SynSpill: Synthetic Industrial Spill Dataset

• SynSpill is a synthetic industrial spill dataset consisting of 2,000 photorealistic images with precise COCO-style annotations, simulating real-world leak scenarios.
• It employs the AnomalInfusion framework, using expert-guided anomaly localization and diffusion-based inpainting to generate context-aware spill scenes.
• The dataset supports both fine-tuning of Vision-Language Models and conventional detectors, enhancing safety-critical detection under low-data, privacy-sensitive conditions.

SynSpill is a curated corpus of 2,000 photorealistic synthetic images of industrial spills introduced by Aaditya Baranwal, Abdul Mueez, Jason Voelker, Guneet Bhatia, and Shruti Vyas in "SynSpill: Improved Industrial Spill Detection With Synthetic Data" (Baranwal et al., 13 Aug 2025). It is designed as a drop-in adaptation dataset for safety-critical detection when real data is scarce or sensitive, and it couples spill-focused synthetic imagery with precise bounding-box labels so that both Vision-LLMs (VLMs) and conventional object detectors can be adapted under low-data, privacy-constrained industrial conditions. The dataset is defined not merely by image release, but by a controllable synthetic generation framework, AnomalInfusion, that produces context-aware spill scenes with human-guided anomaly placement and diffusion-based inpainting (Baranwal et al., 13 Aug 2025).

1. Official meaning and naming scope

In the literature, the official name SynSpill refers to the industrial spill dataset and pipeline introduced in "SynSpill: Improved Industrial Spill Detection With Synthetic Data" (Baranwal et al., 13 Aug 2025). In that work, a “spill” includes liquids, semi-liquids, stains, and discolorations that visually indicate leakage or residue in an industrial context; examples include oil spills and smears, water puddles, chemical discolorations, and rust-like stains.

A separate paper, "Benchmarking Concept-Spilling Across Languages in LLMs" (Badanin et al., 18 Jan 2026), describes a multilingual benchmark for language spilling and concept spilling, and notes that “SynSpill” may be used only as an informal alias for that benchmark. It is not the official dataset name in that second work. For encyclopedic usage, this distinction is important: SynSpill, in its formal sense, denotes the industrial synthetic vision dataset rather than the cross-lingual semantic benchmark (Badanin et al., 18 Jan 2026).

This naming overlap suggests a potential bibliographic ambiguity in secondary sources. In technical discussion, disambiguation by full title or arXiv identifier is therefore preferable.

2. Problem setting and design objectives

Industrial spill detection is presented as difficult to scale with conventional supervised vision methods because real incidents are rare, sensitive, and often proprietary (Baranwal et al., 13 Aug 2025). The resulting scarcity prevents training robust detectors and also limits manual annotation due to privacy and safety constraints. Even when public images exist, they underrepresent domain-specific textures, lighting, facility layouts, and subtle leak signatures that occur in plants.

The paper positions large VLMs as partially effective but insufficient without adaptation. They have good zero-shot recognition, but without domain adaptation they underperform at precise localization and miss fine-grained visual cues (Baranwal et al., 13 Aug 2025). SynSpill is intended to address precisely this gap: it provides a scalable synthetic corpus that enables effective Parameter-Efficient Fine-Tuning (PEFT) of VLMs and improves object detectors such as YOLOv11 and RF-DETR Base.

A central design objective is not generic anomaly synthesis, but high-fidelity, context-aware spill imagery that is physically plausible in the scene and annotated with precise bounding boxes. This emphasis on plausibility is operationalized through human-guided placement in semantically appropriate regions, such as valves, pipe junctions, and areas near serviceable equipment. The paper treats this as causally meaningful scene construction rather than simple visual augmentation.

3. Dataset composition, class scope, and annotation format

SynSpill consists of 2,000 synthetic images at resolution 1024×1024, with COCO-style JSON annotations for spill bounding boxes (Baranwal et al., 13 Aug 2025). The dataset focuses on positive spill instances; clean backgrounds are generated during Stage 1 of the pipeline, but the released synthetic dataset is described as spill-focused and was used entirely for fine-tuning. Background clutter, occlusions, and subtle texture variations are present by construction through the generative process.

