger from 54.10 to 58.36 (real-only vs +Ours synth.).

Limitations

• Identity drift when reference evidence is occluded or incomplete (e.g., nyala overlap, cat vegetation occlusion, tiger fence wires).
• Generation restricted to mild motion and viewpoint changes for ReID augmentation due to reference-evidence guardrail.
• Gains in ReID are measured on five species with a proprietary stoat dataset.

Experimental setup

Datasets: WildlifeVid

Benchmarks: WildlifeVid generation benchmark ATRW Nyala CowDataset IPanda50

Metrics: I2V Subject I2V Background Subject Consistency Motion Smoothness Text Relevance FVD mAP Rank-1

Baselines: Wan2.2-A14B Wan2.2-5B CogVideoX-5B ConsistI2V ConsisID Phantom

Builds on: Wan2.2 flow matching DINO CLIP-ReID

Closest papers by meaning

• Reconstructing Animals and the Wild (CVPR 2025 · similarity 0.86)
• BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes (NeurIPS 2024 · similarity 0.86)
• MammAlps: A Multi-view Video Behavior Monitoring Dataset of Wild Mammals in the Swiss Alps (CVPR 2025 · similarity 0.85)
• The PanAf-FGBG Dataset: Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition (CVPR 2025 · similarity 0.85)
• Web-Scale Collection of Video Data for 4D Animal Reconstruction (NeurIPS 2025 · similarity 0.85)
• WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife (NeurIPS 2026 · similarity 0.85)