pith. sign in

arxiv: 2605.11276 · v1 · submitted 2026-05-11 · 💻 cs.CV · cs.AI

Generative AI for Visualizing Highway Construction Hazards Through Synthetic Images and Temporal Sequences

Trevor Neece , Mason Smetana , Lev Khazanovich This is my paper

Pith reviewed 2026-05-13 06:18 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords generative AIsynthetic imagesconstruction safetyhazard visualizationOSHA reportsimage generationsafety training
0
0 comments X

The pith

Generative AI creates synthetic images of highway construction hazards from injury reports that experts rate 81 percent educationally acceptable.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops methods to convert written OSHA reports of severe highway construction injuries into visual training materials using generative models. It produces either one static image per incident or a four-stage sequence showing how the hazard unfolds over time. Single images receive higher expert marks for fidelity and usefulness than sequences, and both pass tests showing they match the original descriptions in semantic retrieval. This approach sidesteps the difficulty of obtaining real photographs of dangerous events while still supplying concrete visuals for safety lessons.

Core claim

Autoregressive image generation models can synthesize visualizations of highway construction hazards directly from OSHA Severe Injury Report narratives, producing single-pass images that achieve 81.1 percent educational acceptability with fidelity of 4.14 out of 5 and alignment of 4.07 out of 5, while temporal sequences reach 60.9 percent acceptability with alignment of 3.94 out of 5 but lower fidelity of 3.51 out of 5, and both modes demonstrate statistically significant CLIP-based semantic retrieval capabilities.

What carries the argument

Single-pass and temporal generative pipelines that translate incident narratives into one image or four-stage image sequences, measured by expert ratings on educational utility, fidelity, and alignment plus CLIP semantic retrieval.

If this is right

  • Safety trainers can generate visual materials directly from existing narrative reports without photographing active construction sites.
  • The multi-dimensional evaluation method using expert scores and retrieval tests can be reused for synthetic images in other high-risk domains.
  • Both single images and sequences produce outputs that retrieval models can match back to the source descriptions at statistically significant levels.
  • Large collections of injury reports can be turned into scalable visual training libraries.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Pairing the generated images with actual field tests of worker performance would reveal whether the expert ratings translate into measurable safety gains.
  • Generating short video clips instead of static sequences might raise the fidelity scores for the temporal mode.
  • The same narrative-to-image pipeline could be applied to injury reports from manufacturing, mining, or transportation to create domain-specific training sets.

Load-bearing premise

Expert ratings of educational utility and CLIP semantic retrieval scores are sufficient to indicate whether the generated images will actually help workers learn to avoid hazards on the job.

What would settle it

A controlled study that measures actual hazard recognition accuracy and safety behavior changes in workers trained with the synthetic images versus those trained with real photographs or text alone.

Figures

Figures reproduced from arXiv: 2605.11276 by Lev Khazanovich, Mason Smetana, Trevor Neece.

Figure 1
Figure 1. Figure 1: Graphical overview of the proposed contributions of this study. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed methodology, illustrating the flow from OSHA SIR nar [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of highway construction images generated from the SIR database. (A) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of hallucinations by category: (A) processing artifact: head and shoulder [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of cosine similarity scores between CLIP-encoded text descriptions and [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sample image of the evaluation platform used by reviewers to evaluate the single [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Single-pass outputs of the example OSHA narrative. (A) Iteration 1 and (B) Iteration [PITH_FULL_IMAGE:figures/full_fig_p034_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Temporal sequence outputs of the example OSHA narrative. (A) Iteration 1 and (B) [PITH_FULL_IMAGE:figures/full_fig_p036_8.png] view at source ↗
read the original abstract

