AnimalBench: Animal Image Generation Benchmark

• AnimalBench is a benchmark and dataset for personalized animal image generation, emphasizing high-resolution imagery and animal-specific identity preservation.
• It includes over 10,000 training and 1,000 test high-definition animal image–text pairs, with rich annotations to evaluate perceptual, identity, and text-image alignment metrics.
• The benchmark tests image-conditioned personalization methods, revealing trade-offs in preserving texture fidelity, structural details, and text adherence.

AnimalBench is a high-resolution dataset and evaluation benchmark for personalized animal image generation, introduced together with the AnimalBooth framework to address the absence of a dedicated benchmark for animal personalization (Liu et al., 20 Sep 2025). In that formulation, AnimalBench is designed around image-conditioned generation with animal-specific identity preservation, where the central difficulty is that animal subjects combine non-rigid structure, large cross-instance variation, and fine-grained identity cues such as fur textures, coat patterns, facial details, horn or antler structure, and breed-specific morphology. The benchmark therefore targets a regime that differs from generic subject-driven generation by emphasizing fine-grained visual fidelity, identity consistency, and text-image consistency under animal-specific appearance variation.

1. Definition and benchmark scope

AnimalBench is presented as both a dataset and a benchmark protocol. As a dataset, it consists of high-definition, single-subject animal image–text pairs. As a benchmark, it is used to evaluate personalized animal image generation methods with automatic metrics that quantify perceptual distance, image-image similarity, and text-image alignment (Liu et al., 20 Sep 2025).

The task supported unambiguously in the paper is personalized text-to-image generation conditioned on an animal subject. The intended evaluation questions are whether a model can preserve the identity of a reference animal, whether the generated result remains aligned with the text prompt, and whether local appearance details survive generation. The paper does not describe segmentation, retrieval, captioning, or editing as benchmark targets, even though the dataset illustration suggests richer annotations.

The benchmark’s motivating premise is that generic personalization methods do not adequately test or preserve animal identity. The paper attributes this to non-rigid body structure, large morphological variability, and identity-critical local features. A plausible implication is that AnimalBench is less about generic visual consistency than about whether a model can preserve animal-specific biometric texture and structure under prompt-conditioned synthesis.

2. Dataset composition and documentation

The paper reports 10,958 training images and 1,000 test images for AnimalBench, and describes each instance as a “high-definition, single-subject animal image–text pair” (Liu et al., 20 Sep 2025). Training is conducted at 512×512 resolution, but the text does not clarify whether this is also the native stored resolution of the dataset or only the effective training resolution.

Figure 1 introduces an important documentation ambiguity. The figure caption refers to an “Animal-HD dataset”, while the paper elsewhere formally introduces AnimalBench. The figure depicts five associated components for each entry: a high-quality source image, a detailed caption describing the scene and subject, a precise pixel-level semantic segmentation mask of the primary subject, the masked subject for visual verification, and the corresponding paired image. The paper does not resolve the naming discrepancy between “AnimalBench” and “Animal-HD,” but it does suggest that the benchmark is built on richer curation assets than a plain image–text corpus.

Several important dataset properties are explicitly not reported. The paper does not specify the number of unique animal identities, species, or breeds; it does not define how many prompts are associated with each subject; and it does not state whether the train/test partition is subject-disjoint or image-disjoint. It also does not explain where the images were sourced from, how captions were written, how segmentation masks were produced, how the “paired image” is defined, or what filtering criteria were applied during curation. The qualitative examples show multiple animal types, including cheetah, reindeer, Highland cow, and zebra, so multiple species are clearly present, but exact coverage is left unspecified.

Only a train/test split is reported. No validation split is described. Release status is also provisional: the paper says AnimalBench “will be released upon acceptance”, but does not provide a code URL, a data URL, or licensing terms.

3. Evaluation protocol and reported metrics

AnimalBench is used as the central experimental testbed for training and evaluating personalized animal image generation systems (Liu et al., 20 Sep 2025). The reported workflow is straightforward: a model is trained on AnimalBench image–text pairs, conditioned on text and reference-animal information, then evaluated on held-out images using automatic metrics.

Four metrics are reported. LPIPS is treated as a perceptual distance with lower being better. DINO is used to quantify image detail and realism and, in the paper’s interpretation, identity/detail preservation. CLIP-T measures text-image consistency, and CLIP-I measures image-image similarity or identity consistency between the generated image and the reference animal. The paper does not provide exact implementation formulas for these metrics on AnimalBench, so their use is operational rather than formally specified in the manuscript.

