The Science · RooferX.Club
Grown vs. Generated
Two ways a computer can make a 3D thing. One learned from millions of examples. One grows from a seed and a few laws. We use both, and here's how each one really computes.
"NO TRIPO NO NOTHING. LAWS."Thomas Fristed, the day the Reed grew its first colony (09-2026)
The short version
A recipe vs. a very well-read guesser
The Reed: grown
A recipe. You give it a seed and a few dials, and it follows laws you can read, one step at a time. Same seed, same plant, every time, and every part of the plant can tell you which law made it.
Tripo: generated
A very well-read guesser. It studied about 2 million 3D objects, so when you show it one picture it makes its best guess at the whole shape, including the back it never saw.
On a phone, turn it sideways to read the drawing: the same two rows, top grown, bottom generated.
How the Reed computes
Grown by written laws
- One seed, one plant. The same seed always grows the same plant.
- Laws decide the shape, not guesses. A law can even refuse to grow something that breaks it.
- Nothing is trained. It runs on an ordinary computer. Change a dial and it regrows.
- How the Reed builds its meshes stays in the shop. It is Thomas Fristed's own work.
How Tripo computes
From Tripo's own published research
- It learned first. TripoSG was trained on about 2 million cleaned-up 3D objects, filtered from the Objaverse-XL collection, for about 3 weeks on 160 A100 GPUs.2
- A vision model reads your picture and turns it into features the 3D model understands.1
- A "rectified flow" transformer (about 1.5 billion parameters, scaled to about 4 billion) starts from noise and walks it, step by step, toward a shape that fits those features.2
- The shape lives as a distance field: for every point in space, how far it is from the surface. A mesh is then pulled out of that field.2 Newer work handles open, thin things like plants and cloth up to 1024³ detail.3
- Once trained, it's fast. The earlier TripoSR made a mesh from one image in under half a second.1
Side by side
How they compute against each other
| The Reed (grown) | Tripo (generated) | |
|---|---|---|
| What goes in | A seed + a few dials | A picture or a sentence |
| What's inside | Laws you can read in the code | Billions of learned numbers |
| Learning needed | None | ~3 weeks × 160 A100 GPUs, ~2M objects2 |
| Runs on | Any ordinary computer | GPUs in Tripo's cloud |
| Same input twice | The exact same plant, every time | A new guess unless the seed is locked |
| Can it say why? | Yes: every part traces to a law | No: it's a learned guess |
| The side you can't see | Built by the same laws as the front | Imagined from what similar objects looked like |
| Mesh | One clean mesh by construction | Dense and even; we retopologize it |
| Best at | Living, repeating, lawful things: reeds, groves, stone courses | One-off hero shapes from a picture: a character, a prop |
| Fails by | Refusing (a law says no) | Guessing wrong and looking sure |
| Who owns it | Us: our laws, our code | Paid plans grant commercial rights; our finishing work is the human authorship. Training data carries many licenses4 |
Why we use both
Generate the starting block. Grow the world.
Tripo for the one-offs
A character from a T-pose sheet, a hat, a sword. Fast, and it gets you a real starting block from one picture.
Human hands finish it
Segment, retopologize, texture. That finishing makes the model game-ready, and it's the human authorship that makes it ours.
The Reed for everything alive
Groves, riverbanks, walls course by course. Things that repeat, age, and must obey laws should be grown, not guessed.
Receipts on both
Every model ships with where it came from: the seed and dials for a grown one, the source picture and job for a generated one.
What we'll measure next (empty until measured)
Our own numbers, on our own machine
| Test | Reed | Tripo |
|---|---|---|
| Time to make one plant / one model | EMPTY: time a grow on this PC | EMPTY: time one Tripo job |
| Triangles out | EMPTY | EMPTY |
| Same input, 10 runs: how different? | EMPTY (expected: identical) | EMPTY |
| Hours of human finishing to game-ready | EMPTY | EMPTY |
We don't post a number we didn't measure. Testers: help us fill these in.
Sources
- Tochilkin et al., TripoSR: Fast 3D Object Reconstruction from a Single Image, arXiv:2403.02151 (2024). Mesh from one image in under 0.5 s; MIT license; built on LRM.
- Li et al., TripoSG: High-Fidelity 3D Shape Synthesis using Large-Scale Rectified Flow Models, arXiv:2502.06608 (2025). ~1.5B → ~4B parameters; SDF VAE; 2M objects; ~3 weeks on 160 A100 GPUs.
- TripoSF / SparseFlex, arXiv:2503.21732 (2025). Sparse voxels near the surface; up to 1024³; open surfaces like plants and cloth.
- Objaverse-XL dataset card: the dataset as a whole is ODC-By v1.0; "individual objects… are licensed under different licenses."
- Prusinkiewicz & Lindenmayer, The Algorithmic Beauty of Plants (Springer, 1990). The science of growing plants from rules (L-systems, 1968), the family the Reed belongs to.
- The Reed: Thomas Fristed's own grower. Its code stays on his machine and is not published.