The paper works with eight anomaly classes. Examples explicitly named include oil-spill, floor-stain, and chemical-discoloration, while the supplementary section also mentions oil, water, and rust (Baranwal et al., 13 Aug 2025). Exact per-class counts are not provided. The scenes are industrial environments such as control rooms, factory corridors, concrete floors, and metal catwalks.

Resource	Scale and stated use
SynSpill synthetic dataset	2,000 images; all used for PEFT or detector fine-tuning
Public web-scraped spill data	1,520 images; 520 for evaluation/testing and 100 for PEFT/FT
Proprietary factory dataset	150 images total; 50 for ICL support examples and 100 for evaluation

The dataset has no official train/val/test split; it is used entirely for adaptation, while evaluation is conducted on held-out public and proprietary sets to test sim-to-real generalization (Baranwal et al., 13 Aug 2025). This separation is methodologically significant because the core claim concerns transfer from synthetic spill imagery to real industrial imagery rather than within-dataset benchmarking.

Annotation is human-grounded. In Stage 2, human experts place bounding boxes, and in Stage 3 the inpainting masks derive from those boxes so that the final images retain the same bbox annotations used for training (Baranwal et al., 13 Aug 2025). Quality control includes human-in-the-loop placement for plausibility and deduplication of public data. Inter-annotator agreement is not reported.

4. AnomalInfusion: the synthetic generation pipeline

The synthetic generation framework is named AnomalInfusion and is organized into three stages (Baranwal et al., 13 Aug 2025). Its aim is to produce photorealistic, industrially plausible spill scenes while keeping anomaly location and semantics controllable.

Stage 1: Domain-anchored scene generation. The base model is Stable Diffusion XL (SDXL 1.0). Textual prompts define semantics, architecture, and lighting; IP-Adapter conditioning transfers style priors from 150 real factory images to ground textures, colors, and layouts; and LoRA style modules reinforce industrial realism and consistency. The output is a set of diverse, clean factory backgrounds faithful to plant environments.

Stage 2: Expert-guided anomaly localization. Human-in-the-loop annotation places bounding boxes in semantically appropriate regions, including valves, pipe junctions, and areas near serviceable equipment (Baranwal et al., 13 Aug 2025). The paper states that this encodes causal plausibility—“a valve may leak here”—and avoids unrealistic placements. A plausible implication is that the method uses semantic priors about industrial infrastructure to regularize the synthetic anomaly distribution.

Stage 3: Physically-plausible spill inpainting. Differential inpainting with SDXL-Turbo Inpainting is applied within soft masks generated from the annotated boxes. Conditioning combines spill material prompts such as “glistening black-brown oil,” IP-Adapter references of real spill textures, and LoRA reinforcement of lighting, surface, and style. The output is intended to respect scene geometry, illumination, and surface interactions (Baranwal et al., 13 Aug 2025).

The paper is explicit about what the pipeline does not use. Realism is achieved through diffusion models, style transfer from real factory references, and guided inpainting; there is no game-engine PBR or explicit fluid simulation (Baranwal et al., 13 Aug 2025). Typical rendering parameters are reported as resolution 1024×1024, sampler DDPM-SDE-2m-GPU, scheduler Karras, 64 steps, CFG scale 8, LoRA strength 0.2–0.4, and IP-Adapter strength 0.6. For inpainting, the reported settings are SDXL-Turbo Inpainting, differential diffusion enabled, mask feathering 50 px, mask opacity 75%, and denoise strength 0.5–0.6.

Scenario diversity is introduced through varied scenes, lighting, spill morphology, and material appearance, including semi-transparent stains (Baranwal et al., 13 Aug 2025). Text prompts and style conditioning implicitly randomize scene composition, textures, and illumination, while camera viewpoints and weather are not explicitly simulated. This suggests that SynSpill is primarily a texture-, lighting-, and placement-aware synthetic corpus rather than a full physical simulator.

