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

Yuzhuo Li; Di Zhao; Xinyu Zhang; Daniel Wilson; Yun Sing Koh

cs.CVOct 4, 2026

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

Authors: Yuzhuo Li, Di Zhao, Xinyu Zhang, Daniel Wilson, Yun Sing Koh

Organizations: School of Computer Science, University of Auckland Auckland, New Zealand

Abstract

Individual-level wildlife identification often suffers from data scarcity, as varying observations of the same animal under diverse poses, viewpoints, and motions are rarely available. Image-to-video (I2V) generation offers a promising way to mitigate this limitation by synthesizing additional observations from a single reference image. However, existing I2V models mainly emphasize global layout, semantics, and motion, and therefore often fail to preserve fine-grained local appearance cues that distinguish one wildlife individual from another, such as fur texture, stripe boundaries, spot configurations, and contour transitions. We observe that these identity-critical cues are closely related to high-frequency information. To address this challenge, we propose WildIcon, a high-frequency-guided I2V framework for wildlife individual consistency. Specifically, WildIcon introduces a frequency-aware identity encoding branch that extracts individual-specific high-frequency cues from the reference image. Combined with isolated foreground information, the resulting identity tokens are then injected into cross-attention blocks as identity conditioning. Building on a frozen backbone with lightweight identity adaptation, WildIcon preserves fine-grained identity cues visible in the reference image while retaining the motion controllability and semantic fidelity of the base I2V model. In addition, to support the training and evaluation of wildlife individual-consistent I2V, we construct WildlifeVid, a wildlife-centric video dataset with high-quality, temporally coherent clips and individual-level identity labels. Experiments on I2V generation and downstream animal re-identification (ReID) show that WildIcon achieves stronger individual consistency than existing baselines, and that its filtered outputs can serve as useful candidate training augmentations for downstream ReID.

Figures & tables

Figure 1 : Frequency-aware identity preservation in wildlife I2V generation. (a) Low-frequency views preserve coarse body structure but suppress fine-grained identity cues, whereas high-frequency views reveal local identity-discriminative details. (b) Given the same reference image and prompt, Wan2.2 blurs or alters identity-critical stripe patterns across frames, while WildIcon better preserves these fine-grained cues for same-individual consistency.

Figure 2 : Overview statistics of WildlifeVid. (a) Composition of WildlifeVid across the four public data sources, showing the number of videos, identities, and species in each source. (b) Distribution of video durations. (c) Species distribution ranked by the number of videos, with cumulative coverage.

Figure 3 : WildIcon augments a frozen Wan I2V backbone with a foreground-filtered high-frequency identity branch. Given a reference image $I_{\mathrm{ref}}$ , the model first obtains a foreground-isolated image $I_{\mathrm{seg}}$ and its high-frequency map $I_{\mathrm{hf}}=\mathcal{B}(I_{\mathrm{seg}})$ . These complementary cues are encoded into identity tokens $T_{\mathrm{id}}$ , projected and fused with text tokens to form a unified condition sequence, and injected into selected middle and late DiT blocks through native cross-attention.

Method	I2V Subject $\uparrow$	I2V Background $\uparrow$	Subject Consistency $\uparrow$	Motion Smoothness $\uparrow$	Text Relevance $\uparrow$	FVD $\downarrow$
CogVideoX-5B [ 67 ]	82.18	86.74	89.92	95.62	25.63	184
Wan2.2-5B [ 55 ]	86.02	88.01	90.16	95.94	26.57	175
Wan2.2-A14B [ 55 ]	86.23	88.27	90.79	96.43	26.58	152
ConsistI2V [ 39 ]	83.71	86.18	88.58	95.08	25.21	188
ConsisID [ 68 ]	84.06	86.29	89.72	95.56	25.98	168
Phantom [ 31 ]	85.64	87.53	91.02	96.11	26.25	172

Table 1 : Quantitative comparison on wildlife individual-consistent image-to-video generation. We compare WildIcon with representative I2V generation models.

Figure 4 : Examples of individual-consistent video generation by our WildIcon. Given a reference image and a motion prompt, WildIcon generates videos that follow the specified motion while maintaining fine-grained appearance cues visible in the reference image across frames.

