README.md

metadata

pretty_name: IRForge artifacts
license: other
tags:
  - infrared
  - image-to-image
  - rgb-to-infrared
  - infrared-small-target-detection
  - lora
  - flux

IRForge artifacts

Data, checkpoints, predictions, and detection artifacts of IRForge, which adapts image-conditioned foundation flow models (FLUX.2-klein-base-4B, Qwen-Image-2.1) to RGB-to-infrared translation and uses them as data engines for infrared small target detection. The code and the reproduction guide (docs/reproduce.md) are in the IRForge repository.

The layout mirrors the code's IRFORGE_WORK directory. MANIFEST.json lists every file with its size; each *.tar unpacks into the directory given by its unpack_to entry. experiments/tools/fetch_release.py PART ... downloads parts and unpacks them (raw, paper, checkpoints, predictions, detection, backgrounds).

Path	Content
data/raw/20260922pbw/*.tar	paired RGB-infrared images with train.txt / test.txt: FLIR-aligned (LWIR) and VEDAI (NIR), full sets and the 1,000-pair Selected subsets
runs/RUN/saves/RUN/checkpoints/checkpoint-N/	PEFT LoRA adapters (Flow-Factory checkpoints): Stage 1 translators, reward-tuned continuations (checkpoint-0 is the starting point), DMD2 / TDM students
eval/RUN/SETTING/SPLIT/	metrics.json, sampling.json (and merge.json for interpolated translators), predictions.tar with the translated test images
irstd/datasets.tar	synthesized SIRST detection datasets in the BasicIRSTD layout, including the SIRST320 split
irstd/runs/DATASET/MODEL_seedS/	detector scores (test_scores.json, val_scores.json) of every detection run
irstd/backgrounds/*.tar	translated Places365 background pools, 4,200 scenes each
irstd/*.json, irstd/*.npz, runs/*/train.log	background analysis, DINOv2 domain probes, and RL training logs read by the paper tables
release/inputs_make_*.txt	every file the paper's table and figure scripts read

Licenses and attribution

• Adapters of FLUX.2-klein-base-4B: the base model is Apache-2.0.
• Adapters and students of Qwen-Image-2.1 derive from a model under the Qwen Research License (non-commercial research and evaluation only); the same terms apply to them.
• FLIR-aligned derives from the Teledyne FLIR ADAS dataset (aligned by Zhang et al., ICIP 2020), VEDAI from Razakarivony and Jurie (2016), SIRST from NUAA-SIRST (Dai et al., 2021), and the background scenes from the Places365 validation set (Zhou et al., 2018). These images, and the translations and detection data derived from them, remain subject to the original datasets' terms; use them for non-commercial research.