5. Adaptation regimes and evaluation protocol

The experimental framework combines SynSpill with multiple adaptation paradigms for both VLMs and detectors (Baranwal et al., 13 Aug 2025). The VLM backbone family is Qwen2.5-VL at 3B, 7B, and 32B parameter scales, described as using a Swin-style vision transformer and a causal language decoder. The detector baselines are YOLOv11 and RF-DETR Base.

Three adaptation regimes are defined. Zero-shot inference uses the model as-is. In-Context Learning (ICL) prepends a support set of $k \in \{5,10,15\}$ labeled examples to the prompt with weights frozen, so prediction is conditioned on $S = \{(x_i, y_i)\}_{i=1}^{k}$, i.e., $P(y \mid x, S)$. Parameter-Efficient Fine-Tuning (PEFT) uses LoRA adapters applied to the language pathway only, the vision pathway only, or both jointly (Baranwal et al., 13 Aug 2025).

The paper reports the following LoRA parameterization:

$W' = W + \alpha A B$

and also gives the equivalent common form

$W' = W + B A.$

The training recipe for PEFT is fixed across the reported experiments: 800 steps with early stopping, batch size 8, gradient accumulation 4, effective batch size 32, rank $r=8$, scaling $\alpha = 1/8$, learning rate $5 \times 10^{-5}$, and AdamW (Baranwal et al., 13 Aug 2025). Hardware is reported as NVIDIA A100 80GB, though the approach is described as single-GPU feasible.

Evaluation uses two held-out real-image settings: a public web-scraped set of 520 images and a proprietary factory set of 100 images (Baranwal et al., 13 Aug 2025). The primary adaptation metric is mean hit rate at IoU $\ge 0.5$ per image across classes, while the main comparison table reports mAP@50. The paper defines

$\text{IoU} = \frac{|A \cap B|}{|A \cup B|}$

and

$S = \{(x_i, y_i)\}_{i=1}^{k}$0

used here at $S = \{(x_i, y_i)\}_{i=1}^{k}$1 for IoU $S = \{(x_i, y_i)\}_{i=1}^{k}$2 (Baranwal et al., 13 Aug 2025).

For high-stakes inference, the prompting protocol uses a system prompt that emulates a certified industrial safety inspector who only reports verifiable hazards, with decoding parameters $S = \{(x_i, y_i)\}_{i=1}^{k}$3, nucleus $S = \{(x_i, y_i)\}_{i=1}^{k}$4, and repetition penalty 1.2 (Baranwal et al., 13 Aug 2025). Requested outputs are COCO or YOLO v11 JSON boxes for one of eight anomaly classes.

6. Quantitative performance and ablation findings

The reported results separate zero-shot generalization from adaptation-driven performance (Baranwal et al., 13 Aug 2025). On the public evaluation set, mean hit-rate at IoU $S = \{(x_i, y_i)\}_{i=1}^{k}$5 for Qwen2.5-VL rises from 0.25/0.35/0.42 in zero-shot mode for 3B/7B/32B to 0.46/0.63/0.71 with LoRA-(V+L). On the proprietary evaluation set, the corresponding values rise from 0.11/0.15/0.24 to 0.29/0.34/0.49. Across the same experiments, LoRA-V consistently outperforms LoRA-L, and ICL improves quickly but plateaus beyond 10–15 shots.

The main mAP@50 comparison, trained on synthetic plus public data and evaluated on held-out public and proprietary sets, is as follows:

Model/regime	Public	Proprietary
Qwen-VL 7B (Zero-Shot)	0.35	0.15
Qwen-VL 32B (Zero-Shot)	0.42	0.24
YOLOv11 (FT)	0.81	0.64
RF-DETR (FT)	0.83	0.67
Qwen-VL 7B + LoRA (V+L)	0.78	0.66
Qwen-VL 32B + LoRA (V+L)	0.84	0.71

These results support two claims emphasized by the paper (Baranwal et al., 13 Aug 2025). First, without SynSpill, zero-shot VLMs generalize better than detectors to unseen spill scenarios, but still lag in localization and overall AP. Second, with SynSpill, VLMs and detectors become comparable: Qwen2.5-VL-32B + LoRA (V+L) achieves 0.84 mAP@50 on public data and 0.71 on proprietary data, surpassing RF-DETR on proprietary by +0.04 and matching or exceeding detector performance overall.

