Expanded review library — 2026-10-12

Scope

Source design
Completed source library: 100 photographic backgrounds + 100 independently generated counterparts, and 50 photographic + 50 generated objects. The current catalog is the count authority. Photo / GPT pairing matches scene or object category, not pixels, exact geometry, illumination, difficulty or human visibility. Pure-text generation prompts and native originals are retained; the tool does not expose the exact image-model version. Generated does not mean photographed. Composition, movement and event manipulation are programmatic in both source groups.

Photographic backgrounds use licensed Poly Haven sources and retained project inputs. Full panoramas are projected to perspective using recorded yaw, pitch and field of view; no 16:9-to-square stretching. Near-overhead placement views are derived from the same panorama, share its source group and do not increase the source count. Generated background inputs are independently generated images, not panels cut out of contact sheets. Reused source files are identified explicitly.

Object photographs retain the downloaded publisher input and original publication URL. A publisher-provided derivative is labelled with its actual dimensions; it is not described as the full original. No small source is upscaled and called 2K. The prepared RGBA is obtained by source crop, foreground mask and alpha-bounds crop, without resizing source RGB. Source images, review canvas, effective alpha area and model processor inputs are different quantities. Processor inputs for this version remain unmeasured.

Task suitability
Current shared-background protocol (shared_background_v1):
All five tasks reuse the shared catalog of 100 photographic + 100 generated background source identities; active pages use the explicit pools below.
T1: multiplicative page 50+50, convolution page 100+100; no separate object pool.
T2: backgrounds 100+100; objects 49+49.
T3: backgrounds 100+100; objects 17+17.
T4: backgrounds 100+100; objects 38+38.
T5: backgrounds 100+100; objects 24+24.
These are available pools, not an instruction to generate their Cartesian product. Twelve paired scene categories (24 background images) per task have curated background/object presets; manual independent selections remain available.
Full source scene is now the default. Previously automatic near-overhead placement projections are explicit optional conditions for sources with retained views. T1/T4 square canvases and T2/T3/T5 3:2 canvases still have recorded different crops; shared source identity does not imply identical rendered pixels.
The default object pool uses the intersection of each photo/GPT pair's individual eligibility, so both provenance groups have equal numbers of the same categories. Individual candidate_tasks remain in the manifest; unpaired eligibility is not evidence that a source itself failed.
T1: all accepted backgrounds can be treated as a controlled image foreground inside a stationary aperture. The new RGB-convolution branch and older multiplicative-stripe branch remain separately named.
T2: use complete shared scenes with controlled foreground composites; local brightness increment is applied in linear RGB and reported as actual digital Y, including baseline and event timing. Ask which side brightened. A glass cutout whose transmitted background was removed is retained for source review but excluded from controlled tasks.
T3: use designated visible straight shafts or edges, with source-coordinate axis annotations. Apples and balls are excluded from angular discrimination because rotating a nearly symmetric appearance does not define a robust physical orientation. Chairs can support directional masking; fine tilt requires a separately designated edge. The question remains which side has the vertical axis. No reference-orientation JND is inferred from this accuracy task.
T4: use large complete objects with identities fixed within a clip. Available candidate identities and identities actually used among N objects are separate counts. Identity selection uses a seed independent of LEFT/RIGHT, and candidate sets are sampled rather than always taking the first N. Classic controls remain dots, with matched center tracks and alpha area. Object silhouettes are an optional additional control. N = 5, 10, 15, 20; achievable coherence = signal_count/N; zero coherence is listed separately. The 48 px fitting-box review excludes objects whose opaque core short side is <12 px; this is an engineering visibility screen, not a measured human threshold.
T5: use L squares, hammers, side-view chairs and other inspected asymmetric shapes. All four quarter-turn appearances are reviewed, with the target shown separately. Keep four spatial positions, two events, real masks and source-dependent effective object size. The task asks whether the specified target appeared at either event. Historical second-target hit rates do not become results for this readout.

