This paper is devoted to studying the relationship between two fundamental objects in physics, associated to finite graphs: the two-dimensional Ising model and the spin network evaluations on planar graphs.Given a graph Γ and a coloring of the edges of Γ, i.e. a map c : E → N where E is the set of edges, the classical spin network evaluation Γ, c is a rational number which is the result of contracting some tensors over irreducible representations of SU(2). Spin networks arise in many areas, in particular related to physics. Since they come from the representation theory of SU(2), they are objects of prime interest in the theory of quantum angular momentum [1] where they are often called Wigner symbols. As such, they have applications in atomic/molecular physics, chemistry, quantum information and so on [2]. More recent applications stem from quantum gravity as spin network equipped with holonomies are the states of loop quantum gravity [3,4] while their evaluations provide quantum gravity amplitudes, known as spin foams, [5][6][7]. The latter are intimately related to lattice topological invariants, such as the Reidemeister torsion [8-10] and the Turaev-Viro invariant of 3-manifolds.Spin network evaluations can be computed in many different ways. The most famous is certainly the combinatorial definition due to R. Penrose [11]. In quantum gravity one often uses contractions of SU(2) intertwiners (notice that it requires an orientation on the edges while Penrose's definition does not -we will prove the equivalence between those evaluations in the main text). In the last years it has been understood [12-16] that a good way to study spin network evaluations on a fixed graph Γ is to organize it in a single generating series Z spin (Γ) := c Γ, c Y c where the symbols Y c stand for a suitable multivariate monomial in formal variables Y e , e ∈ E (full details will be provided later).On a seemingly different side of physics (and mathematical physics), the 2D Ising model can be defined on the same graph Γ. It probably is the most famous statistical model, based on the configuration space of maps from the set V of vertices of Γ into {±1}. The Hamiltonian (energy function) of the model is a sum of interactions between nearest-neighboring sites of Γ, hence associated to the set E of edges of Γ, and weighted by couplings y e , e ∈ E. The partition function Z Ising (Γ) was proved by van den Waerden to be proportional to a sum over even subgraphs of Γ weighted by some monomials in tanh(y e ), e ∈ E(Γ) (see [17] for instance).The present paper is motivated by the following observation made in [21]; if Γ is a planar trivalent graph and for each edge e ∈ E we set Y e := tanh(y e ) then the following equality holds: Unless explicitly stated the contrary we will assume that Γ is planar, i.e. embedded (up to isotopy) in S 2 . This automatically equips Γ with the datum of a cyclic, counter-clockwise ordering of the edges around each vertex. For each angle α around a vertex, following the cyclic ordering around that vertex, we call s(α...