In asymmetric, pairing-based cryptography $\mathbb G_1$ is usually a subgroup of an elliptic curve over a prime field $\mathbb F_q$. Elements of this group are usually expressed as a pair of numbers $(x,y)\in(\mathbb F_q)^2$. Computationally, both values are needed, but as $y$ can be recovered from $x$ up to sign elements are often compressed to an $x$ value and an additional bit for transmission purposes. This requires $\lceil\lg q\rceil+1$ bits.
$\mathbb G_2$ is usually a subgroup of the elliptic curve with the same equation but with points of $\mathbb F_{q^k}$ where $k$ is such that $\#\mathbb G_1|(q^k-1)$. Generically, such $k$ are hard to find, but there are various special constructions that parameterise suitable $q$ and curves for particular values of $k$. There is an especially nice family found by Barreto and Nehrig for $k=12$ which allows the whole elliptic curve group to be used for $\mathbb G_1$, which is particularly efficient. An earlier more general construction by Barreto, Lynn and Scott is almost as efficient with $k=12$ and $k=48$. In both the BN and BLS cases, elements of $\mathbb G_2$ can be expressed as a pair $(x,y)\in\mathbb (F_{q^k})^2$. Again compression is possible so that only $x$ and a sign bit need to be transmitted. This would require $k\lceil\lg q\rceil+1$ bits. In the BLS and BN cases, we can choose $\mathbb G_2$ in such a way that $x$ and $y$ and can be derived from a point on a related curve over $\mathbb F_{q^{k/6}}$. In such circumstances, it suffices to transmit a single element of $\mathbb F_{q^{k/6}}$ and a sign bit. This would require $\frac k6\lceil\lg q\rceil+1$ bits. However this choice of $\mathbb G_2$ is not compatible with all uses of pairing-based cryptography.
With such choices of $\mathbb G_1$ and $\mathbb G_2$ the various cryptographic pairing all map to $\mathbb G_T$ which is a multiplicative subgroup $\mathbb F_{q^k}$ of order $\#\mathbb G_1$. Elements of this group can be written as elements of $\mathbb F_{q^k}$ which takes $k\lceil\lg q\rceil$ bits.
The choice of $q$ and $k$ will depend on the level of security that you want your pairing-based system to have. The size of $\mathbb G_1$ needs to be large enough enought to block generic ``square root attacks'' and the size/structure of $\mathbb G_T$ needs to be sufficient to block the TNFS attack of Kim and Barbalescu. A 2019 draft from the IETF suggests the following in section 4.
Security (in bits) |
Size of $\mathbb G_1$ (uncomp./comp.) |
Size of $\mathbb G_2$ (uncomp./comp./BN-BLS comp.) |
Size of $\mathbb G_T$ |
100 |
512/257 |
6144/3073/513 (BN256, $k=12$) |
3072 |
128 |
924/463 |
11088/5545/925 (BN462, $k=12$) |
5544 |
128 |
922/462 |
11064/5533/923 (BLS12-461, $k=12$) |
5532 |
128 |
762/382 |
9144/4573/763 (BLS12-381, $k=12$) |
4572 |
256 |
1162/582 |
55776/27889/4649 (BLS48-581, $k=48$) |
27888 |
Note that this is a purely classical security estimate and like all pairing systems these should be considered vulnerable to a cryptanalytically relevant quantum computer.