NABLA: Bird Identity Preservation Benchmark

• NABLA is a benchmark defined to evaluate identity-preserving bird image generation by testing models on varied poses, viewpoints, and backgrounds.
• It comprises 4,759 expert-curated image pairs and 1,073 true identity pairs from iNaturalist, providing a rigorous dataset for fine-grained evaluation.
• Proxy identity training and control modalities like fill, depth, and keypoints highlight challenges in preserving subtle plumage and structural details.

Searching arXiv for the primary NABLA paper and closely related bird-identification work. arxiv_search(query="(Sun et al., 4 Dec 2025)", max_results=5) arxiv_search: (Sun et al., 4 Dec 2025) NABirds Look-Alikes, abbreviated NABLA, is a benchmark for identity-preserving bird image generation designed to test whether controllable generative models can keep a bird’s apparent identity consistent while changing pose, viewpoint, or background (Sun et al., 4 Dec 2025). It targets a failure mode of current zero-shot identity-preserving systems in non-rigid, fine-grained domains: birds exhibit high visual diversity across species, substantial within-class variation, and many poses and deformations, while accessible same-subject reference data—especially videos or multi-view captures of the same bird—remain scarce (Sun et al., 4 Dec 2025). Within this benchmark, “identity” is not always literal biological individual identity; it is an apparent identity / look-alike identity criterion based chiefly on plumage and structure, curated to support evaluation of whether a model preserves fine-grained bird-specific detail under transformation (Sun et al., 4 Dec 2025).

1. Problem setting and motivation

NABLA was introduced in the context of controllable image generation, where increasingly rich control modes have made zero-shot identity-preserving generation practical for applications such as virtual try-on without requiring additional fine-tuning (Sun et al., 4 Dec 2025). The primary claim motivating the benchmark is that models such as Insert Anything and OminiControl work reasonably well for humans and rigid everyday objects but still struggle in non-rigid, fine-grained domains like birds (Sun et al., 4 Dec 2025).

Birds constitute a demanding test case for three reasons stated explicitly in the benchmark description. First, they have high visual diversity across species. Second, identity depends on subtle details such as plumage, body proportions, and markings. Third, they appear in a wide range of poses and deformations, including perching, flying, swimming, turning, and stretching (Sun et al., 4 Dec 2025). The benchmark therefore shifts the focus of identity-preserving generation away from comparatively rigid objects toward a category where pose variation and fine-grained morphology are tightly coupled.

The benchmark’s broader significance follows from this design choice. The paper argues that such domains are essential for moving beyond content creation toward applications that demand accuracy and fine detail (Sun et al., 4 Dec 2025). In that sense, NABLA functions not merely as a dataset of bird pairs, but as an evaluation regime for whether controllable generation can preserve discriminative, ornithologically relevant structure under controlled transformations.

2. Dataset composition and the meaning of “identity”

NABLA is built from several sources, combining expert-curated look-alike pairs with true same-observation pairs from iNaturalist (Sun et al., 4 Dec 2025).

Source	Composition	Role
NABirds expert-curated pairs	4,759 image pairs	Core NABLA test pairs
iNaturalist observations	1,073 image pairs	True identity pairs
Videos	A small set	Broader evidence base

The core benchmark consists of 4,759 expert-curated image pairs from NABirds. Bird experts manually selected pairs that look like the same individual, and the benchmark spans 539 classes across 401 species (Sun et al., 4 Dec 2025). The auxiliary iNaturalist component contributes 1,073 image pairs, divided into 677 pairs from species seen in NABirds (iNat-Seen) and 396 pairs from species not in NABirds (iNat-Unseen) (Sun et al., 4 Dec 2025). Because images within a single iNaturalist observation usually correspond to the same individual, these pairs are treated as true identity pairs (Sun et al., 4 Dec 2025).

