SIGMA-SET27K: Synthetic Multi-Subject Dataset

• SIGMA-SET27K is a synthetic multi-subject image dataset with 26,435 images and over 100k unique identities, designed for identity-preserving generation.
• It offers comprehensive annotations including per-subject identity images, masks, depth maps, and 2D/3D bounding boxes, supporting various control granularities.
• The dataset underpins SIGMA-GEN training through a curriculum that scales subject count, enhancing identity preservation and structural alignment over iterative baselines.

Searching arXiv for papers mentioning SIGMA-SET27K and related SIGMA terminology. SIGMA-SET27K, also written as SIGMA-Set27K, is a synthetic dataset introduced by the paper "SIGMA-GEN: Structure and Identity Guided Multi-subject Assembly for Image Generation" for multi-subject identity-preserving image generation under explicit structural and spatial control (Saha et al., 7 Oct 2025). Its stated purpose is to provide a supervision regime that prior personalization datasets do not support well: a single image may contain multiple distinct subjects whose identities must all be preserved while obeying user guidance ranging from coarse 2D or 3D boxes to pixel-level segmentations and depth. The dataset is described as providing identity, structure, and spatial information for over 100k unique subjects across 27k images, with the appendix giving the exact finalized count as 26,435 images and 105,756 unique identities (Saha et al., 7 Oct 2025).

1. Dataset definition and intended problem setting

SIGMA-SET27K was created to address a missing data regime in controllable generative modeling. The motivating task is not generic image synthesis, but multi-subject insertion and personalized generation in which several identities must be preserved simultaneously within one scene. The paper states that existing personalization datasets are mostly single-subject, or at best contain very few identities per image, while even recent multi-subject datasets are limited to around two identities per image or narrow domains such as virtual try-on (Saha et al., 7 Oct 2025).

Within that framing, SIGMA-SET27K is a general-purpose synthetic corpus of multi-subject scenes. The abstract describes it as a dataset with "over 100k unique subjects across 27k images," while the appendix specifies 26,435 images and a total of 105,756 unique identities (Saha et al., 7 Oct 2025). The distinction between images and identities is structurally important: one image may contain multiple subjects, and those subjects are counted as separate identities.

The paper repeatedly emphasizes that the dataset supports up to 10 subjects per image. In the appendix, prompts are generated with 3 to 10 subjects, and filtering retains only samples with more than 2 subjects per image, so the retained corpus is explicitly multi-subject by construction (Saha et al., 7 Oct 2025). For quantitative testing, the authors also generate a separate evaluation set using the same synthetic process, containing 710 images and 2,102 unique identities, with 200 single-subject examples and 510 multi-subject examples (Saha et al., 7 Oct 2025).

2. Annotation modalities and conditioning schema

Each target image in SIGMA-SET27K provides per-subject identity image, mask, depth, and 2D/3D bounding box (Saha et al., 7 Oct 2025). The dataset is therefore not merely a collection of RGB scenes and captions; it is an aligned supervisory corpus tailored to the conditioning interface used by SIGMA-GEN. The paper states that "the key to our approach is a large-scale synthetic data generation pipeline that automatically produces aligned RGB images, depth maps, masks, and identity descriptors" (Saha et al., 7 Oct 2025).

The formal representation begins with the subject set

$\mathbf{S} = \{s_i\}_{i=1}^N$

and the corresponding spatial regions

$\mathcal{R} = \{R_i\}_{i=1}^N.$

A mapping

$f:\mathbf{S} \longrightarrow \mathbb{M}$

assigns each subject $s_i$ a unique pixel intensity $m_i$. The routing control image is then defined as

$I^{\mathcal{R}}(x) = f(s_i)\mathbf{1}[x \in R_i].$

This encodes spatial layout as a single-channel image whose pixel values identify which subject occupies each region, with 0 reserved for background (Saha et al., 7 Oct 2025).

Per-subject identity images are denoted

$I^{s_i} \in \mathbb{R}^{H' \times W' \times 3},$

and the subject identity condition image is formed by concatenating them along height:

$$I^{\mathcal{S}} \in \mathbb{R}^{(N \cdot H') \times W' \times 3}.$$

The paper states that the $i$-th $H' \times W'$ block in $\mathcal{R} = \{R_i\}_{i=1}^N.$0 corresponds directly to region $\mathcal{R} = \{R_i\}_{i=1}^N.$1, so the identity control image functions as a compact visual dictionary of subject identities (Saha et al., 7 Oct 2025).

