Synthetic underwater imagery generation and sim-to-real object detection. Generated benthic-survey scenes are domain-randomised through a Jerlov water-column model, then used to train YOLO detectors evaluated against the real DUO dataset.
| Path | Contents |
|---|---|
imggen/ |
The package. All active pipeline code. |
imggen/prompts/ |
Active prompt engines. Superseded ones in prompts/legacy/. |
imggen/generate/ |
Image generation entry points (flux2dev v9, hunyuan v8). |
imggen/dr/ |
Domain randomization and Jerlov water-column physics. |
imggen/train/ |
YOLO training entry points. |
imggen/eval/ |
Evaluation against DUO and real imagery. |
imggen/data/ |
Manifest building, dataset assembly, annotation. |
imggen/analysis/ |
Water statistics, class balance, depth/range, montages, plots. |
archive/generators/ |
Superseded generators, kept as the reproducibility record. |
tools/ |
Standalone dev utilities (preflight, model smoke tests). |
slurm/ |
Stanage batch scripts. |
manifests/, configs/ |
Generation inputs — the reproducibility record. |
configs/experiments/ |
Experiment variant registries consumed by the consolidated runners. |
reports/ |
Measured outputs: class_counts/, water_stats/, analysis/. |
docs/ |
Path-migration map, model comparison notes, the HiDream abandonment record, and runbooks/ for HPC and RunPod. |
runs/ |
Trained models, evals and detections — the source of truth for results. |
logs/ |
Run logs, grouped train/ eval/ annotate/ assemble/ misc/. Git-ignored. |
weights/, outputs/, dataset/, scratch/, literature/, thesis/, tmp/ |
Local artefacts, git-ignored. |
Everything runs as a module from the repository root — no PYTHONPATH, no install step:
python -m imggen.generate.hunyuan_v8 --manifest manifests/benthic-survey-1000-hunyuan.json --out outputs/hunyuan/v8Superseded generators still run, from the same root:
python -m archive.generators.generate_flux2dev_v5 --manifest manifests/pilot.json --out outputs/flux2dev/v5Experiment families that differ only by dataset are driven by a variant flag rather than duplicated scripts:
python -m imggen.train.v9_combined --variant duo_scatter
python -m imggen.eval.v9_combined_on_duo --variant duo_scattertools/ scripts are standalone and take a plain path:
python tools/pod_preflight.py
tools/export_run_bundle.sh v9_flux2dev_duo_dr_v3See docs/PATH_MIGRATION.md for the full old-to-new path map.
- Dataset and model checkpoints: huggingface.co/datasets/Rithish007/MScDissertation (CC-BY-NC-4.0)
- Companion repository: Shallow_Seabed — the descoped Unreal Engine 5.8 simulator arm
Code in this repository is released under the MIT License.
If you use this code, please cite it:
@misc{ramamoorthysathya2026imggen,
author = {Ramamoorthy Sathya, Rithish},
title = {{ImgGen\_AI: Synthetic Underwater Imagery Generation and Sim-to-Real Object Detection}},
year = {2026},
howpublished = {\url{https://github.com/rithish007/ImgGen_AI}},
note = {Code repository. University of Sheffield MSc Robotics Dissertation}
}See CITATION.cff for the machine-readable citation record (also surfaced by GitHub's "Cite this repository" button).