Deep-learning models for PET/CT are typically trained on small, single-centre datasets and degrade when applied to new scanners, sites or rare presentations. In the previous allocation (NAISS 2025/22-1177) we established a benchmarking framework comparing latent-diffusion and VAE-based augmentation with classical and mixup-style baselines for PET/CT image quality assessment (AUROC, AUPRC, ECE) and lesion segmentation (Dice, HD95) across multiple centres and label budgets, using 3,955 GPU-hours (133 % of the allocation). This continuation moves from benchmarking to targeted, trustworthy and deployable generative augmentation, organised in four work packages: WP1 (months 1-4) - Completion of the multi-centre benchmark: additional ablations and multi-seed runs to obtain confidence intervals, and preparation of a manuscript on generative versus classical augmentation for multi-centre PET/CT. WP2 (months 3-9) - Conditional and domain-aware generative augmentation: latent-diffusion models conditioned on lesion attributes (size, location, uptake) to over-sample small and rare lesions, and on acquisition site/scanner to perform style transfer between centres for domain generalisation. The key question is which synthetic samples improve downstream performance, rather than whether generic augmentation helps. WP3 (throughout) - Trustworthiness: calibration and uncertainty estimation of downstream models, fidelity metrics for synthetic PET/CT, and memorisation checks to ensure that generated images do not reproduce training subjects. An optional reader study with clinical collaborators will assess realism. WP4 (months 8-12) - Lightweight and deployable models: knowledge distillation and compression of the best augmented models into lightweight networks, evaluated for accuracy-efficiency trade-offs (parameters, FLOPs, latency) on GPUs and edge hardware, building on our related work on lightweight edge-AI models for surgical image analysis. All data are de-identified and covered by KI ethical approvals; only authorised project members have access, and only aggregated metrics are shared publicly. Software is PyTorch/MONAI in containerised environments. Expected outputs are 2-3 publications acknowledging NAISS and an open-source conditional augmentation toolkit.