The spatial control image combines routing and structure. Because coarse controls such as bounding boxes can overlap ambiguously, the authors define two routing maps, $\mathcal{R} = \{R_i\}_{i=1}^N.$2 and $\mathcal{R} = \{R_i\}_{i=1}^N.$3, obtained by different compositing orders, together with a structure image $\mathcal{R} = \{R_i\}_{i=1}^N.$4 that is primarily depth. The resulting control image is

$\mathcal{R} = \{R_i\}_{i=1}^N.$5

For exact masks, the ascending and descending routing maps are identical (Saha et al., 7 Oct 2025). This unified representation allows the same sample to instantiate fine control with masks and depth, intermediate control with 3D-box masks and depth, and coarse control with 2D boxes.

3. Synthetic generation pipeline

SIGMA-SET27K is fully synthetic and generated through a fully automatic pipeline (Saha et al., 7 Oct 2025). The pipeline has six stated stages.

First, prompt creation is performed with an LLM. The paper specifies Qwen-3-8B and states that it is used to generate a compositional image prompt together with full-scene, background, and subject captions. The appendix further states that the LLM is prompted to generate image-generation prompts with 3 to 10 subjects (Saha et al., 7 Oct 2025).

Second, a target image is generated for each prompt using an off-the-shelf text-to-image model. The paper does not name that generator in the dataset section itself, but it is explicit that this stage is synthetic image generation rather than real-image collection (Saha et al., 7 Oct 2025).

Third, individual subject masks are obtained with Grounded-Segment-Anything, using the subject captions as grounding cues (Saha et al., 7 Oct 2025). Fourth, depth is estimated with MoGe-2 (Saha et al., 7 Oct 2025). Fifth, identity images are created by reposing subject crops with Flux.1-Kontext-dev, producing identity references under different poses and lighting; the paper highlights this as a key step because it makes the reference image function as an identity cue rather than merely a crop of the final pose (Saha et al., 7 Oct 2025). Sixth, 2D boxes are fit to segmented subjects and oriented 3D bounding boxes are estimated with Open3D on the depth for each subject (Saha et al., 7 Oct 2025).

The annotations are therefore algorithmic rather than manual. Masks come from grounded segmentation, depth from a depth estimator, 3D boxes from geometric fitting on estimated depth, identity images from generative reposing, and prompts and captions from an LLM (Saha et al., 7 Oct 2025).

The paper also specifies filtering heuristics. During segmentation, boxes smaller than 1% of total image area or larger than 40% are removed; duplicate or overlapping masks are pruned so that only one is kept; and only samples with more than 2 subjects per image are retained for the main SIGMA-SET27K corpus (Saha et al., 7 Oct 2025). These filters serve both annotation quality control and enforcement of the dataset's multi-subject character.

4. Role within SIGMA-GEN training

SIGMA-SET27K is presented not as an auxiliary resource but as the foundational supervision layer of SIGMA-GEN (Saha et al., 7 Oct 2025). The method section states that to enable controllable generation with many subjects in one shot, the authors create a high-quality dataset containing multiple subject identities in each image, and the model is then fine-tuned using that dataset and the proposed conditioning representation (Saha et al., 7 Oct 2025).

The training schedule is explicitly aligned with subject-count complexity. Stage 1 trains on a subset with up to four subjects per image; stage 2 trains on images with three or more subjects; and stage 3 trains on images with more than four subjects (Saha et al., 7 Oct 2025). This curriculum reflects the structure of the dataset itself rather than an external benchmark convention.

The dataset also supports unified training across multiple control granularities. During training, one of three structural inputs is randomly sampled for each example: precise masks with depth, 3D bounding-box masks with depth, or 2D bounding boxes (Saha et al., 7 Oct 2025). One spatial condition channel is randomly dropped with probability 0.1, and masks or boxes are augmented through dilation and aspect-ratio perturbation (Saha et al., 7 Oct 2025). A plausible implication is that the dataset is sufficiently annotation-rich to let a single model learn a common latent interface across fine and coarse control modalities.

The paper additionally notes that AnyInsert and MUSAR-Gen are processed to obtain spatial conditions and captions for single- and double-subject examples, but only for the first stage of training; later stages use SIGMA-SET27K (Saha et al., 7 Oct 2025). This places SIGMA-SET27K at the center of the model's many-subject regime rather than at the periphery.

5. Evaluation protocol and empirical significance

SIGMA-SET27K is primarily presented as a training dataset, while quantitative evaluation is performed on a separately generated 710-example test set produced through the same synthetic process (Saha et al., 7 Oct 2025). The reported metrics are DINO-I and SigLIP-I for identity preservation, SigLIP-T for text-image alignment, Depth MSE when precise depth is given, and CLIP-IQA and MUSIQ for perceptual quality (Saha et al., 7 Oct 2025). Baselines include UniCombine, OminiControl, Flux Kontext + depth ControlNet, Insert Anything*, and MSDiffusion, depending on the control regime (Saha et al., 7 Oct 2025).