The ablations sharpen the interpretation. For Qwen-7B on the public set, zero-shot hit-rate decreases from 0.35 at IoU $S = \{(x_i, y_i)\}_{i=1}^{k}$6 to 0.02 at 0.9, whereas LoRA-(V+L) decreases from 0.63 to 0.13 (Baranwal et al., 13 Aug 2025). The paper interprets this as tighter boxes under stricter criteria after adaptation. It also reports that even with a modest synthetic dataset size of 2,000 images, high-fidelity generation plus human-guided localization yields substantial gains on real-world tests.

The domain-gap analysis is similarly specific. Gains on the proprietary evaluation set, described as industrial CCTV-like scenes, are presented as evidence that SynSpill can bridge sim-to-real gaps (Baranwal et al., 13 Aug 2025). At the same time, failures are more common on structurally novel layouts not well covered by the style priors. This suggests that the synthetic pipeline improves texture-level generalization more reliably than structural generalization.

7. Availability, deployment considerations, and limitations

The project page is listed as synspill.vercel.app, and the paper states that the SynSpill dataset will be made public upon acceptance (Baranwal et al., 13 Aug 2025). The supplementary material indicates release under CC BY-NC 4.0 for non-commercial research use with attribution. The project page is described as hosting documentation of the AnomalInfusion generation pipeline, parameters, prompts, and implementation details used in the study.

The paper also gives practical guidance for adoption (Baranwal et al., 13 Aug 2025). A recommended workflow is to acquire SynSpill with COCO bounding boxes, optionally add a small set of relevant public images such as 100 images, choose either Qwen2.5-VL-7B or 32B for flexible prompting or YOLOv11/RF-DETR for pure detection pipelines, and use LoRA-(V+L) with rank $S = \{(x_i, y_i)\}_{i=1}^{k}$7, AdamW, learning rate $S = \{(x_i, y_i)\}_{i=1}^{k}$8, batch size 8, gradient accumulation 4, and 800 steps with early stopping. Evaluation should use mAP@50 and mean hit-rate at IoU $S = \{(x_i, y_i)\}_{i=1}^{k}$9, with stricter IoU thresholds considered for localization audits.

Several limitations are explicit. SynSpill models static imagery only; temporal dynamics of leak progression are not represented (Baranwal et al., 13 Aug 2025). It uses no explicit fluid physics or PBR, so edge cases involving complex fluid behavior may be underrepresented. Human-guided box placement introduces subjectivity, although the paper also treats that subjectivity as essential for plausibility. Exact per-class counts are not provided, inter-annotator agreement is not reported, and structural variance remains challenging when layouts are not well captured by the style priors.

Ethical and operational notes are also part of the dataset’s framing (Baranwal et al., 13 Aug 2025). SynSpill avoids sensitive real incident footage, and synthetic generation can be tuned without exposing proprietary imagery. When integrated with CCTV feeds, the paper advises adherence to site privacy policies and auditing false positives and false negatives before automating alarms. It further recommends confidence thresholds, rate-limiting for alerts, logging detections with time and location for maintenance triage, and periodic regeneration of site-tuned SynSpill variants followed by renewed PEFT.

Taken together, SynSpill is best understood as a synthetic adaptation dataset plus generation methodology for spill-focused industrial vision. Its significance lies in showing that high-fidelity, plausibly placed synthetic anomalies can support both PEFT of foundation VLMs and fine-tuning of conventional detectors in safety-critical settings where real incident data is scarce, sensitive, or impractical to obtain (Baranwal et al., 13 Aug 2025).