Highway construction workers face a high risk of serious injury or death. Image-based training materials depicting hazardous scenarios are essential for engaging safety instruction but remain scarce due to ethical and logistical barriers. This study develops and evaluates a generative AI methodology for producing synthetic visualizations of highway construction hazards from OSHA Severe Injury Report narratives. Two modes were developed: a single-pass approach yielding one image per incident, and a temporal approach producing a four-stage sequence. A sample of 75 incident records yielded 750 images, evaluated using CLIP-based semantic retrieval and expert assessment across dimensions such as educational utility, fidelity, and alignment. Single-pass images achieved 81.1% educational acceptability with fidelity and alignment scores of 4.14/5 and 4.07/5, respectively, while temporal sequences achieved 60.9% acceptability with comparable alignment (3.94/5) but lower fidelity (3.51/5). CLIP-based retrieval revealed that both modes produce images with statistically significant retrieval capabilities. This is among the first studies to leverage modern autoregressive image generation models for visualizing construction hazards from reported severe injuries and to generate temporally sequenced hazard imagery, and a new multi-dimensional evaluation framework was developed to support future research in this domain. The work enables safety trainers to pair narrative storytelling with visual learning material without photographing real-world hazards, and the framework could be applied to datasets across diverse domains, enabling synthetic image generation tailored to new application areas.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript develops two generative AI pipelines (single-pass and temporal) to synthesize images of highway construction hazards from 75 OSHA Severe Injury Report narratives, producing 750 images total. These are evaluated via CLIP-based semantic retrieval and expert ratings on educational utility, fidelity, and alignment, yielding 81.1% educational acceptability (fidelity 4.14/5, alignment 4.07/5) for single-pass images and 60.9% (alignment 3.94/5, fidelity 3.51/5) for temporal sequences, with statistically significant retrieval for both. The authors position this as enabling visual safety training materials without real-world photography and introduce a multi-dimensional evaluation framework.

Significance. If the proxy metrics translate to real training gains, the work provides a scalable, ethical method for generating hazard visualizations in a domain with scarce authentic images, along with a reusable evaluation framework. The temporal sequencing aspect and application to construction safety are distinctive contributions that could extend to other high-risk fields. However, the absence of direct outcome measures substantially limits the assessed significance.

major comments (2)
  1. [Abstract] Abstract: The central claim that the generated images are suitable replacements for scarce real hazard photographs in safety training is load-bearing but rests solely on expert Likert-scale acceptability (81.1% single-pass) and CLIP retrieval; no controlled measurement of downstream worker outcomes (hazard recognition accuracy, retention, or behavioral change) or head-to-head comparison against authentic OSHA photographs is reported, leaving the mapping from reported scores to training efficacy unestablished.
  2. [Evaluation] Evaluation section: The reported expert educational acceptability percentages and fidelity/alignment scores lack accompanying details on expert count, qualifications, selection criteria for the 75 incidents, inter-rater agreement statistics, or the precise statistical tests and p-values underlying the 'statistically significant' CLIP retrieval claims, which are required to substantiate the quantitative results.
minor comments (2)
  1. [Abstract] Abstract: The description of image generation ('750 images' from 75 records) would benefit from explicit clarification on whether multiple variants were produced per incident and how prompt engineering or model parameters were controlled.
  2. [Discussion] The manuscript would be strengthened by adding a dedicated limitations paragraph discussing the proxy nature of the current metrics and outlining planned or recommended follow-up studies with actual trainees.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help clarify the scope and limitations of our work. We address each major point below and will revise the manuscript to improve transparency and temper claims where appropriate.

read point-by-point responses

  1. Referee: [Abstract] The central claim that the generated images are suitable replacements for scarce real hazard photographs in safety training is load-bearing but rests solely on expert Likert-scale acceptability (81.1% single-pass) and CLIP retrieval; no controlled measurement of downstream worker outcomes (hazard recognition accuracy, retention, or behavioral change) or head-to-head comparison against authentic OSHA photographs is reported, leaving the mapping from reported scores to training efficacy unestablished.