A central definitional point is that NABLA is an apparent identity / look-alike identity benchmark rather than a strict biological individual-identification dataset (Sun et al., 4 Dec 2025). Annotators were instructed to choose a second image in which the bird looks like the same bird as the anchor image, based mainly on plumage—colors, patterns, and textures—and structure—body shape and proportions such as bill length, tail length, and head shape (Sun et al., 4 Dec 2025). The annotation explicitly does not require the same pose or lighting. This constraint is important because the intended task is to preserve identity while allowing variation in pose and scene, matching the actual objective of controllable generation.

The paper also emphasizes a data-quality distinction. NABLA images are high-quality single-subject bird images, whereas many iNaturalist observation images can be blurry, lower quality, or contain multiple birds (Sun et al., 4 Dec 2025). Expert curation is therefore used to filter out misleading within-species pairs that are not actually good identity matches, even when the class label is shared (Sun et al., 4 Dec 2025).

3. Evaluation protocol and control modalities

NABLA evaluates zero-shot identity-preserving generation under a paired-image protocol. For each test pair, one image is used as the subject and the other as the target, generation is run in both directions, and the generated image is compared with the target after masking to reduce background effects (Sun et al., 4 Dec 2025). More specifically, the generated image and the target image are masked using the subject mask to remove background effects, and the output is then scored with four benchmark metrics (Sun et al., 4 Dec 2025).

The evaluation metrics are DINOv2 feature similarity, SigLIP feature similarity, LPIPS, and MSE (Sun et al., 4 Dec 2025). DINOv2 and SigLIP are used because they better capture semantic/class-level identity and pose similarity, whereas LPIPS and MSE measure pixel/perceptual similarity; lower LPIPS and MSE are better, while higher DINOv2 and SigLIP are better (Sun et al., 4 Dec 2025). The benchmark does not introduce a special evaluation loss function; these are the standardized metrics used to assess whether identity is preserved under transformation.

Three control types are evaluated:

1. Fill: inpainting from a mask of the target region, where the subject is inserted into the target’s masked-out shape.
2. Depth: control by the target’s depth map, which provides more pose information than a mask alone.
3. Keypoints: control by 11 bird keypoints from NABirds, converted into a sparse skeleton (Sun et al., 4 Dec 2025).

For the depth and keypoint settings, captions are added because the control image alone may not specify the background well enough (Sun et al., 4 Dec 2025). This design makes the benchmark sensitive not only to identity retention, but also to the adequacy of different geometric control modalities for highly deformable biological subjects.

4. Baselines, failure modes, and proxy-identity training

The primary empirical finding is that state-of-the-art zero-shot baselines fail to reliably preserve identity on NABLA (Sun et al., 4 Dec 2025). The baselines include Insert Anything, OminiControl, and proprietary systems shown in figures such as Nano Banana and GPT-4V (Sun et al., 4 Dec 2025). Reported failure modes include changing plumage details, altering head, wing, and bill patterns, failing to follow the target pose, and producing unrealistic bird shapes when the subject and target are different species (Sun et al., 4 Dec 2025).

The paper’s main improvement strategy is proxy identity training. Because true same-subject bird data is scarce, the models are fine-tuned on pairs of images grouped by species, age, sex, and breeding status, with those labels used as a proxy for identity (Sun et al., 4 Dec 2025). At each training step, two images are sampled from the same class; one becomes the subject, the other the target, the control signal is derived from the target image, and the model is trained to generate the subject in the target control setting (Sun et al., 4 Dec 2025). This is not true individual-identity supervision, but the paper reports that it provides enough structure for the model to learn bird-specific shape priors, fine-grained appearance consistency, and better pose transfer (Sun et al., 4 Dec 2025).

The fine-tuning experiments cover Insert Anything and OminiControl with two backbones, FLUX.1-Schnell (“Omini-S”) and FLUX.1-Kontext (“Omini-K”), trained on 1024×1024 images using LoRA-style fine-tuning; the models are trained separately for each control mode (Sun et al., 4 Dec 2025). Quantitatively, the paper reports that fine-tuning with proxy identity training improves performance across NABLA, iNat-Seen, and iNat-Unseen, including about a 41% reduction in MSE on NABLA over the baseline model (Sun et al., 4 Dec 2025).