Current generation / review units
Source library: source image or source object identity. A crop, a new phase, a trajectory seed and a parameter variant are not a new source.
T1 multiplicative branch: 50 photographic + 50 GPT sources. The original ten pairs and their 40 encoded videos are retained unchanged; forty additional semantic pairs reuse shared-library originals. Non-square masters are center-square cropped and isotropically resized to 512; the exact crop, source/master/input dimensions and SHA remain in sources.json. No new generation or encoding is implied by this expansion. Its 6 spatial frequencies × 5 stripe depths × 3 temporal frequencies × 2 directions × 2 phases × 3 durations are a possible interaction grid, not a requirement to encode every source-condition cell.
T1 RGB convolution: 3 review frequencies (4, 16, 32 cycles per 384 px reference aperture), 3 contrast gains (.25, .5, 1), two constructions and two directions. The library interaction uses prepared atlases; manifest.json identifies actually encoded short videos separately. 512×512 encoded canvas, 60 fps, 2 s review duration, 24 px/s foreground translation. The 2 s duration is a review choice, not a newly accepted formal human protocol.
Earlier placement review: 12 classic/scene pairs, 24 videos and 1,728 frames remain in review_manifest.json. Frozen renderer/configuration are review/render.js and review/build.py.
New shared-context review: 20 classic/scene pairs referencing 20 natural + 18 unique classic lossless RGB videos (38 unique files); 2,880 decoded frame references / 2,736 unique-video frames, in review_shared_manifest.json. T2 6 pairs, T3 6, T5 8. These retain backgrounds/objects across paired flips, with tables, streets, workshops, living rooms, indoor activity rooms and coast. Additional combinations are rendered interactively; no full-grid generation or model calls.

Convolution and multiplication are different manipulations
Multiplication: each linear RGB channel is multiplied by a moving positive sinusoidal gain field while the source scene remains stationary. Stripe depth is not whole-image Michelson or RMS contrast.
Convolution: F_c = N_c * K_f for each linear RGB channel c. K_f is a zero-DC Gabor kernel with recorded bandwidth, orientation, spatial frequency and finite support. A pure-bandpass-plus-DC branch and a low-pass-color-plus-bandpass branch are both retained. The latter preserves coarse color/appearance but also contains extra low-frequency motion cues. It is not called a single-frequency natural image. The processed foreground is translated inside a stationary aperture. Actual output means, RMS, spectrum, clipping and motion are stored with the processing parameters. No per-frame normalization is used. See ../t1_convolution_review/protocol.json for the exact reconstruction and fixed-gain rules.

Formal experiment recommendation — not launched
First lock source groups, then sample conditions. Keep a development set distinct from a later, unseen source group; all sources shown here are review/development material. Photo/GPT status is an analysis factor. Split shared capture/location groups together. Do not claim a crop-based split is unseen scenery.
Six log-spaced frequencies and five contrast levels are a reasonable coarse scan, not a precise threshold estimate. Add levels around a measured transition if the initial curve brackets floor and ceiling; do not densify empty extremes merely to create more points. More backgrounds support transfer estimates; repeated phases support nuisance robustness. They answer different questions.
A finite starting allocation for T1 is 40 sources, balanced 20 photo / 20 GPT and across scene families, at 6 frequencies × 5 contrasts × 2 directions × 2 phases: 4,800 natural sequences for one fixed temporal protocol. That is a proposed upper grid, not a mandated run; a smaller pilot should first verify sampling and response validity. Classic unique inputs must be counted from the actual source-dependent calibration and file/content SHA. A calibrated classic reused by multiple natural scenes is one-to-many, never duplicated to inflate independent n.
For T2–T5 select a balanced incomplete background–object design rather than 200×100 exhaustive composites. Let B_t and O_t denote task-eligible source pools, C_t the selected compatible combinations, L_t the tested condition cells, A_t balanced answer / nuisance cells and R_t seeds. N_natural = C_t × L_t × A_t × R_t; N_classic is the number of unique calibrated classic inputs after SHA deduplication; calls per model = N_natural + N_classic for separate-domain evaluation. Report every factor and actual pair mapping before a formal run. Confidence intervals should resample source groups, not treat correlated phases from one photo as independent people/scenes. With n independent binary cases, the worst-case standard error is 0.5/sqrt(n); the familiar n≈384 for ±5 percentage points ignores clustering and is not a source-count justification here.

