caption fed to the text modality
The same S3 checkpoint generating the same held-out object from a single image, from four views, and from a text caption alone — full cascade each time (SS → shape → texture), no ground-truth structure, identical seed. Sorted by how much multi-view loses to single-image. Checkpoints: SS 40,000 · shape 40,000 · texture 40,000 cumulative steps of the tri-modal run (I1 0.5 / IM 0.3 / T 0.2).
Reading the numbers. SS-IoU is occupancy overlap with ground truth at 64³. Text has no image to anchor scale or pose, so its IoU is low by construction — judge that column on whether the object is the right thing, not on the number. edade2fc is memorised: IoU is exactly 1.000 from all three modalities including a 57-token text-only cond, and the render is pixel-identical to ground truth, so a content-duplicate of it sits in the training corpus under a different hash. It is excluded from the "clean" means above. Move with ↑↓ or JK.
About the 3D. The I1/IM/T columns can swap their render for the mesh the cascade actually produced — drag to orbit, and with link cameras on all three follow each other so you are comparing the same viewpoint. Meshes load only for the asset on screen (42 meshes, 23 MB total, 0.56 MB each on average), and they are decimated previews: ~25k faces and a 768² baseColor against the ~199k-face, 2048² originals, with the metallic/roughness map folded into scalar factors where it was constant. Silhouette IoU against the undecimated mesh is printed under each viewer; anything below 0.95 is flagged. Ground truth has no mesh here — that column is a render of the reference asset, which is not in this cache, so it stays a render.