miguelemosreverte/caucasus-branch-closeups-leafwood

Branch close-ups, and a semantic-colour label carrier — rung 1 of a staged dataset ladder

26 Gaussian-splat reconstructions built to answer one question, with the pass/fail gates written down before any data existed, and the honest answer reported whether it passed or not.

• 20 natural branch close-ups — 12 summer, 4 winter (bare by construction), 4 autumn
• 6 semantic-colour "label carrier" assets — the same branches recoloured leaf-flat-red / wood-flat-violet before reconstruction, so the label rides into the cloud in f_dc

Total cost: $3.38. $0.13 per asset ($0.08 image + $0.05 reconstruction).

---

The verdict, first

The leaf/wood boundary IS present in a close-range Gaussian-splat reconstruction — a geometry-only classifier orders the seasons correctly with a 33–42 point span where the previous best was 5 points. But it is not present sharply enough to pass a per-splat AUC gate of 0.85 (measured 0.717 ± 0.074), it does NOT survive transfer to whole plants, and the apparently spectacular 0.963 "ceiling" measured on the colour-carrier assets is a property of the flat-colour domain, not of the material — it falls to 0.615 the moment it is applied to a natural cloud.

Every learning curve in this dataset is FLAT. More assets of this kind buy nothing. We are not asking for the next budget rung.

---

Why this dataset exists

Seven prior routes in the parent project tried to recover a leaf-vs-wood label from Gaussian-splat reconstructions of whole plants. All seven failed. The most recent named a cause that is a resolution limit, not an annotator limit:

a 24×24 cell is 30 px and a twig is 1–2 px, so no cell ever contains a leaf beside the branch it hangs on, and the model is never asked the question.

Every asset in the parent corpus is a whole plant reconstructed at 262,144 gaussians, so a 2 mm leaf is built from splats of roughly its own size and cannot carry its own shape. A close-up of one branch, filling the frame at the same gaussian budget, gives each leaf 20–40× the linear resolution. These 20 assets are that intervention, held to the parent corpus' own conventions in every other respect.

---

The KPIs, stated in advance

	gate	measured	verdict
KPI-0 twig structure present in the cloud at all	bare more linear than leafed	see below	corrects the premise
KPI-1 separability at close range	AUC ≥ 0.85, slabness < 0.3, wood connectedness beats a size-matched random control	AUC 0.717 ± 0.074, slabness −0.128, connectedness 0/6 (degenerate)	FAIL
KPI-2 transfer to whole plants	winter whole plant under 30% foliage (number to beat: 79.4%)	20.9% — but summer 19.2%, span −1.7 points	PASSES THE LETTER, FAILS THE POINT
KPI-3 autumn	autumn leaf share clearly above winter, winter→summer span > 20 points	close-ups: winter 8.4%, autumn 56.3%, summer 41.2%, span +32.8 pts	PASS on close-ups, FAIL on whole plants

Colour is forbidden as a feature everywhere a cross-season claim is made. The feature function reads position (cols 0:3) and the splat's own scales (cols 10:13) and nothing else — enforced in code, not promised in prose. Opacity is banned too, and for a measured reason: across 20 specimens median opacity is 0.085 on bare whole plants and 0.192 on leafed ones (AUC 1.000, a perfect season detector).

---

KPI-0 — a correction to the premise, and a confound nobody had caught

A sibling lane reported, from one specimen per season, that a leafed plant is more locally linear than a leafless one at every neighbourhood size, and concluded the shape is not in the cloud at all. We reproduced that table exactly through its own code path (winter 72.9 / 62.6 / 54.2, spring 73.9 / 65.6 / 59.4 at k = 8/16/32) — and then measured the same quantity on four bare and four leafed whole plants:

	k=8	k=16	k=32
whole plants, bare	77.0% ±2.8	68.8% ±6.7	62.2% ±8.6
whole plants, leafed	68.2% ±2.2	55.1% ±0.7	46.2% ±1.0
gap (bare − leafed)	+8.9 pts	+13.7 pts	+16.1 pts (AUC 1.000)

The reported inversion does not replicate. It rested on one spring specimen sitting 1.0 point above one winter specimen, inside the ±6.7 point spread four bare plants actually show.

But that does not rescue the measurement, because the close-ups broke it a different way. Four bare branches — identical by construction — split into two populations 40 points apart, and what separates them is opacity, not botany:

bare close-up	median opacity	k=32 linearity
cu-w03-bare-whippy-straight	0.019	88.5%
cu-w01-bare-fine-twiggy-smooth	0.036	78.2%
cu-w02-bare-gnarled-fissured	0.995	44.3%
cu-w04-bare-opposite-fork	0.975	28.9%

Across all 20 specimens, linearity tracks opacity at Spearman −0.70 to −0.88. On whole plants "bare > leafed" and "low-opacity > high-opacity" both order the specimens at AUC 1.000 — they are perfectly confounded there and cannot be told apart. On close-ups, where opacity comes apart from season (AUC leafed-over-bare = 0.500, exactly chance), linearity stops tracking season (AUC 0.469–0.594) and keeps tracking opacity (AUC 0.667–0.806).

shape_control linearity is an opacity/density detector, not a shape detector. The close-ups are the specimens that de-confound it. Neither the original claim nor its negation should be quoted from whole plants.