Human comparison and scientific claim
Preserve original task, exposure, viewing geometry, units and response definition. Cycles/aperture and digital Y do not become cycles/degree or cd/m² without calibration. Ramirez-style direction accuracy is a candidate T1 anchor; historical contrast sensitivity, thresholds and JNDs use adjacent figure panels unless a compatible readout is actually implemented. Showing every frame to a model is not equivalent to a human 250 ms exposure. Current historical human curves remain separate from these untested scene stimuli.
Candidate claim to test: controlled classical measurements predict direction/event judgments on held-out natural source groups beyond trivial answer bias and sampled-pixel signal baselines; adding background/object factors may improve predictions. The claim is supported only by held-out predictive improvement with source-group uncertainty. It is weakened if simple pixel or sampling baselines explain the curves, or if classical measures do not predict unseen sources. Classical transfer working well is also an informative result. More tasks, more photos and parameter curves alone are not a novelty claim.
A later world-model evaluator check can use objective event labels from programs, tracks or masks, comparing VLM judgments on real and generated clips at matched perceptual conditions. Keep evaluator error distinct from generator error. Do not use the evaluated VLM as its own truth source; do not equate VBench quality subscores with physical correctness. No new generation queue, training or robot deployment was started for this review.

Recent dataset review — reviewed candidates, not silently adopted sources
HouseCat6D (CVPR 2024): 194 real objects / 10 categories / 41 scenes, RGB and instance masks. Most relevant next source for real tabletop/background-object validation. https://sites.google.com/view/housecat6d/dataset ; https://github.com/Junggy/HouseCat6D
HOT3D (2024 data release, CVPR 2025): 33 objects, egocentric interactions, Aria RGB plus masks; hand occlusion and lens geometry need separate handling. Better for later realistic validation than clean isolated cutouts. https://facebookresearch.github.io/hot3d/
TRADE (2026): 17 transparent/deformable objects; a difficult-material stress set, not a replacement for broad source coverage. https://researchdata.tuwien.ac.at/records/j72kx-6fw97
IRIS (2026): real industrial-part photographs and masks alongside CAD/GenAI material; inspect source type before adopting. https://huggingface.co/datasets/Carraskito/IRIS
OmniObject3D (CVPR 2023), Google Scanned Objects and YCB are useful older resources, but rendered views of scans must not be labelled captured photographs. None of these repositories was bulk-downloaded into the accepted current library simply to raise the count.

Reproduction / maintenance
prepare.py is the source-discovery, retention and catalog entrypoint; its selection JSON files preserve the accepted metadata. review.py exports a bounded T2–T5 review set using the shared scene renderer; --only allows a selected case to be rebuilt. T1 convolution processing remains in ../t1_convolution_review/render.py. Use the existing qy environment. Keep originals, prompts, source licenses, source groups, selected-source records, renderer/config SHA and finite review outputs. Candidate contact sheets and decoded-frame temporary files are disposable after QA; do not discard native inputs or expensive original experimental outputs.

Reviewed RGBA files and masks are retained canonical inputs for exact stimulus reproduction. Some photographic masks include recorded manual alpha repairs; rerunning IS-Net alone does not reproduce those repairs. prepare.py validates all retained raw/output hashes before retention and refuses to overwrite changed, missing or differently sourced reviewed inputs.

Shared-background counterbalancing
T2: each scene/object/strength cell includes LEFT and RIGHT increments while holding the baseline and mask/ROI fixed. Optional background-only horizontal mirror is crossed independently with the label. Local background Y mean and RMS are recorded, not silently equalized.
T3: each scene/object/angle cell crosses LEFT/RIGHT vertical side and tilt sign ±, holding source, scale, background and exposure fixed. Background-only mirror changes neither foreground orientation nor the answer. The matched control copies one central patch into two fields and is explicitly distinct from full-scene viewing.
T5 shared schedule: foil_variant 0/1/2 selects (90,180)/(180,270)/(270,90). PRESENT_FIRST and PRESENT_SECOND compare with ABSENT_A by changing only the designated event to 0 degrees. Other event, positions, mask pixels, duration and SOA stay fixed. ABSENT_B uses the alternate foil order; the one-event toggle is disabled for that auxiliary schedule. First LEFT/RIGHT and second TOP/BOTTOM locations, foil variant and mask seed are independently selectable. Outlined location guides retain the four-box/cross layout on bright scenes. Earlier linked-seed/schedule remains available through the earlier-source pool.
Formal allocation: choose a balanced incomplete crossing of compatible source backgrounds and objects. Within every selected combination, balance answer, position and nuisance controls; perform source-group splits before deriving conditions. Report background-context strata and side-specific outcomes. More independent scenes improve coverage; they do not prove that differing local clutter is harmless. Mirror/crop/phase/seed variants count as repeated conditions of the same source.

