SIGGRAPH acceptance turns on evidence proportional to a graphics claim: a technique that claims
to be faster must be timed against a real baseline on stated hardware; one that claims higher
quality must be compared, quantitatively and visually, against the strongest prior method. This
skill matches evaluation to claim shape and pre-empts the domain-expert reviewer's first
objections. Anchor policy to resources/official-source-map.md.
| Claim shape | Evidence the reviewer expects | Common failure |
|---|---|---|
| "Higher quality" | Head-to-head vs SOTA with a metric (PSNR/SSIM/LPIPS/FLIP; Hausdorff/normal error for geometry) + side-by-side visuals + video | Only one's own results shown; no baseline |
| "Faster / real-time" | Wall-clock vs baseline at equal quality, with GPU/CPU, driver, resolution | Timing at unequal quality; no hardware stated |
| "More general / robust" | Results across a broad, non-cherry-picked scene set incl. hard cases | Works only on the paper's three easy inputs |
| "New capability" | Demonstrations prior methods provably cannot produce | Capability asserted, not shown against a method that fails |
| "Perceptually better" | A user/perceptual study with enough participants and a valid protocol | "Looks better" with no study |
In graphics, the head-to-head comparison against the strongest prior method is not optional:
reimplement, say so and match their reported numbers where possible. A weakened baseline is the
objection that sinks the paper.
yourself or the baseline an advantage, disclose it.
numbers and pixels persuade different reviewers.
the region of interest. No single metric is sufficient — report several and show the images.
the alignment and units.
residuals; a plot over time, not a single frame.
unequal sample counts or resolutions without saying so.
Timings are claims a reviewer will check:
for every timing.
When the claim is about perceived quality or usability:
correction for multiple comparisons where relevant).
size and confidence intervals, not just significance.
the part you claim, not an incidental engineering detail.
check** so test scenes are not in training.
siggraph-supplementary).
[Claim -> evidence] each claim matched to comparison/metric/timing/study? yes/no
[Baselines] strongest prior method compared, faithfully, at equal conditions? yes/no
[Metrics] appropriate metrics + visuals/video for each quality claim? yes/no
[Timing] hardware/resolution/settings reported, matched-quality? yes/no
[Ablations] each component isolated; contamination checked (if learned)? yes/no
[Gaps] <ordered, with the reviewer objection each closes>