The factor relating CO emission to molecular hydrogen column density, XCO, is still subject to uncertainty, in particular at low metallicity. In this paper, to quantify XCO at two different spatial resolutions, we exploited a dust-based method together with ALMA 12-m and ACA data and H I maps of three nearby metal-poor starbursts, NGC 625, NGC 1705, and NGC 5253. Dust opacity at 250 pc resolution was derived based on dust temperatures estimated by fitting two-temperature modified blackbodies to Herschel PACS data. By using the HI maps, we were then able to estimate dust-to-gas ratios in the regions dominated by atomic gas, and, throughout the galaxy, to infer total gas column densities and H2 column densities as the difference with HI. Finally, from the ACA CO(1–0) maps, we derived XCO. We used a similar technique with 40 pc ALMA 12-m data for the three galaxies, but instead derived dust attenuation at 40 pc resolution from reddening maps based on VLT/MUSE data. At 250 pc resolution, we find XCO ∼ 1022 − 1023 cm−2/K km s−1, 5–1000 times the Milky Way value, with much larger values than would be expected from a simple metallicity dependence. Instead, at 40 pc resolution, XCO again shows large variation, but is roughly consistent with a power-law metallicity dependence, given the Z ∼ 1/3 Z⊙ metal abundances of our targets. The large scatter in both estimations could imply additional parameter dependence, which we have investigated by comparing XCO with the observed velocity-integrated brightness temperatures, ICO, as predicted by recent simulations. Indeed, larger XCO is significantly correlated with smaller ICO, but with slightly different slopes and normalizations than predicted by theory. Such behavior can be attributed to the increasing fraction of CO-faint (or dark) H2 gas with lower spatial resolution (larger beams). This confirms the idea the XCO is multivariate, depending not only on metallicity but also on the CO brightness temperature and beam size. Future work is needed to consolidate these empirical results by sampling galaxies with different metal abundances observed at varying spatial resolutions.