GSL-Diffusion adds LSTM temporal conditioning and an SNR-gated shortcut to cut MIMO channel-estimation error and high-SNR diffusion steps

A time-series-conditioned diffusion estimation framework that denoises the channel in the angular domain and exploits inter-frame temporal correlation by encoding a short sequence of LS observations with an LSTM, going beyond per-snapshot estimation. Existing diffusion-based estimators typically operate on a single snapshot and incorporate SNR as a conditioning variable or positional embedding, leaving temporal correlation underused; this work injects sequence-level conditions into a 3D U-Net through multi-scale cross-attention. Trained and tested on a QuaDRiGa-generated 3GPP TR 38.901 UMa/UMi, LOS/NLOS time-varying dataset (50,000 training and 10,000 test samples), with adjacent-snapshot normalized correlation used to quantify temporal structure; ablations show that removing full LSTM temporal conditioning makes the model worse than separable LMMSE beyond about 10 dB.

A learnable SNR-gated late-fusion shortcut that re-injects the network input into the final decoding stage through a sigmoid gate with trainable center and scale, preserving observation detail at high SNR and relying on the generative prior at low SNR to mitigate over-smoothing. Prior diffusion estimators lack an explicit mechanism to modulate the weight of observation fidelity versus generative prior across a wide SNR range; the gate parameterizes this tradeoff as an SNR-aware, learnable trust mechanism. Ablations show that removing the gated shortcut from the LSTM-conditioned model degrades the high-SNR regime and falls behind LMMSE beyond about 15 dB; keeping the gate but lacking full LSTM temporal conditioning falls behind LMMSE beyond about 10 dB.

Deterministic DDIM-style reverse updates with SNR-adaptive truncation and step allocation, matching observation SNR to the diffusion schedule's effective SNR to set the starting timestep, so high SNR uses few steps and low SNR more steps, improving the accuracy-latency tradeoff. Unlike diffusion estimators with fixed sampling budgets, this strategy allocates reverse steps per sample according to SNR, often reaching the minimum of one step at high SNR while capping the maximum number of steps. Latency measured on an AMD 9950X CPU with an NVIDIA RTX5090 GPU at batch size 1, averaging 50 repeated timed runs after warm-up; per-step runtime is 13.2 ms versus 1.7 ms for the lightweight Fesl et al. baseline, but overall latency is dominated by the number of reverse updates rather than per-step cost.

On standardized time-varying channel simulations, the full GSL-Diffusion achieves the best NMSE almost across the full SNR range from about -10 to 20 dB, outperforming LS, LMMSE and diffusion baselines. Relative to existing diffusion-based channel-estimation baselines, the work combines temporal conditioning and the SNR-gated shortcut, both shown by ablation to be necessary for avoiding high-SNR over-smoothing while retaining low-SNR generative denoising. Results are reported as NMSE (dB) versus SNR and can be grouped by mobility (velocity/Doppler) and scenario; comparisons include LS, LMMSE, and diffusion ablations with different conditioners/shortcuts under the same sampling strategy.

The result targets narrowband MIMO links with omni, single-polarized ULAs, with channels generated by QuaDRiGa under 3GPP TR 38.901 UMa/UMi, LOS/NLOS urban scenarios, snapshot intervals and speed ranges spanning pedestrian to high mobility; the method suits time-varying channels with energy-concentrated angular structure and strong inter-frame correlation, and assumes LS observations and SNR information are available. For physical-layer researchers and implementers seeking to reduce diffusion-estimator inference latency, the work offers reusable design patterns: sequence-level condition encoding, SNR-gated fusion, and SNR-based sampling truncation.

The original presents NMSE and inference-latency curves in Fig. 4 and Fig. 5, but the text does not give per-SNR numeric values, so the precise gaps to each baseline cannot be independently checked from the text alone; the learned gate center and scale values and the actual step-count distribution across SNR are also not listed in the body. The dataset is QuaDRiGa simulation without measured channels or hardware-prototype validation, so behavior when moving to real deployments remains an open question. In addition, the specific maximum step cap and the low-SNR latency bound need to be confirmed against the figures.