---

KPI-1 — separability at close range: FAIL at 0.717

Leave-one-asset-out over the 12 summer close-ups. Training label = green-excess colour proxy (validated: summer 38–59% green, winter 0.5–3.2%, autumn 0.0–1.3% — it correctly refuses to fire on bare wood). Features geometry-only.

Per-asset AUC ranged 0.559 (conifer needle shoot) to 0.812 (leathery evergreen); mean 0.717 ± 0.074. The gate was 0.85.

Gates. Slabness mean −0.128, max +0.038 — comfortably under 0.3, so the mask is not a half-space; this gate passes, unlike the earlier mask that scored 0.825. The connectedness gate came back degenerate on close-ups and we report it as degenerate rather than as a pass: at the cell size plant_cell selects, a single branch is one component and so is any random subset of it (1 piece / 100% for both wood and the random control, 0/6 "beats"). The test discriminates on whole plants and carries no information at branch scale.

Free control. The four winter close-ups, all wood by construction, held out: predicted leaf 1.9%, 27.4%, 3.0%, 17.8%. The two failures are exactly the two high-opacity reconstructions from the KPI-0 table.

---

KPI-2 — transfer to whole plants: passes the letter, fails the point

whole-plant season	predicted foliage
winter (5 specimens)	20.9%
summer (2)	19.2%
autumn (3)	28.0%
spring (1)	32.2%

The gate was "winter under 30%", against a vision annotator's 79.4%, and 20.9% clears it by a wide margin. We do not claim this as a pass. The winter→summer span is −1.7 points: the model calls everything about 20% leaf, so the winter number is low for the same reason every number is low. This is precisely the failure mode the previous lane documented — a 5-point span amplified twentyfold — and reporting 20.9% as a KPI-2 pass would be the most dishonest sentence available here, and arithmetically true.

The boundary does not survive whole-plant resolution. That is the finding, and it points at the reconstruction rather than at any annotator.

---

KPI-3 — autumn: PASS on close-ups, and it is the one real win

Autumn assets were never trained on by anything, and colour was never a feature. This is the only non-circular cross-season number in the study.

	winter	autumn	summer	span
close-ups	8.4%	56.3%	41.2%	+32.8 pts
whole plants	20.9%	28.0%	19.2%	−1.7 pts

At branch scale the ordering is correct and the span is +32.8 points against the previous best of 5.0. Autumn scores above summer, which is what a real leaf detector should do (autumn branches here are densely clothed and not green). At whole-plant scale the same model produces nothing.

---

The colour-carrier experiment: a 0.963 ceiling that is not what it looks like

Six assets were made by asking an image-edit model to recolour an accepted close-up by semantic part — every leaf flat red, every twig flat violet — and reconstructing that. The label then arrives per-splat in f_dc, with no annotator, no back-projection and no cross-view vote.

The label survived cleanly. Ambiguous-hue share 0.0–0.6%; desaturated share 0.9–24.3%. The bare branch — whose mask image contains 0.0% red pixels by measurement — reconstructed to 0.00% leaf splats. The bleed detector passes.

The ceiling, held out by asset, geometry only: AUC 0.963 ± 0.019 (min 0.934, max 0.992; self-fit upper bound 0.980).

Read alone, that says the shape is in the cloud and seven annotators were the problem. It is not alone. The same model, applied to natural close-up clouds it has never seen:

	AUC
mask → mask (same domain, held out by asset)	0.963
mask → natural close-up (cross-domain)	0.615 ± 0.050
natural → natural (trained natively, LOO)	0.717

A flat saturated colour step is a stronger cue than a tonal one, so triposplat reconstructs a red leaf beside a violet twig as two geometrically distinct objects in a way it does not for a green leaf beside brown bark. Coloured assets come back with 0.87× the median splat size of their natural siblings. The 0.963 is a fact about the synthetic domain. A mask-trained model is worse on natural clouds than one trained on the natural clouds directly, and its whole-plant winter control degrades to 38.0% foliage.

An independent blinded read of the masks (no numbers given) found the classes recognisable but the boundary imperfect — "thin RED twig-like spurs appear near the central and right forks; some leaf stalks appear VIOLET" — so the label is clean in the middle of a leaf and noisy at the leaf-to-twig join. Label noise depresses a measured ceiling, so 0.963 is if anything understated within its domain, which only sharpens the point: the domain is the problem.