    Authors: We acknowledge that our evaluation uses proxy metrics (expert ratings and CLIP retrieval) rather than direct measures of training outcomes or comparisons to real photographs. This study was designed as an initial demonstration of generative feasibility and a new evaluation framework; controlled worker studies measuring hazard recognition or behavioral change were outside its scope due to resource and ethical constraints. In revision we will update the abstract, introduction, and discussion to frame the results more cautiously as preliminary indicators of utility, explicitly note the absence of downstream efficacy data, and call for future validation studies. revision: partial

  2. Referee: [Evaluation] The reported expert educational acceptability percentages and fidelity/alignment scores lack accompanying details on expert count, qualifications, selection criteria for the 75 incidents, inter-rater agreement statistics, or the precise statistical tests and p-values underlying the 'statistically significant' CLIP retrieval claims, which are required to substantiate the quantitative results.

    Authors: We agree these details are necessary and were inadvertently omitted. The revised Evaluation section will specify: five experts with professional experience in construction safety and OSHA reporting; random selection of the 75 incidents from the OSHA Severe Injury Report database; inter-rater agreement via Fleiss' kappa; and exact statistical methods and p-values for CLIP retrieval (all p < 0.01). These additions will be included in the next version. revision: yes

Circularity Check

0 steps flagged

Empirical evaluation with external human and CLIP benchmarks; no derivation reduces to self-inputs

full rationale

The paper is an applied empirical study that generates images from external OSHA narratives using off-the-shelf autoregressive models, then measures outputs via independent expert Likert ratings and CLIP semantic retrieval. No equations, fitted parameters, or self-citations are load-bearing; the reported 81.1% acceptability and retrieval statistics are direct measurements against external benchmarks rather than quantities defined or forced by the generation pipeline itself. The work contains no mathematical derivation chain that could exhibit self-definition, fitted-input renaming, or uniqueness smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that current text-to-image models can translate narrative descriptions of construction incidents into educationally useful visuals; no new free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Modern autoregressive image generation models can produce images that align with textual descriptions of real-world hazards
    Invoked when the paper states that single-pass and temporal outputs were generated from OSHA narratives.

pith-pipeline@v0.9.0 · 5564 in / 1239 out tokens · 34829 ms · 2026-05-13T06:18:06.770992+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [1]

    Identify which worker is in the danger zone and should be highlighted with a red outline or glow

    work page

  2. [2]

    Watch for pinch points

    Generate a concise safety warning phrase (5–7 words max) based on the upcoming hazard from the OSHA data:{json data}. Safety Warning Examples: - “Watch for pinch points” - “Watch for tipping vehicles” - “Maintain safe distance from equipment” - “Watch for overhead hazards” - “Stay clear of swing radius” - “Watch for struck-by hazards” - “Beware of fall ha...

    work page

  3. [3]

    No machines or workers

    THE INFRASTRUCTURE: Describe the base roadway or work zone layout (e.g., pavement, lane markings, traffic control). No machines or workers

    work page

  4. [4]

    THE ACTIVITY: Introduce the construction equipment and workers performing their routine task within that space

    work page

  5. [5]

    Constraints: 1–2 sentences per step

    THE HAZARD: Describe the peak of the event, focusing on the physical interaction between the worker and the hazard ({event keyword}). Constraints: 1–2 sentences per step. Start the response immediately with the Step 1 description. No dramatization or inventions. Output Format: Output as a single cohesive paragraph. Plain text only. No preamble. 27 A.2 Ima...

    work page

  6. [6]

    Style and Purpose: Use realistic photography with bright, even lighting suitable for an educational manual

    work page

  7. [7]

    Only include equipment or structures explicitly mentioned in the description

    Minimize Clutter: The background should be relatively clean and minimalist. Only include equipment or structures explicitly mentioned in the description. Avoid random debris or complex textures that distract from the main subject area

    work page

  8. [8]

    Text is only permitted if it appears naturally on environmental objects, such as safety signs, labels on equipment, or vehicle branding

    Text Rules: Do not include overlayed captions, watermarks, or UI elements. Text is only permitted if it appears naturally on environmental objects, such as safety signs, labels on equipment, or vehicle branding

    work page

  9. [9]

    28 State 2: Construction Activity Image (VT2 ) You are generating an image

    Do not have the GUI of Google Streetview overlayed on the image. 28 State 2: Construction Activity Image (VT2 ) You are generating an image. Acting as a VFX editor for an educational series, modify the provided input image to integrate the next step of the safety sequence. This state introduces construction equipment and workers performing their routine t...

