Update 2: a more streamlined version for 2D graphs:

ClearAll[g]

g = GraphComputation`ExpressionGraph[ConstantArray[x, #], VertexLabels -> None] &;

Examples:

g[Range[2, 4]]

SetProperty[g[{3, 1, 3, 1, 2, 1, 4}],

  {GraphLayout -> "RadialEmbedding", EdgeShapeFunction -> "Line"}]

Original answer:

ClearAll[f]

f[g_: Graph][n_List, o : OptionsPattern] := g[UndirectedEdge @@@ EdgeList@

    GraphComputation`ExpressionGraph[ConstantArray[x, n]], 

     o, GraphLayout -> {"BalloonEmbedding"}, ImageSize -> Large]

f[g_: Graph][n_Integer, o : OptionsPattern] := f[g][Range[n], o]

Examples:

f[6]

f[6, GraphLayout -> {"RadialEmbedding"}]

g1 = f[Graph3D][6]

g2 = f[Graph3D][6, GraphLayout -> {"RadialEmbedding"}]

Use a list for number of vertices on each layer as the argument:

f[{3, 5, 2, 4}, GraphLayout -> {"RadialEmbedding", "LayerSizeFunction" -> (# &)}]

Update: ... to take the Graph3D and somehow improve the layout on the high-n layers: There seems to be lots of wasted space.

One way to change the box ratios without distorting the vertex shapes is to modify the VertexCoordinates using ScalingTransform:

SetProperty[g1, VertexCoordinates -> ScalingTransform[{1, 1, 3}][GraphEmbedding@g1]]

SetProperty[g2, VertexCoordinates -> ScalingTransform[{1, 1, 3}][GraphEmbedding@g2]]

Or add the suboption "LayerSizeFunction" in "RadialEmbedding":

g3 = f[Graph3D][6, GraphLayout -> {"RadialEmbedding", "LayerSizeFunction" -> (# &)}];

SetProperty[g3, VertexCoordinates -> ScalingTransform[{1, 1, 3}][GraphEmbedding@g3]]

SetProperty[#, VertexCoordinates -> ScalingTransform[{1, 1, 3}][GraphEmbedding@#]] &@

 f[Graph3D][{3, 5, 2, 4}, GraphLayout -> {"RadialEmbedding", "LayerSizeFunction"->(#&)}]

IGraph/M already has this built-in as IGSymmetricTree. You can specify the number of branches at each level.

IGSymmetricTree[

 Range[2, 4],

 DirectedEdges -> True,

 GraphLayout -> "LayeredEmbedding"

]

The implementation is mostly in C (not Mathematica) and not from igraph this time. This is simply easier to implement procedurally, for which C is a good fit. This is why I did not do it in pure WL.

Here's another structure, with a different number of branches at each level.

IGSymmetricTree[{5, 4, 3, 2}]

I don't know if you find this elegant. But I give it a try.

maxdepth = 5;

Graph[

 Transpose[{

   Join @@ MapIndexed[

     {x, i} [Function] Join @@ Transpose[ConstantArray[Range @@ x, (i + 1)]], 

     Join[{{1}}, Partition[Accumulate[Range[maxdepth - 1]!], 2, 1] + ConstantArray[{1, 0}, maxdepth - 2]]

     ],

   Range[2, Total[Range[maxdepth]!]]

   }],

 DirectedEdges -> True,

 GraphLayout -> "BalloonEmbedding"

 ]

Edit

Out of curiosity, I adapted the algorithm above to produce also other symmetric trees.

SymmetricTree[branchlist_?(VectorQ[#, IntegerQ] &)] := 

 Module[{levelnodecounts}, 

  levelnodecounts = FoldList[#1 #2 &, 1, branchlist];

  Graph[Transpose[{

     Join @@ MapIndexed[

       {x, i} [Function] Join @@ Transpose[ConstantArray[Range @@ x, branchlist[[i[[1]]]]]],

       Join[

        {{1}},

        Partition[Accumulate[Most[levelnodecounts]], 2, 1] + ConstantArray[{1, 0}, Length[branchlist] - 1]

        ]

       ],

     Range[2, 1 + Total[Rest[levelnodecounts]]]}], 

   DirectedEdges -> True

   ]

  ]

Regarding speed, it seems to be on par with IGSymmetricTree. Of course, I cannot provide such a detailed user interface as Szabolcs so I would suggest to use IGraphM whenever possible.

Edit 2

Adapting my (slow) code for fractal trees, here is another way to embedd the tree:

BoccoliEmbedding[branchlist_] := 

 Module[{data0, data, θ, stem, thickness, s1, s2, f, F},

  θ = Pi/4.;

  s1 = 1/GoldenRatio // N;

  s2 = 1/GoldenRatio // N;

  stem = {0., 0., 1.};

  thickness = 0.15;

  data0 = {Join[

     {{0., 0., 0.}},

     {stem},

     {{thickness, 1., 0.}},

     Table[

      RotationMatrix[2. k Pi/branchlist[[1]], {0, 0, 1}].{Cos[θ], 0.,Sin[θ]}, 

      {k, 0, branchlist[[1]] - 1}]

     ]

    };

  f = {U, n} [Function] Table[

     Join[

      {U[[1]] + U[[2]]},

      {U[[i]]},

      {s2 U[[3]]},

      Dot[

       s1 Table[RotationMatrix[2. Pi j/n, U[[2]]].U[[i]], {j, 0, n - 1}], 

       RotationMatrix[{U[[i]], U[[2]]}]

       ]

      ],

     {i, 4, Length[U]}];

  F = {data, n} [Function] Join @@ (f[#, n] & /@ data);

  data = Join @@ FoldList[F, data0, Join[Rest[branchlist], {1}]];

  data[[All, 1]] + data[[All, 2]]

  ];

And this is how we apply it:

b = Range[2, 7];

plot = Graph[

  EdgeList[SymmetricTree[b]],

  VertexCoordinates -> BoccoliEmbedding[b]

  ]

Based on @kglr's answer, but avoiding the use of undocumented functions (ExpressionGraph):

These trees correspond to the expression structure of a full array of dimensions {2, 3, 4, ...} (think TreeForm).

To recover this tree, we walk the array expression using Position and record the positions of subexpressions. We will use these positions as graph vertices.

The position specification is such that we can always compute the parent node of a position by dropping its last element (implemented as [[;;-2]]). For example, the parent subexpression of a subexpression at position {2,1} is found at position {2}.

Position walks the expression in post-order, meaning that it returns the root vertex (position of full expression, {}) last. We reverse the vertex list to get the root as the first vertex instead. Then we drop this root from the edge computation (MapThread) as it has no parents.

Finally, we replace vertex names with integer vertex indices using IndexGraph.

symmetricTree[levels_List] :=

 Module[{vertices, edges},

  vertices = Reverse@Position[ConstantArray[0, levels], _, {0, Infinity}, Heads -> False];

  edges = MapThread[DirectedEdge, {vertices[[2;; , ;;-2]], vertices[[2;;]]}];

  IndexGraph@Graph[v, e]

 ]

I find this method clear and readable.

symmetricTree[{2, 3, 4}]