---

The learning curves — why we are not asking for the next rung

Every point is held out by asset, never by splat, with stratified draws so no draw lacks a winter asset.

Natural close-up route (12 draws per N):

N training assets	KPI-1 AUC	whole-plant winter	whole-plant summer
4	0.710 ± 0.009	19.7%	16.9%
8	0.722 ± 0.017	24.3%	24.0%
12	0.714 ± 0.045	20.5%	19.1%
15	0.739 ± 0.056	20.9%	19.1%

Slope over the last segment: +0.008 AUC per asset. The entire rise from N=4 to N=15 is 0.029 AUC — smaller than the ±0.056 spread at the top point. The curve is flat within its own noise.

Colour-carrier route (10 draws per N):

N	same-domain AUC	cross-domain AUC	whole-plant winter
1	0.949 ± 0.006	0.609 ± 0.028	41.9%
2	0.954 ± 0.008	0.601 ± 0.012	37.4%
3	0.963 ± 0.011	0.607 ± 0.015	41.7%
5	—	0.620	35.8%

Same-domain performance is already at 0.949 with a single training asset. Cross-domain is flat at 0.60–0.62 across a five-fold increase in data. Neither curve is limited by how many assets exist.

A flat curve is a successful rung. It answers the rung's question for the price of the rung. Asking for 50 assets on these curves would be spending budget to avoid admitting a measurement.

---

What we recommend instead

1. The framing intervention worked and is worth keeping. A 32.8-point season span against a 5.0-point previous best is real, and it came from changing what is photographed, not from any new estimator.
2. The remaining limit is the reconstruction, not the label. Both routes converge on the same wall: whole-plant clouds do not carry the boundary at any framing of the classifier, and the only way to make it easy was to change the image in a way that changes the reconstruction.
3. Fix opacity before anything else. A 10× season-locked opacity difference is currently sufficient to fake every whole-plant result in this area, and at least one published claim in the parent project rests on it. Any future work here must force unit opacity or report the opacity confound alongside the number.
4. The genuinely promising route is upstream of all of this: take the label from the generator that produced the plant, not from anything downstream of the cloud.

---

Contents

closeups/
  requests/     40 request JSONs (stage-0 image + stage-3 reconstruction), corpus-convention shape
  source/       20 PNG source images, one per asset
  splats/       20 PLY reconstructions + run.json, spec.json, pricing, hashes per asset
  analysis/     kpi0_shape_control.py, kpi0_opacity_confound.py, features.py, kpis.py + results JSON
  manifest.json what was varied across the 20, and what was held fixed
synth-color/    THE LABEL CARRIERS - synthetic flat-colour masks, NOT appearance assets
  requests/ source/ splats/  6 assets, all prefixed synthcolor-
  gen.py ceiling.py transfer.py audit.py + results JSON
  MANIFEST.json hue→class mapping, recolour prompts, measured ambiguous-hue share per asset

Bucket discipline. synth-color/ assets are synthetic flat-colour label carriers and must never enter a natural appearance pool — a flat-red plant would silently poison any fidelity or uniqueness metric computed over it. synth-color/audit.py re-runs the parent corpus' own file globs against these directories and checks prefixes, symlinks and published listings. It passes: 0 of the 100 PLYs the corpus enumerates fall inside either new bucket.

Generation

All 20 close-ups hold the parent corpus' stage-0 conventions fixed (fal-ai/nano-banana-2, 1K, plain neutral warm-gray seamless studio backdrop, museum-catalog lighting, strict level elevation, long lens) and its stage-3 reconstruction settings fixed (tripo3d/triposplat, num_gaussians: 262144, num_inference_steps: 20, guidance_scale: 3, output_format: ply). The only thing that changes is the subject: one detached branch, both ends in frame, filling it.

Twelve different branch habits were varied across the summer set — ovate/alternate, pinnate-compound, conifer needle shoot, drooping lanceolate, deeply lobed, palmate on long petioles, small round dense, cordate on a zig-zag twig, ribbed/doubly-serrate, sparse leathery evergreen, prickly arching cane, and small opposite decussate pairs. The four autumn assets are habit-matched to four of the summer ones, so autumn differs from summer in leaf colour only — which the classifier is forbidden to see.

Limitations

• Per-splat leaf labels on natural summer assets come from a colour proxy, not ground truth. Colour is never a feature, and the proxy is never applied to autumn, winter or whole plants — the three places it would be circular — but KPI-1's 0.717 is an AUC against a proxy, not against truth.
• 5 two-class colour-carrier assets is a small pool for a ceiling estimate; the curve is flat across it, but flatness over N=1..5 is weaker evidence than flatness over N=1..20 would be.
• Whole-plant evaluation uses 11 specimens from three canopy-tree groups, not the full corpus.
• The connectedness gate is reported as degenerate at branch scale, not as a pass.