For multi-subject generation with mask and depth control, SIGMA-Gen reports DINO-I 74.54 versus Insert Anything* 72.72, SigLIP-I 77.82 versus 75.58, SigLIP-T 17.73 versus 17.66, Depth MSE 26.35 versus 203.4, CLIP-IQA 72.64 versus 44.41, and MUSIQ 73.21 versus 48.86 (Saha et al., 7 Oct 2025). For multi-subject bounding-box control, it reports DINO-I 71.90 versus MSDiffusion 63.28, SigLIP-I 73.15 versus 69.06, SigLIP-T 17.21 versus 11.20, CLIP-IQA 68.83 versus 61.99, and MUSIQ 70.96 versus 69.05 (Saha et al., 7 Oct 2025). For multi-subject 3D-box-mask plus depth control, the paper reports DINO-I 73.48, SigLIP-I 75.27, SigLIP-T 18.19, CLIP-IQA 72.45, and MUSIQ 72.55 (Saha et al., 7 Oct 2025).

The paper presents these results as evidence that the dataset and training setup let the model preserve identity and structure more effectively as subject count increases (Saha et al., 7 Oct 2025). A plot over number of subjects is described as showing that iterative baselines degrade in quality and speed much more sharply, whereas SIGMA-Gen remains relatively stable, with especially strong gains for scenes with five or more subjects (Saha et al., 7 Oct 2025).

Ablations also expose the value of the dataset's structural annotations. On multi-subject evaluation, "Mask (bg)" yields DINO-I 74.17, SigLIP-I 77.52, SigLIP-T 17.62, MSE 40.10, CLIP-IQA 71.26, and MUSIQ 72.27, whereas "Mask + depth (bg)" improves to 74.54 / 77.82 / 17.73 / 26.35 / 72.64 / 73.21 (Saha et al., 7 Oct 2025). This suggests that depth supervision is not ornamental but materially participates in the learned control interface.

6. Limitations, release status, and nomenclatural disambiguation

Several limitations are explicit. First, the dataset is synthetic, so there is an implied synthetic-to-real gap, although the paper does not deeply analyze domain transfer (Saha et al., 7 Oct 2025). Second, the identity images are generated by reposing subject crops with a generative model, so "identity" is operationally defined through synthetic visual consistency rather than human-verified persistent identity (Saha et al., 7 Oct 2025). Third, the paper reports failure cases when coarse controls have very high overlap, in which one subject may be ignored (Saha et al., 7 Oct 2025). Fourth, large viewpoint changes relative to the identity image can hurt consistency (Saha et al., 7 Oct 2025). Finally, the paper explicitly observes weakness in human facial identity, stating that "the training data is not specifically designed for this task" (Saha et al., 7 Oct 2025).

Release status is more limited than the model presentation. The abstract announces "Code and visualizations" at the project webpage, but the paper does not explicitly state that SIGMA-SET27K itself is publicly downloadable, and it does not provide a dataset repository URL, benchmark protocol document, or license (Saha et al., 7 Oct 2025). The most precise statement is therefore that code and visualizations are announced, while public dataset release is not clearly confirmed in the provided text.

The name also requires disambiguation. Although "SIGMA" appears in several unrelated arXiv contexts, the exact identifier SIGMA-SET27K is tied to SIGMA-GEN's synthetic multi-subject image-generation dataset and not to the resonant anomaly-detection method SIGMA (Das et al., 2024), the Short-spacing Interferometer Array for Global 21-cm Signal detection (Zhao et al., 7 Apr 2025), or reduced SIGMA basis sets for molecular calculations (Ema et al., 7 Jul 2025). A plausible source of confusion is that those works use the same surface acronym while referring to unrelated methodological or instrumental programs.

In summary, SIGMA-SET27K is a synthetic, automatically generated, multi-subject supervision corpus built to support single-pass identity-preserving generation under mixed structural and spatial controls. Its defining properties are 26,435 images, 105,756 unique identities, 3 to 10 subjects per prompt, and aligned per-subject identity images, masks, depth, and 2D/3D boxes (Saha et al., 7 Oct 2025). Within the SIGMA-GEN framework, it functions as the data substrate that makes unified training across masks, depth, 3D boxes, and 2D boxes technically feasible, while its limitations remain those typical of synthetic control-oriented corpora: synthetic-to-real uncertainty, operational rather than verified identity, and performance degradation under severe overlap or large viewpoint shifts (Saha et al., 7 Oct 2025).