TamAko783/TPDD_Honglin_CLS_Synthetic

TLA Tobacco Disease — SYNTHETIC variant (AI-generated)

⚠️ These images are AI-GENERATED, not real data

Each image was produced by Stable Diffusion img2img (stabilityai/sd-turbo) seeded from a real long-tail training image of the TLA dataset (one source image → multiple generated variants). They are provided only as an experimental augmentation pool for the rare classes.

Do NOT use them for evaluation. Train-time augmentation only. For honest benchmarking use the real val/test of the main dataset.

Relationship to the real dataset

This is a separate, derived companion to the real, credited dataset: TamAko783/TPDD_Honglin_CLS.

The real source images were collected & annotated by Hong Lin, Rita Tse, Su-Kit Tang et al. (Macao Polytechnic University) — TLA / TPDD. Please cite:

• TPDD — ICDIP 2022, DOI 10.1117/12.2644288
• TLA / FREN — Frontiers in Plant Science 2024, DOI 10.3389/fpls.2024.1333236

How it was generated

• Pipeline: AutoPipelineForImage2Image("stabilityai/sd-turbo"), fp16, on an NVIDIA L4. img2img strength = 0.30, 8 steps (~2 effective), guidance 0, disease-aware prompt per class, deterministic seeds. The low strength keeps the variants faithful to the real leaf morphology and lesions (a higher strength of 0.55 was rejected — it drifted off-tobacco and invented symptoms).
• Targets: the long-tail classes (anthracnose, black_shank, tswv, genetic_abnormality, pvy). K variants per real source image.
• Layout: train/<slug>/<source>_synth{k}.jpg. Full provenance (source image, prompt, seed, strength) in provenance.csv.

⚠️ Quality caveat

sd-turbo is a fast, general model with no plant-pathology grounding. At the low strength used here variants stay close to the real leaf, but some may still slightly distort symptoms. Review before use and treat as a synthetic augmentation experiment (Tier-4 in the project plan). For higher fidelity, fine-tune a class-conditional diffusion / StyleGAN2-ADA model on the real images and quality-gate by FID/KID + manual review.