One photo is enough for a 3D plant: Plant Orbiter orbits a video diffusion model around the specimen and holds mean height error to 5.9%

A single frontal photograph yields a 3D Gaussian splat of a plant that already sits in the camera frame of the input photo, with no extra calibration or manual placement. Multi-view methods need tens to hundreds of calibrated photographs around the plant, and single-image generators, though they need only one image, return output in their own arbitrary frame where realignment and rescaling are a known challenge. Under input-view placement, Plant Orbiter reaches IoU 0.552 on PlantDreamer and 0.556 on Botany-Bot, above the best compared values of 0.529 and 0.504; on PLANesT it reaches 0.343, slightly below SV3D at 0.348.

The reconstructed 3D shape matches real plant point clouds more closely, and on most datasets that advantage does not depend on ideal alignment. Generative single-image methods improve markedly under oracle placement, with TRELLIS.2 overtaking on some metrics, which indicates their single-image alignment is itself unreliable. Plant Orbiter's 3D F-score is highest on all four groups, at 0.443, 0.540, 0.252 and 0.257 for PlantDreamer, Botany-Bot, pepper and rose; its Chamfer distance is best everywhere except PlantDreamer, where TRELLIS.2 is slightly better at 17.7 versus 18.3.

Prepending a text-to-image model to the same pipeline generates 3D plants from text and preserves the fruit count named in the prompt. General text-to-3D models control thin, overlapping plant structures poorly, and most match a requested fruit count at or below chance. In the tomato and lemon fruit-count experiment, Plant Orbiter reaches 50.0% exact number generation with mean absolute count error 0.60, the best and lowest of all methods; 40 of its 42 assets (95.2%) preserve the fruit count of their input image, against 54.8% for TRELLIS.2 lifting the same images.

The demonstrated scope is a single frontal photo input, a single horizontal orbit video, and the bean, mint, kale, pepper, rose, quinoa and sequoia species in four public multi-view plant datasets, PlantDreamer, Botany-Bot, PLANesT-3D and GrowSplat; height is evaluated on the GrowSplat time series. The authors propose that covering the plant's reverse side or dense structures would require the video diffusion model to take an additional rear-view image, which would still need to overcome plant consistency and the determination of relative camera placement.

The authors note that a monocular approach captures only details visible in the input image, so the reverse side and occluded features must be inferred, making it best suited to measurements such as height and to plants that do not self-occlude heavily; the current orbit covers a single fixed height, and reconstructions often show artifacts when viewed from above or below. Fruit-count errors originate mainly in the text-to-image stage, with tomato images running at most one fruit short and lemon images up to two too many, more often as the requested count grows; on realism, the Inception-feature kernel distance ranks last, which the authors attribute to mint morphology and pot color, while the DINOv2-feature ranking is second.