Source
Abdal et al. (2023), LIT-113 — ARXIV-2311.17857.
The rule
When a system is too slow to run at full resolution and the workaround is to run it small and upsample, check what the upsampler costs in correctness before treating the speed problem as the problem.
In the case this comes from, the cost is categorical rather than incremental: a 2D upsampler is multi-view inconsistent by construction. It works on one rendered view at a time and has no way to know what the neighbouring view will contain. No amount of training fixes that, because it is a property of where the operator sits.
So the fix is not a better upsampler. Making the native path affordable — here, Gaussians on articulable shells instead of volume rendering — removes the defect by removing the stage that caused it.
The rule comes from LIT-113, Gaussian Shell Maps. Prior 3D GANs used volume representations too slow to render at training resolution, so they rendered small and upsampled in 2D; the paper samples 3D Gaussians on articulable shells around a template body, renders natively at 512x512 with no 2D upsampler, and gets multi-view consistency because the upsampler is gone. The paper presents the change as an efficiency gain; reading it as a correctness fix is this record’s framing of its result.
Why this is a rule and not an anecdote
The shape recurs whenever a resolution or precision shortcut is introduced to buy throughput:
- The system cannot afford the honest path
- A cheaper stage is inserted
- That stage has a defect the honest path does not have
- Effort goes into mitigating the defect
The recommendation is to ask at step 4 whether step 1 is still true. Hardware and representation both move; a shortcut adopted when the honest path cost 10x is worth re-examining when it costs 1.2x.
Conditions
- The defect must be structural, not statistical. This is the part that makes the rule usable: “the upsampler is multi-view inconsistent by construction” is a statement about what it can see. A stage that is merely worse on average does not qualify — that is an ordinary quality trade
- The native path has to actually become affordable. The paper earns this with a representation change, not with patience
- One paper, one setting. See
consensus_note. The 512x512 figure is 3D human generation from single-view data
Where else the record has this shape without naming it
SOTA-249 is the mirror image and worth reading beside this one: there, recomputation is the right answer because the trade buys memory the system genuinely does not have. The distinction is step 3 — recomputation has no correctness defect, and a 2D upsampler does.