Figure 5 : Representative challenging cases under incomplete or ambiguous reference evidence. Red boxes highlight local regions where overlap, occlusion, or scene structures in the reference image lead to unsupported or drifting appearance details during generation.

Setting	Subj.	Back.	Cons.	Mot.
FG-only	86.81	88.39	91.33	96.53
HF-only	86.69	88.34	91.48	96.28
FG+HF	87.40	89.36	92.94	96.85

Table 2 : Ablation studies of WildIcon. Subj., Back., Cons., and Mot. denote I2V Subject, I2V Background, Subject Consistency, and Motion Smoothness. FG denotes foreground-only input, HF denotes high-frequency-only input, LoRA-FT denotes LoRA fine-tuning of Wan2.2 on WildlifeVid, and Full denotes the default WildIcon setting. Default configurations are marked in gray .

Model	Setting	Tiger		Nyala		Cow		Panda		Stoat
Model	Setting	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1
SigLIP-FT [ 69 ]	Real only	55.13	90.21	25.84	41.83	62.57	91.53	28.97	45.43	61.16	92.06
	+ Data Aug.	55.41	90.08	26.11	42.06	62.74	91.47	29.22	45.31	61.38	92.14
	+ Wan synth.	57.24	91.36	29.08	43.02	64.08	92.09	31.14	47.82	62.94	92.57
	+ Ours synth.	59.72	93.41	32.01	46.36	65.84	92.11	32.26	49.36	65.18	94.71
DINO-FT [ 48 ]	Real only	54.72	92.61	22.94	35.87	66.31	95.08	29.34	44.56	63.07	93.02

Table 3 : Downstream animal ReID results on five species. All settings use the same real training split, synthetic-data construction pipeline, and unchanged real-only query/gallery splits. We report mAP and Rank-1 (R1) (%) on each species. Best results within each backbone block are shown in bold. Results are averaged over ten runs; 95% confidence intervals are reported in the Appendix.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Animal	Dataset	Train		Query		Gallery		Total
Animal	Dataset	Imgs	IDs	Imgs	IDs	Imgs	IDs	Imgs	IDs
Tiger	ATRW [ Li et al., 2020 ]	3,678	148	424	33	521	33	4,623	181
Nyala	Nyala [ Dlamini and Van Zyl, 2020 ]	1,213	179	354	58	375	58	1,942	237
Cow	CowDataset [ Fu et al., 2022 ]	919	8	210	5	356	5	1,485	13
Panda	IPanda50 [ Wang et al., 2021 ]	4,114	31	1,022	19	1,738	19	6,874	50
Stoat	Stoat (Proprietary)	1,390	37	403	46	780	46	2,573	83

Appendix

Table 4 : Animal ReID datasets used for downstream evaluation. We report the processed ReID splits used in our experiments. These statistics correspond to our curated splits and may differ from the raw dataset statistics reported in the original papers.

Figure 6 : Qualitative comparison with representative video generators. Given the same reference image and text prompt, we compare WildIcon with representative general I2V and identity-preserving video generation baselines.

Method	I2V Subject $\uparrow$	I2V Background $\uparrow$	Subject Consistency $\uparrow$	Motion Smoothness $\uparrow$	Text Relevance $\uparrow$	FVD $\downarrow$
Wan2.2-A14B	86.23	88.27	90.79	96.43	26.58	152
Identity branch input
Full image	86.57	88.41	91.07	96.45	26.56	159
Foreground only	86.81	88.39	91.33	96.53	26.49	148
High-frequency only	86.69	88.34	91.48	96.28	26.43	150
Injection strategy

Appendix

Table 5 : Quantitative ablation studies of WildIcon. We study three factors: the input to the identity branch, the strategy for injecting identity conditioning, and the effect of the WildlifeVid dataset. The final row reports the full WildIcon model as the common reference for all ablations.

Figure 7 : Downstream ReID ablation of WildIcon identity-branch inputs. We visualize the absolute performance gains over the real-only baseline for different identity-branch variants, including foreground-only conditioning (FG-only), high-frequency-only conditioning (HF-only), and the full WildIcon identity branch (Ours). Darker colors indicate larger downstream ReID gains.

