SpeedTree to Nanite foliage: the whole chain, with the numbers

If you search for how to get SpeedTree vegetation into Unreal 5 as Nanite foliage, you will find a handful of paid courses, two or three YouTube videos, and a pirated copy of one of the courses. What you will not find is a free, current, written account of the whole chain with numbers attached.

That is a strange gap, because this is not a niche problem — it is the core technical question for anyone doing natural environments, and the answer changed substantially with Nanite foliage support. So this is that account. It is written against UE 5.7 and SpeedTree for Games 9, and every number in it was measured on the same forest scene.

The decision that dominates everything else

Before any export settings: masked or opaque?

Traditional game foliage uses alpha-masked cards — a quad with a leaf texture and an opacity mask. It is cheap in triangles and it has been the standard for twenty years. Nanite changes the arithmetic, because Nanite does not love masked materials. Masked geometry forces a separate pass, defeats some of Nanite’s rasterisation advantages, and interacts badly with Lumen, where every masked surface is additional work for the ray tracer.

The alternative is modelled leaves: actual geometry per leaf cluster, opaque material, no opacity mask at all. Twenty years ago that was absurd. With Nanite it is often faster, because Nanite’s whole premise is that triangle count stops being the thing you budget.

Here is what that swap actually did on the test scene.

The cost is real: disk size roughly doubles, and the SpeedTree work is more involved because you are modelling leaf clusters rather than mapping cards. But frame time is what ships a scene, and the shadow and Lumen savings compound with every additional light.

The export chain, step by step

1. In SpeedTree

2. On import into Unreal

Enable Nanite on the trunk and branch meshes immediately. For the leaf clusters, enable Nanite and set the material to Opaque — if it is still Masked you have not actually made the swap and you will get the worst of both approaches.

Then the setting almost everyone misses:

3. Wind that does not look like jelly

Wind is where good foliage becomes convincing and where most foliage becomes faintly ridiculous. The problem is almost always that everything moves together — one global sine wave, and the forest breathes in unison like a single organism.

The fix is per-instance phase offset. Three layers, each at a different frequency:

Nodes in this graph · Unreal Engine 5.7

1. PerInstanceRandom — a stable 0–1 value per foliage instance — this is what desynchronises the forest
2. Multiply — scale that random into a phase offset in radians
3. Time — scaled three times, at roughly 0.4, 1.6 and 5.0 — trunk sway, branch bend, leaf flutter
4. Sine — one per frequency layer, each with the per-instance phase added
5. Add — sum the three layers, weighted heavily toward the slowest
6. VertexColor — the SpeedTree wind weights — R for trunk, G for branch, B for leaf — masking each layer to the right part of the tree
7. WorldPositionOffset — the final displacement

Weight the layers heavily toward the slowest. My defaults are roughly 0.6 / 0.3 / 0.1 for trunk, branch and leaf. Trees are mostly still; it is the occasional slow lean that reads as wind, not constant fluttering.

Where the frames actually go

The single most useful thing I can give you is not a settings list, it is the order in which to look for problems. From the same test scene, ranked by how much a bad decision costs you:

That ordering is the opposite of most people’s instinct. Triangle count is the thing artists worry about and, with Nanite, the thing that matters least. Shadow casting on grass is the thing nobody thinks about and it can cost several milliseconds.

Five milliseconds for two checkboxes.

What I still have not solved

Two honest gaps, because a guide that claims to have everything solved is usually lying:

• Seasonal variation from one asset. I have a working mask-and-LUT approach that handles green to autumn convincingly. Winter — bare branches with snow load — genuinely needs different geometry, and I have not found a way around that which does not look wrong. That is a later part in this series, when I have something worth publishing rather than a workaround.
• Billboard transition at extreme distance. The pop is reduced but not gone. It is currently the weakest part of my forest and I would rather say so than pretend the cull distance solves it.

Next

Part two takes the masked-to-opaque decision apart properly: exactly when the swap loses, how to measure it on your own scene rather than trusting my forest, and what happens to your memory budget at the scale where it starts to hurt.