    work page

  10. [11]

    The existing background should remain distinct but secondary to the action

    Focal Point: The new elements (workers and construction equipment) must be the sharpest and most distinct part of the image. The existing background should remain distinct but secondary to the action

    work page

  11. [12]

    Clean Integration: Add the new elements seamlessly into the existing geometry with- out adding unnecessary environmental clutter around them

    work page

  12. [13]

    Text must only exist on physical objects within the scene, such as hazard signs or worker PPE

    Text Rules: No overlayed text, labels, or captions. Text must only exist on physical objects within the scene, such as hazard signs or worker PPE

    work page

  13. [14]

    29 State 3: Safety Warning Overlay Image (VT3 ) You are generating an image

    Although the critical positioning requirement describes the hazard, this image must not depict the hazard event itself. 29 State 3: Safety Warning Overlay Image (VT3 ) You are generating an image. Acting as a VFX editor for an educational safety training series, modify the provided input image to add a safety warning overlay. This state highlights the wor...

    work page

  14. [16]

    The highlight should be clearly visible but not obscure the worker’s details

    Worker Highlight: Add a distinct red outline, glow, or semi -transparent red overlay around the worker identified as being in danger. The highlight should be clearly visible but not obscure the worker’s details

    work page

  15. [17]

    SAFETY W ARNING

    Safety Warning Banner: Display the safety warning text prominently in the upper or lower third of the image. Use a high-contrast format: white or yellow bold text on a red or dark background banner. The text should be large and readable. The safety warning text should start with “SAFETY W ARNING”

    work page

  16. [18]

    Only add the highlight effect and text overlay

    Preserve Scene: Do not alter the existing workers, equipment, or environment. Only add the highlight effect and text overlay

    work page

  17. [19]

    30 State 4: Hazard Event Image (VT4 ) You are generating an image

    Educational Tone: The overall effect should resemble a training video freeze-frame or safety manual illustration with clear visual callouts. 30 State 4: Hazard Event Image (VT4 ) You are generating an image. Acting as a VFX editor for an educational series, modify the provided input image to integrate the next step of the safety sequence. This state intro...

    work page

  18. [20]

    Maintain a realistic eye-level perspective

    Move Camera: Feel free to adjust the camera angle so that the health and safety haz- ard and worker interaction are clearly visible. Maintain a realistic eye-level perspective. Do not change background info

    work page

  19. [21]

    The existing background should remain distinct but secondary to the action

    Focal Point: The new elements (the hazard or the incident) must be the sharpest and most distinct part of the image. The existing background should remain distinct but secondary to the action

    work page

  20. [23]

    Text should only exist on physical objects within the scene, such as hazard signs or worker PPE

    Text Rules: No overlayed text, labels, or captions. Text should only exist on physical objects within the scene, such as hazard signs or worker PPE. Single-Pass Hazard Image (VSP ) You are generating an image taken at eye-level, as if an inspector took the photo while standing a few feet away. Generate a photorealistic educational visualization of a const...

    work page

  21. [24]

    Do not zoom, pan, or reframe the shot

    Lock Camera: Maintain the exact camera angle, perspective, and lens focal length of the input image. Do not zoom, pan, or reframe the shot

    work page

  22. [25]

    The existing background should remain distinct but secondary to the action

    Focal Point: The new elements (workers, the hazard, or the incident) must be the sharpest and most distinct part of the image. The existing background should remain distinct but secondary to the action

    work page

  23. [26]

    Clean Integration: Add the new elements seamlessly into the existing geometry without adding unnecessary environmental clutter around them

    work page

  24. [27]

    No Issues — Fully Acceptable

    Text Rules: No overlayed text, labels, or captions. Text should only exist on physical objects within the scene, such as hazard signs or worker PPE. 31 B Evaluation Interface Figure 6 provides a sample screenshot of the Google Form used as the evaluation platform for the expert-based review. Reviewers filled out this form twice for each OSHA narrative rev...

    work page 2000