Figure 8 : Qualitative ablation studies of WildIcon. We compare WildIcon variants under the same reference image and prompt. Red boxes mark representative artifacts, including tail deformation, back texture drift, head appearance changes, motion inconsistency, and limb distortion. The full WildIcon model better preserves local identity cues and structural consistency across frames.

Figure 9 : Confidence intervals for downstream animal ReID gains. For each backbone and species, we report the change in mAP and Rank-1 relative to the corresponding real-only baseline. Points denote mean gains over ten runs, and horizontal bars denote 95% confidence intervals. The dashed vertical line marks zero gain.

Query source	Tiger	Nyala	Cow	Panda	Stoat
Real reference	64.38 / 96.97	12.87 / 11.11	71.11 / 100.00	30.60 / 68.42	64.49 / 82.61
Wan2.2-A14B	42.73 / 78.79	4.18 / 3.97	54.62 / 85.71	21.87 / 47.37	42.36 / 60.25
LoRA-FT	43.91 / 82.68	4.95 / 5.29	54.74 / 88.57	21.68 / 49.62	46.83 / 67.08
WildIcon	53.72 / 89.61	10.84 / 10.32	64.27 / 94.29	26.14 / 58.65	54.61 / 75.78

Appendix

Table 6: Direct generated-to-real retrieval. Each entry reports frame-level mAP / Rank-1 (%). Generated frames are used as queries against a gallery containing held-out real images of the source individual and same-species negatives. No DINO-based filtering or manual frame selection is applied.

Selection filter	Tiger	Nyala	Cow	Panda	Stoat
DINOv3	+1.95 / +0.93	+2.86 / +2.65	+1.47 / +0.24	+1.09 / +1.88	+1.61 / +0.94
SigLIP	+1.81 / +0.68	+2.94 / +2.11	+1.12 / +0.08	+1.57 / +2.03	+1.44 / +1.06

Appendix

Table 7: Effect of the synthetic-sample selection representation. Each entry reports the absolute mAP / Rank-1 difference (%) between WildIcon and Wan2.2 using CLIP-ReID as the downstream evaluator.

Figure 10 : Frequency-cue utilization analysis. (a) Texture importance and high-frequency (HF) magnitude maps localize identity-related structures, such as stripe boundaries, tail patterns, and contour transitions. (b) ID-token attention and token occlusion sensitivity maps show that WildIcon attends to and depends on these high-frequency identity regions during generation.

Method	GPUs	Train h/1K samples	Infer. s/video	Train mem.	Infer. mem.
Wan2.2	1 GPU	–	693.3	–	51.36
Wan2.2	4 GPUs	–	300.5	–	45.03
WildIcon	1 GPU	5.2	701.5	51.67	55.73
WildIcon	4 GPUs	1.3	315.8	54.48	46.47

Appendix

Table 8 : Efficiency and memory profile. We report 81-frame training and inference costs for Wan2.2-I2V-A14B and WildIcon at

832\times 480

resolution. Training time is reported as hours per 1K training samples. Memory is reported as peak per-GPU memory in GiB.

Dataset	Year	Domain	# Videos	Avg. Len (s)	Duration (h)	Resolution
UCF-101 Soomro et al. [2012]	2012	Human (Action)	13.3K	7.2	26.7	240p
NTU RGB+D Shahroudy et al. [2016]	2014	Human (Action)	114K	–	37	1080p
MSP-Avatar Sadoughi et al. [2015]	2015	Human (Action)	74	–	3	1080p
Taichi-HD Siarohin et al. [2019]	2019	Human (Action)	3K	–	–	256p
SkyTimelapse Xiong et al. [2018]	2018	Sky	35K	–	–	360p
FaceForensics++ Rossler et al. [2019]	2019	Human (Face)	1K	–	–	Diverse

Appendix

Table 9 : Comparison of existing domain-specific video generation datasets and our WildlifeVid.

Figure 11 : Random examples from WildlifeVid.

Panel of Figure 12 — Figure 12 : Instruction template for initial wildlife caption generation.