Several additional empirical patterns are singled out. Fine-tuned models beat baselines consistently; Omini-K and fine-tuned Insert Anything are especially strong; Fill slightly outperforms depth-based control overall; and Keypoint control performs worst, likely because bird pose is too deformable and ambiguous to be captured well by a sparse skeleton (Sun et al., 4 Dec 2025). The paper also states that performance on NABLA correlates strongly with performance on true identity pairs from iNaturalist, and that improvements on iNat-Seen and iNat-Unseen are similar, indicating generalization beyond seen species (Sun et al., 4 Dec 2025).

5. Relation to fine-grained bird recognition and comparative explanation

NABLA belongs to a broader line of work on bird look-alikes, where visually similar classes must be separated using subtle visual or contextual cues. One directly relevant direction incorporates habitat information into fine-grained bird identification (Nguyen et al., 2023). That work argues that habitat is one of the “four keys” to bird identification used by ornithologists and shows that, on NABirds, habitat-aware augmentation improves ResNet-50 from 80.23 to 81.06 with Mixed-G, a gain of +0.83, while habitat-enriched prompts improve zero-shot CLIP accuracy on a 267-class NABirds subset by up to +0.99 with SSC+habitat (Nguyen et al., 2023). In relation to NABLA, this establishes that the bird look-alike problem is not purely about local appearance: ecological context can reduce confusions among visually similar species when that context is visible and relevant (Nguyen et al., 2023).

A second related direction is comparative explanation rather than direct generation. “Neural Naturalist” introduces the Birds-to-Words dataset of 40,969 sentences describing fine-grained differences between pairs of bird photographs and trains a model with a joint image encoding, a comparative module, and a Transformer decoder to generate comparative paragraphs (Forbes et al., 2019). The paper reports that such descriptions can help humans distinguish image pairs, though the hardest categories remain visual similarity and same species (Forbes et al., 2019). This suggests that NABLA’s identity-preservation objective has a natural explanatory complement: in difficult look-alike regimes, systems may need not only to preserve a bird’s identity under transformation, but also to express which visual differences remain invariant and which are pose-dependent.

Taken together, these related works locate NABLA at the intersection of fine-grained recognition, context-aware disambiguation, and comparative visual reasoning. NABLA differs from both by focusing specifically on controllable generation, yet it inherits the same central difficulty: small perturbations in plumage, structure, or context can determine whether two birds are perceived as the same apparent individual or as different look-alikes.

6. Scope, applications, and terminological notes

The paper presents NABLA as a benchmark intended to support domains where identity fidelity matters more than unconstrained visual plausibility (Sun et al., 4 Dec 2025). The significance claimed for the benchmark is that it enables researchers to measure whether a model truly preserves identity for fine-grained, non-rigid categories, compare systems on a high-quality and human-expert curated dataset, study generalization to unseen species, and develop controllable generation tools for scientific visualization, education, machine teaching, species comparison, and individual re-identification (Sun et al., 4 Dec 2025).

At the same time, the benchmark incorporates several explicit caveats. The main task is constructed from still image pairs even though a small set of videos is mentioned as part of the broader evidence base (Sun et al., 4 Dec 2025). The “identity” target is an apparent look-alike judgment rather than biological individual identity for every pair (Sun et al., 4 Dec 2025). The baseline analysis also implies a boundary on current model reliability: such systems may be useful for creative tasks, but the paper states that they are not sufficiently reliable for scientific or fine-grained applications where identity and detail matter (Sun et al., 4 Dec 2025).

A further terminological note is useful because the word nabla has an established and unrelated meaning in mathematics. In symmetric-function theory, the nabla operator denotes a linear operator acting diagonally on the modified Macdonald basis, and it is the subject of work such as “A combinatorial formula for the nabla operator” (Carlsson et al., 2020). That mathematical usage is unrelated to NABirds Look-Alikes. In the bird-generation literature, NABLA refers specifically to the benchmark introduced for identity-preserving bird image generation (Sun et al., 4 Dec 2025).