PAPER-NOTES.md tracks findings that cleared the bar of surprised us + measured + others would hit it. Each entry records what a publication-grade version still needs. Collected here so they're visible as work rather than only as footnotes.
§1 — per-texel relative-difference metrics are biased under jitter. Needs a minimal synthetic reproduction (checkerboard + jitter, no upscaler) demonstrating the bias analytically and numerically, plus a comparison against FSR 3.1.5's own signed-difference pyramid on identical input. Currently the claim rests on measured floors inside our own detector.
§2 — the price of skipping the scatter. Needs an A/B of the repaired cross-frame gather against the true atomic scatter on the same grazing-plane scene. The structural candidate bundle still implements the scatter, so the rig exists. The open question is sharp: does the repaired gather match the scatter's mask exactly, or only its stability?
§3 — source-faithful pass graphs vs fused re-derivations. 168 capture pairs sit ungraded under bench/results/raw/CANDIDATES/. The "no measurable visual win" claim currently rests on our own inspection; blinded grading would make it a result. Cross-device timings also needed (#4).
§6 — EMA temporal AA cannot converge if rectification writes back. Needs a formal fixed-point argument (under what box statistics is the converged mean a fixed point of clip∘blend?) and a comparison against FSR2's actual still-scene behavior on the same scene. The relax sensitivity study is tracked separately in #5.
§4 — temporal guides. This one is now satisfied: the consumer's M6 A/B supplied the measured win the entry was waiting on. Worth folding into the entry.
Each of these is independent — happy to see them picked off one at a time.
PAPER-NOTES.mdtracks findings that cleared the bar of surprised us + measured + others would hit it. Each entry records what a publication-grade version still needs. Collected here so they're visible as work rather than only as footnotes.§1 — per-texel relative-difference metrics are biased under jitter. Needs a minimal synthetic reproduction (checkerboard + jitter, no upscaler) demonstrating the bias analytically and numerically, plus a comparison against FSR 3.1.5's own signed-difference pyramid on identical input. Currently the claim rests on measured floors inside our own detector.
§2 — the price of skipping the scatter. Needs an A/B of the repaired cross-frame gather against the true atomic scatter on the same grazing-plane scene. The structural candidate bundle still implements the scatter, so the rig exists. The open question is sharp: does the repaired gather match the scatter's mask exactly, or only its stability?
§3 — source-faithful pass graphs vs fused re-derivations. 168 capture pairs sit ungraded under
bench/results/raw/CANDIDATES/. The "no measurable visual win" claim currently rests on our own inspection; blinded grading would make it a result. Cross-device timings also needed (#4).§6 — EMA temporal AA cannot converge if rectification writes back. Needs a formal fixed-point argument (under what box statistics is the converged mean a fixed point of clip∘blend?) and a comparison against FSR2's actual still-scene behavior on the same scene. The relax sensitivity study is tracked separately in #5.
§4 — temporal guides. This one is now satisfied: the consumer's M6 A/B supplied the measured win the entry was waiting on. Worth folding into the entry.
Each of these is independent — happy to see them picked off one at a time.