Explore similar work

May 17, 2026cs.CV

Image-to-Video Diffusion: From Foundations to Open Frontiers

Diffusion-based \textit{image-to-video} (I2V) generation has become a central direction in generative models by turning a reference image, with optional conditions, into a temporally coherent video. Compared with broader video generation settings, this task places stricter demands on content consistency, identity preservation, and motion coherence. Although the literature grows rapidly, existing works mostly discuss I2V generation within broader topics and still lack a dedicated taxonomy together with a systematic analysis centered on this field. This work addresses that gap by treating diffusion I2V generation as a standalone subject. It first reviews the task formulation, model architectures, datasets, and evaluation metrics, and then organizes existing methods through a taxonomy based on architecture and training paradigm. It further distills four core designs, namely condition encoding, temporal modeling, noise prior design, and spatial-temporal upsampling, and discusses representative application scenarios together with major open challenges.

Xianlong Wang, Wenbo Pan, Shijia Zhou +6

Department of Computer Science, City University of Hong Kong, Hong Kong SAR, China · School of Cyber Science and Engineering, Huazhong University of Science and Technology, Wuhan 430074, China · School of Cyber Science and Technology, University of Science and Technology of China, Hefei 230026, China +3

Text-To-Video Diffusion Models Long Video Generation Generative Models Motion Coherence Frontiers

Ask PDF

Jul 22, 2026cs.CV

Vera: Identity-Faithful Human Subject-to-Video Generation

Subject-to-video (S2V) generation has made substantial progress in preserving reference subjects across diverse categories, yet generic subject consistency remains insufficient for human-centric generation. A video may appear globally consistent while identity-critical human details still drift across frames, poses, and interactions. This issue becomes more severe in multi-person scenarios, where incorrect identity-role binding leads to subject confusion, attribute swapping, and excessive copying of reference-specific appearance cues. We propose Vera, a unified human-centric S2V framework for single- and multi-person generation. We first construct a million-pair identity-aligned human image-video dataset through person-level cross-clip retrieval, providing explicit identity correspondence and diverse references. Built on this dataset, Vera introduces two complementary designs. Identity-Focal Masked Supervision (IFMS) strengthens identity-aware learning with spatially focused supervision while reducing interference from irrelevant artifacts. Reference-Aware Layer-wise Attention (RALA) regulates how video tokens interact with reference identity cues in the DiT backbone, preserving stable identity anchors and enhancing layer-aware identity readout. Extensive experiments demonstrate that Vera improves human identity consistency, multi-person subject binding, and motion naturalness, while reducing identity confusion and excessive reference-image copying.

Yulong Xu, Xinyue Liu, Shujuan Li +6

Kuaishou Technology · Institute of Automation, Chinese Academy of Sciences · Tsinghua University

Video Dataset

Ask PDF

May 7, 2026cs.CV

SwiftI2V: Efficient High-Resolution Image-to-Video Generation via Conditional Segment-wise Generation

High-resolution image-to-video (I2V) generation aims to synthesize realistic temporal dynamics while preserving fine-grained appearance details of the input image. At 2K resolution, it becomes extremely challenging, and existing solutions suffer from various weaknesses: 1) end-to-end models are often prohibitively expensive in memory and latency; 2) cascading low-resolution generation with a generic video super-resolution tends to hallucinate details and drift from input-specific local structures, since the super-resolution stage is not explicitly conditioned on the input image. To this end, we propose SwiftI2V, an efficient framework tailored for high-resolution I2V. Following the widely used two-stage design, it addresses the efficiency--fidelity dilemma by first generating a low-resolution motion reference to reduce token costs and ease the modeling burden, then performing a strongly image-conditioned 2K synthesis guided by the motion to recover input-faithful details with controlled overhead. Specifically, to make generation more scalable, SwiftI2V introduces Conditional Segment-wise Generation (CSG) to synthesize videos segment-by-segment with a bounded per-step token budget, and adopts bidirectional contextual interaction within each segment to improve cross-segment coherence and input fidelity. On VBench-I2V at 2K resolution, SwiftI2V achieves performance comparable to end-to-end baselines while reducing total GPU-time by 202x. Particularly, it enables practical 2K I2V generation on a single datacenter GPU (e.g., H800) or consumer GPU (e.g., RTX 4090).

YaoYang Liu, Yuechen Zhang, Wenbo Li +3

HKUST · CUHK · Joy Future Academy +2

Topic pending

Ask PDF