Thermal infrared (TIR) cameras offer robust perception in adverse conditions, but standard vision backbones cannot directly ingest raw 14- or 16-bit thermal data without compressing it into an 8-bit, multi-channel format. Existing tone-mapping methods either use rigid heuristics that fail to generalize across environments or require expensive radiometric hardware constrained by single-task trade-offs. To resolve this, we introduce a lightweight pre-encoder for adaptive tone-mapping of non-radiometric thermal imagery. By parameterizing sigmoid curves via learnable center and width variables, the module dynamically maps raw frames into a 3-channel representation optimized end-to-end via task loss without requiring radiometric data. This prevents thermal outliers from washing out background contrast while retaining target details. Evaluations across multiple benchmarks ($\text{MS}^2$, ViViD++, and FLIR ADAS) for monocular depth estimation (under frozen, fine-tuned, and from-scratch regimes) and 2D object detection demonstrate that our method consistently improves performance and enhances zero-shot cross-domain generalization. ... Read more