Synthesizing realistic radio-frequency signals is important for data augmentation in wireless systems, yet labeled in-phase/quadrature (I/Q) measurements are expensive to collect, and diffusion models typically demand large per-class training sets. This paper studies whether a diffusion model trained on one modulation scheme can be transferred to unseen schemes with minimal data and a minimal parameter budget. We formulate the adaptation as a minimal-cardinality parameter-increment problem and adapt a conditional diffusion transformer, pretrained on QPSK signals from the RML2018.01a corpus, to three target modulations (OQPSK, 16PSK, and 32QAM) through low-rank adaptation (LoRA) of its attention projections under a composite objective that couples waveform reconstruction with spectral and autocorrelation constraints, updating only 0.496% of the 9.9-million-parameter backbone. With only 50 target signals, the adapted models attain a periodogram cosine similarity of 0.985 against 0.705 for from-scratch training and match from-scratch models given 1000 signals: a twenty-fold gain in sample efficiency. The transferred models also match or exceed reference target experts trained on up to 333 times more target data on every spectral and autocorrelation metric. Ablations on two target schemes show that sources with dense phase manifolds transfer best irrespective of family labels and that the generation quality is insensitive to the LoRA rank between 1 and 16, pointing to a low-dimensional adaptation subspace. Downstream classifier experiments confirm that the synthetic signals help few-shot modulation classification, and they also bound the claim: conventional label-preserving transforms remain stronger at very small sample sizes, downstream utility does not track spectral fidelity, and the generated signals under-represent the noise floor in the low-SNR regime. The main study uses RML2018.01a; a cross-corpus check on the independently generated Sig53 corpus reproduces the transfer advantage with compressed margins, and validation on over-the-air recordings remains future work.

Evaluation of generative models for RF (Radio Frequency) signals remains largely ad hoc, with existing approaches relying on uncalibrated metrics imported from computer vision without systematic justification or baseline establishment. We present a multi-dimensional, distribution-level evaluation framework comprising ten metrics across six layers, calibrated against real-real baselines. A real-real baseline is computed by splitting authentic signals into two independent subsets and measuring the same metric between them; the resulting value sets the achievable ceiling for that metric, against which generated-vs-real scores are normalized. The framework is grounded in four design principles: distribution-level aggregation, real-real baseline normalization, signal-to-noise ratio (SNR)-modulation stratification, and multi-dimensional coverage. Application of the framework reveals two previously unreported phenomena. First, linear Short-Time Fourier Transform preprocessing creates a gradient imbalance between frequency-domain and time-domain loss components that causes catastrophic generation failure for analog amplitude modulation; logarithmic compression resolves this. Second, the widely adopted noise-prediction training objective exhibits systematic gradient suppression for amplitude-modulated signals due to SNR-dependent implicit loss weighting; switching to signal-prediction substantially improves temporal structure fidelity for amplitude-modulated signals, with only modest trade-offs for digital communication signals. The framework and calibration methodology establish reproducible standards for comparative assessment of RF generative models.