This paper presents a mixture-gas detectable edge-computing device with a generative learning framework for selectivity and accuracy. Mixture-gas detection capability is enabled through two proposed schemes of temperature modulation and cross-iterative-tuning artificial neural network (CIT-ANN). Their related computations are facilitated inside the edge device level, applying analog normalization concepts in the readout integrated circuit (ROIC). This proposed edge platform provides generative training data for mixture-gas detection, allowing much less empirical data for its learning process, especially under mixture gas environment. An edge-computing IoT device prototype was manufactured based on a fabricated ROIC and in-house metal-oxide-semiconductor sensor arrays embedding heater modulation function. Under mixture-gas experiments of NO2 and CO gases, the proposed CIT-ANN together with the heater modulation demonstrated 44% higher recognition performance than in the conventional ANN. The proposed generative learning method showed higher relative label coincidence, achieving 17% higher correlation with real training data than in the conventional mathematical interpolation method