Abdal et al. (2023) — ARXIV-2311.17857
Key takeaways
- Puts 3D Gaussians on articulable shells — inflated and deflated copies of a template human body in a canonical pose — with a CNN generating a 3D texture stack whose features become the Gaussians’ attributes
- A speed fix that repays in correctness. Prior 3D GANs use volume representations too slow to render at training resolution, so they render small and upsample in 2D — and 2D upsamplers are multi-view inconsistent by construction. Efficient Gaussian rendering makes native 512×512 affordable, so the upsampler disappears and the inconsistency with it
- The shells being articulable matters twice: during GAN training, and at inference for posing a body arbitrarily
- Generates 3D humans from single-view datasets (SHHQ, DeepFashion), the template prior standing in for multi-view supervision
Standing in the anthology
Read — NOTE-032. Among the furthest documents in this corpus from the anthology’s subject, and nothing is sourced to it.
The observation that is not about graphics: an accumulated workaround has a cost that disappears when its cause does. The 2D upsampler existed only because rendering was slow, and it was responsible for a quality defect nobody attributed to speed. The record’s efficiency practices justify themselves by cost; this is a case where a cost fix repaid in correctness.
The takeaways it carried listed four properties without the connection between them, and the connection is the paper.
Audited by #121; unsourced. Downstream of LIT-108
within the record — it builds on those Gaussian primitives — so its
contribution to SOTA-205 is adoption rather than an independent
result, and ADR-017 puts adoption in consensus_note rather than
source:. R2 (a strong structural prior can substitute for supervision) is the
separable claim and the record holds one instance of it.