2026-10-13 current minimal review
The display precedes independent background/object source selectors, grouped by photographic/GPT provenance. All selects the eligible source group; Half selects alternating stable catalog entries. Every selected combination is represented, with viewport-limited rendering buffers. T4 partitions selected object identities into mixed sets of up to N; short sets repeat identities to fill N without changing identity mid-clip. This UI grid is not a committed statistical sample or encoded-video count. T5 presents only canonical target orientation as reference. One earlier web/source snapshot is web_before_minimal_20261013.zip.
T1 primary display: fixed original-color source, moving RGB multiplicative carrier; 50 Photo and 50 GPT sources, independent source columns. Same-source fixed-aperture v3 convolution review uses the existing RGB filter atlases but selects the same source as stationary exterior and translated filtered interior. The aperture stays fixed, so no moving alpha silhouette supplies a broadband shortcut. The independent-object v2 is retained in foreground.py/foreground/manifest.json (8 clips); old independent-background fixed-aperture v1 remains in render.py/protocol.json (42 clips). Neither old encoded set is relabelled as v3. The color+bandpass adaptation retains low frequencies; fixed gain across SF does not equalize output RMS.
Relevant method precedents: Perfetto et al.2020, Effects of Spatial Frequency Filtering Choices on the Perception of Filtered Images (https://doi.org/10.3390/vision4020029), grayscale natural images with Butterworth filtering and explicit contrast-normalization choices; not a color-motion protocol. Wallis/Dorr/Bex2015, Sensitivity to gaze-contingent contrast increments in naturalistic movies (https://doi.org/10.1167/15.8.3), modifies selected Laplacian frequency-band contrast and reconstructs original natural movies; not a fixed-background left/right task. These support distinguishing band isolation from structure-preserving reconstruction, not an assertion of exact replication.
Display defaults: T2 single scene_gain (default0.9), separately selectable, not physical luminance; T3 tilt16 degrees; T4 N15 with15 distinct source identities. Existing URL values override defaults. Earlier classic panels reuse7Oct encoded media with their own parameter labels and independent controls. Current matched controls and earlier classical samples are distinct protocols.
T4 review adds 32 px/s alongside historical16. New live UI defaults32; missing speed in old calls still defaults16 and existing videos are unchanged. Twelve numerical renderer checks passed: N=5/10/15/20 at16 and32, plus N20 at32 with zero/full coherence and seeds0/1. Checks covered equal signal/noise speed magnitude, shared paired starts and bounds/no overlap across endpoint envelopes. These bounded cases do not certify all 200 backgrounds or every arbitrary object set. Pilly and Seitz2009 Exp1 used12 deg/s at85Hz; Exp2 used4 and12 deg/s. Screen calibration is required to relate these to pixels/s. Author PDF: https://faculty.ucr.edu/~aseitz/pubs/Pilly_Seitz09.pdf .
T5 static interface diagnosis is separately frozen in ../t5_static_controls_20261013; 32 calls and two matched visual inputs are recorded there and summarized in ../../README.md. This is not a new human curve or a full benchmark evaluation.

Current-turn file accounting: added 22 persistent files: T1 foreground.py plus foreground/manifest.json and8 encoded examples (10), T5 static-control inputs/results/runtime (11), and one prior-web snapshot zip (1). Existing preview.png was replaced with the reviewed minimal T5 screenshot; its predecessor is inside the same snapshot. No new dependency or environment. T1 now has304 owned files, exceeding the normal100-file guideline because retained review assets comprise204 atlases,42 historical videos,42 historical keyframes and8 new videos; code/config/manifest/frontend account for8. No20,000-combination video grid was written. These existing assets remain used by current or explicitly retained historical review branches; broader consolidation is outside this UI/diagnosis revision.


V3 bounded formula QA (2026-10-13): photo_bg_001/gpt_bg_001 × two branches × SF4/16/32, frames0/30/119 only. All12 conditions: stationary exterior identical, measured interior displacement0.5s=12px, admissible gamut. Prepared core first-frame linear-Y RMS for photograph bandpass: .010799/.013341/.010778; color+bandpass: .059286/.058636/.055984. Generated bandpass: .009087/.009076/.009165; color+bandpass: .033813/.033471/.031943. These quantify the v3 CPU construction with Lanczos source preparation; browser background scaling is separately described and not claimed byte-identical. No model inference or new encoded batch.