The benchmark compares BLIP-Diffusion, OmniGen, IP-Adapter, and AnimalBooth. No human evaluation is reported. The paper also notes a metric-interpretation caveat: it states that AnimalBooth achieves 49.08 on LPIPS while treating lower as better, which suggests a scaled or noncanonical LPIPS reporting convention. The safest reading is therefore relative comparison rather than absolute interpretation.

The experimental setup that accompanies the AnimalBench benchmark uses Stable Diffusion v1.5 as the backbone, 8 × NVIDIA A100-SXM4-80GB, AdamW, learning rate $1\times10^{-5}$, batch size 4, training resolution 512×512, DDIM sampling, and 50 inference steps. The paper also reports a two-stage training procedure, but it does not provide the total training schedule length or epoch count.

4. AnimalBooth as the benchmarked reference method

AnimalBench is introduced jointly with AnimalBooth, and the benchmark results are inseparable from that reference system (Liu et al., 20 Sep 2025). AnimalBooth combines three named components: Animal-Net, an Adaptive Attention Module, and a Frequency-Controlled Feature Integration Module.

The model uses a trainable Animal-Net together with a frozen Photography-Net derived from SD v1.5. Animal-Net encodes the reference animal image, while the Photography-Net handles generation. The paper states that the Animal-Net is solely used for encoding reference images, and that during diffusion no noise is added to the reference image, which undergoes only a single forward pass. This suggests an image-conditioned personalization regime in which the reference branch is not jointly denoised with the generated sample.

The adaptive attention mechanism injects animal identity features into the generation backbone. The paper gives the attention output as

$$\mathbf{O}_{h} = \operatorname{Softmax}\!\left(\frac{\mathbf{Q}\mathbf{K}^{\top}}{\sqrt{d}}\right)\mathbf{V} + \lambda\, \operatorname{Softmax}\!\left(\frac{\mathbf{Q}(\mathbf{K}_{ID})^{\top}}{\sqrt{d}}\right)\mathbf{V}_{ID},$$

where $\lambda \in [0,1]$ controls animal identity strength, $\mathbf{Q},\mathbf{K},\mathbf{V}$ come from Photography-Net features, and $\mathbf{K}_{ID}, \mathbf{V}_{ID}$ are derived from Animal-Net features. The paper sets $\lambda = 0.4$ in experiments. It also contains a wording inconsistency: the equation suggests that larger $\lambda$ should emphasize identity cross-attention more strongly, while the text says larger $\lambda$ biases the output toward the diffusion prior.

The frequency-controlled module applies DCT filtering in latent space and uses four masks: $Mask_{mini}$, $Mask_{low}$, $$\mathbf{O}_{h} = \operatorname{Softmax}\!\left(\frac{\mathbf{Q}\mathbf{K}^{\top}}{\sqrt{d}}\right)\mathbf{V} + \lambda\, \operatorname{Softmax}\!\left(\frac{\mathbf{Q}(\mathbf{K}_{ID})^{\top}}{\sqrt{d}}\right)\mathbf{V}_{ID},$$0, and $$\mathbf{O}_{h} = \operatorname{Softmax}\!\left(\frac{\mathbf{Q}\mathbf{K}^{\top}}{\sqrt{d}}\right)\mathbf{V} + \lambda\, \operatorname{Softmax}\!\left(\frac{\mathbf{Q}(\mathbf{K}_{ID})^{\top}}{\sqrt{d}}\right)\mathbf{V}_{ID},$$1. The paper’s stated interpretation is that mini-pass corresponds to texture, low-pass to texture and structure, mid-pass to layout, and high-pass to contour consistency. Training is divided into two denoising stages,

$$\mathbf{O}_{h} = \operatorname{Softmax}\!\left(\frac{\mathbf{Q}\mathbf{K}^{\top}}{\sqrt{d}}\right)\mathbf{V} + \lambda\, \operatorname{Softmax}\!\left(\frac{\mathbf{Q}(\mathbf{K}_{ID})^{\top}}{\sqrt{d}}\right)\mathbf{V}_{ID},$$2

and

$$\mathbf{O}_{h} = \operatorname{Softmax}\!\left(\frac{\mathbf{Q}\mathbf{K}^{\top}}{\sqrt{d}}\right)\mathbf{V} + \lambda\, \operatorname{Softmax}\!\left(\frac{\mathbf{Q}(\mathbf{K}_{ID})^{\top}}{\sqrt{d}}\right)\mathbf{V}_{ID},$$3

The paper does not provide a separate explicit identity-preserving loss, feature-alignment loss, or contrastive loss. Identity preservation is attributed to conditioning architecture rather than to a standalone objective.

