I have different species, I'm interested in both DE and exploring the variance of expression in their genes which I calculate from the rlog values, e.g. more variance in genes for one species that may reflect their adaptiveness to heterogenous environment.

I am trying to understand what blind parameter is actually doing, the current recommendation is to set blind=FALSE for downstream analysis, that is NOT to blind the transfrom to the experimental design (which in my case is species). From the help file:

blind: logical, whether to blind the transformation to the experimental design. blind=TRUE should be used for comparing samples in an manner unbiased by prior information on samples, for example to perform sample QA (quality assurance). blind=FALSE should be used for transforming data for downstream analysis, where the full use of the design information should be made. blind=FALSE will skip re-estimation of the dispersion trend, if this has already been calculated. If many of genes have large differences in counts due to the experimental design, it is important to set blind=FALSE for downstream analysis.

I am also read this discussion: https://stat.ethz.ch/pipermail/bioconductor/2014-February/057965.html, where Mike Love said:

When blind=FALSE, the experimental design is only used by the VST and rlog transformations in calculating the gene-wise dispersion estimates, in order to fit a trend line through the dispersions over the mean. Only the trend line is then used by the transformations, not the gene-wise estimates. Therefore, for visualization, clustering, or machine learning applications I tend to recommend blind=FALSE.

The downside of setting blind=TRUE, is that large differences due to the experimental design (e.g., cell types in your case, or different water columns in the linked discussion above), will inflate the gene-wise dispersion estimates. When most of the genes contain such large differences across conditions, this will raise the trend-line, and then the transformed values will be greatly shrunken toward each other for most genes, which is an undesirable loss of signal.

Since I'm interested in gene-wise dispersion itself, it's not clear to me how the transform affects the signal I want to detect. From the discussion I linked, it appears I should set blind=FALSE to avoid shrinking of values toward each other? But I'm not sure how the transform affects the variance?

Or should I calculate variances from just the normTransform or cpm or FPKM?

If you are interested in modeling (?) gene-wise dispersion, then, is that not what DESeq2 and other RNA-seq programs are already doing, starting from the raw counts? - you may have to elaborate on exactly what you want to do. Take a look at the estimateDispersions() function, too.