5. Reported benchmark performance

On AnimalBench, AnimalBooth is reported to outperform BLIP-Diffusion, OmniGen, and IP-Adapter on all four automatic metrics (Liu et al., 20 Sep 2025). The reported scores are:

• BLIP-Diffusion: LPIPS 68.21, DINO 62.96, CLIP-T 19.68, CLIP-I 82.38
• OmniGen: LPIPS 71.68, DINO 50.05, CLIP-T 19.41, CLIP-I 72.95
• IP-Adapter: LPIPS 62.91, DINO 72.88, CLIP-T 19.39, CLIP-I 89.75
• AnimalBooth: LPIPS 49.08, DINO 75.66, CLIP-T 20.73, CLIP-I 90.00

Relative to the strongest baseline on each metric, AnimalBooth improves LPIPS by 13.83 over IP-Adapter, DINO by 2.78 over IP-Adapter, CLIP-T by 1.05 over BLIP-Diffusion, and CLIP-I by 0.25 over IP-Adapter. The paper’s qualitative interpretation is that the gains reflect better identity fidelity, better preservation of coarse structure and fine-grained texture, and slightly better text alignment.

The ablation study on frequency conditioning is one of the clearer mechanistic findings. Without frequency conditioning, the model reports LPIPS 69.40, DINO 58.52, CLIP-T 21.37, and CLIP-I 78.17. High-Pass yields 56.72 / 64.72 / 21.21 / 80.02; Mid-Pass yields 61.66 / 66.97 / 20.84 / 82.79; Mini-Pass yields 68.21 / 59.97 / 21.20 / 80.89; and Low-Pass yields 49.08 / 75.66 / 20.73 / 90.00. The paper therefore selects Low-Pass because it is best on LPIPS, DINO, and CLIP-I, even though CLIP-T is slightly higher without frequency conditioning or with High-Pass. This indicates a reported tradeoff between text alignment and identity/detail preservation.

The qualitative discussion emphasizes four example animals: cheetah, reindeer, Highland cow, and zebra. The paper attributes failures in baseline methods to distortions in facial features, fur patterns, antler details, winter coat texture, color accuracy, stripe sharpness, and overall texture refinement, while claiming that AnimalBooth better preserves both coarse structure and fine-grained texture.

6. Relation to the wider animal benchmark landscape and current limitations

AnimalBench occupies a narrow but distinct position within the broader landscape of animal-oriented benchmarks. Other resources in the same ecosystem target different problems: OpenAnimals systematizes animal re-identification across datasets such as HyenaID, LeopardID, SeaTurtleID, and WhaleSharkID (Hou et al., 2024); MammalNet benchmarks mammal video recognition, behavior recognition, compositional low-shot generalization, and behavior detection (Chen et al., 2023); AnimalTrack targets multi-animal tracking in the wild (Zhang et al., 2022); AP-10K and APTv2 focus on animal pose estimation and pose tracking (Yu et al., 2021, Yang et al., 2023); PetFace addresses animal face identification with seen-individual re-identification and unseen-individual verification (Shinoda et al., 2024); and BEBE benchmarks behavior classification from animal-borne sensor data (Hoffman et al., 2023). This comparison suggests that AnimalBench is not a general-purpose animal perception benchmark, but a benchmark for one specific generative problem: subject-driven animal image personalization.

That narrow scope is accompanied by substantial documentation gaps. The paper does not provide subject-wise breakdowns by species or breed, does not report the number of unique animal identities, does not describe prompt templates or test-set prompt construction, and does not analyze robustness across viewpoint, occlusion, or other controlled difficulty factors (Liu et al., 20 Sep 2025). It also omits source provenance, annotation workflow, segmentation-mask production details, licensing terms, code URLs, data URLs, training schedule length, and metric implementation details. The paper explicitly acknowledges additional presentation inconsistencies, particularly the “AnimalBench” vs “Animal-HD” naming mismatch and the inconsistent interpretation of $$\mathbf{O}_{h} = \operatorname{Softmax}\!\left(\frac{\mathbf{Q}\mathbf{K}^{\top}}{\sqrt{d}}\right)\mathbf{V} + \lambda\, \operatorname{Softmax}\!\left(\frac{\mathbf{Q}(\mathbf{K}_{ID})^{\top}}{\sqrt{d}}\right)\mathbf{V}_{ID},$$4.

These omissions do not negate the benchmark contribution, but they constrain its present scientific utility. The most accurate characterization is therefore that AnimalBench is an important but sparsely documented dataset-plus-benchmark contribution for personalized animal image generation. It establishes a dedicated evaluation setting for animal-specific identity preservation in diffusion-based personalization, but its long-term benchmark value depends on the completeness and reproducibility of the promised release.