Lab home
Round-by-round research notebook
The landing page explains the experiment, shows evidence, and keeps the research legible instead of hiding it in scripts.
Visual field guide · 34 screenshots
ForgeScene + Forge Valley
One research system, two real-time engines, and one hard rule: the scene has to survive an impartial visual judge, not just look good from a hero camera.
Three.js construction sitesUnity 6 natural landscape0–100 visual judgeNavigable views count
Two Arms, One Mission
The shared objective is not “generate a nice 3D scene.” It is to rebuild a photoreal source image in a real-time renderer and have an impartial judge score the match. ForgeScene uses construction-site photographs and a Match Score rubric. Forge Valley uses a natural-environment rubric, four camera views, and the weakest navigable view as the real score.
2rendering engines
0–100absolute visual score
4 viewsrequired for Forge Valley
85reference target score on the calibrated rubric
A hero shot is not the deliverable.
If the reconstruction breaks when the camera moves, it is not a real scene. That rule is why alternate views, orbit shots, fly-throughs, and the weakest-view score matter so much.
What The App Actually Contains
ForgeScene is already more than a renderer. The production surface exposes the experiment history, render gallery, interactive arena, semantic segmentation, hypotheses, progress, and the Unity build proof. Click any screenshot to open its full-page capture.
Gallery
Source and render inventory
A visual catalog of scene attempts, source photos, and generated evidence.
Arena
Interactive reconstruction
The browser surface is where a scene proves it can be explored, not merely screenshotted.
Segmentation
Photo decomposition
The source image is broken into meaningful regions such as earth, rock, tree, sky, snow, and water.
Hypotheses
What should improve the next round
Promising ideas and failed approaches remain visible so the system does not repeatedly pay for the same lesson.
Progress
Tournament telemetry
Tracks the source set, generated assets, match activity, and judge-scale context.
Detailed lab
The experiment is inspectable
Long-form evidence turns a black-box generation loop into a reviewable research record.
Unity proof
Build, budget, render, verdict
The page documents the Unity output and review evidence, but its iteration copy is stale and is not the current score authority.
The Three.js Arm
Each source photo becomes a “genome”: camera, environment, textures, and placed objects. The renderer captures alternate views. The judge compares them against the source on composition, inventory, lighting, texture realism, and depth. The orchestrator keeps what scores and rejects what does not.
View 01
Front
Closest to the source composition. Necessary, but never sufficient.
View 02
Right
Tests object count, parallax, projection seams, and whether the world survives rotation.
View 03
Top
Exposes flat cards, repeated textures, floating geometry, and fake depth immediately.
Photo Decomposition
Semantic segmentation converts one flat image into a working inventory. Regions become masks, cutouts, texture candidates, depth hints, and scene-placement evidence.
Input
Source
The untouched image.
Interpretation
Overlay
Class boundaries reveal the scene inventory.
Machine surface
Class map
The structured map feeds downstream reconstruction work.
Projection test
Depth from a single image
A useful experiment, but edge stretching and non-parallax backdrops remain visible failure modes.
Alternative pipeline
Tracer render
Another reconstruction path used to test whether different geometry and texture strategies score better.
The Unity Arm: Forge Valley
A Unity 6 High Sierra landscape assembled entirely through code, rendered headless, and validated before the command succeeds. The build includes terrain, erosion, creek, terrain layers, vegetation, production models, sky, atmosphere, lighting, camera framing, four captures, and hard budget checks.
Navigation view
Lateral
The historic weak view. It reveals silhouette repetition, terrain tiling, and shallow atmosphere.
Navigation view
Opposing oblique
Tests whether the ecology, creek, and depth work from the other side of the valley.
Navigation view
Low fly-through
Exposes foreground material scale, shadow crush, water width, and sparse detail.
Asset + Material Experiments
The landscape is procedural, but the visual investigation still uses authored/scanned assets, generated source images, cutout studies, and model conversion. These tests are evidence about silhouette, coverage, tiling, and whether assets survive multiple viewpoints.
Selection test
Pine candidates
Candidate imagery for replacing repetitive single-family vegetation.
Cutout test
Alpha extraction
Tests whether branches and needles retain enough canopy mass after isolation.
3D conversion
Foreground cluster
Model evidence for high-value near-camera vegetation.
3D conversion
Broad crown
A silhouette family intended to break the monoculture.
Atmosphere
Panoramic sky
The environment map used to establish Sierra scale and distant depth.
Terrain source
Granite scree
Generated source imagery feeding a PBR terrain layer.
Terrain source
Pine duff
Ground material intended to tie forest cover to believable surface ecology.
Production asset
Granite cluster
More credible silhouette and surface variation than stock geometry.
Ecology
Manzanita
A needed understory family for ecological variety.
Detail
Deadwood
Near-field breakup that also signals age, disturbance, and habitat.
What “Tests” Means Here
This app does not currently have a conventional automated unit-test suite. Its strongest tests are executable evidence loops: compile, build, render, inspect budgets, capture alternate views, measure image properties, append the ledger, and then run the calibrated visual judge. The distinction matters: a green build proves the pipeline ran; it does not prove the image is good.
1 · Build
TypeScript typecheck + Vite build for the web arm; Unity batch-mode scene generation for Forge Valley.
2 · Validate
Reject missing assets, invalid captures, triangle overruns, broken navigation views, and failed render output.
3 · Measure
Record triangles, stems, ground detail, dynamic range, highlight band, temperature separation, and tiling peaks.
4 · Judge
Score the render with the fixed visual rubric. For Forge Valley, the weakest of four views controls the outcome.
Metrics detect problems; they are not the goal.
Forge Valley has twice improved every recorded metric while making the image visibly worse. The image still has to pass human review and the calibrated visual judge.
| Evidence | What it proves | What it does not prove |
|---|---|---|
| Successful build | The code compiled and the pipeline completed. | Photorealism or navigability. |
| Metric improvement | A measurable image property moved. | The image looks better overall. |
| Hero score | One composition improved. | The world holds up when explored. |
| Four-view judge | The scene survived multiple cameras. | A production 90; the rubric itself caps normal real-time work near 85. |
Current Honest Status
The closed loop works and the project has a real body of evidence. The scene is not photoreal yet. The last completed judge run is iteration 83: 37 hero and 30 weakest navigable view. Iterations 91–95 built and measured but were never judged, so they cannot be claimed as improvements.
37last verified hero score
30last verified weakest-view score
85reference target score
91–95built, measured, not judged
| Priority | Known visual blocker | Why it matters |
|---|---|---|
| 1 | Creek reads as a painted ribbon | Water has scored 1–2/10 and the authored path cuts through terrain instead of following hydrology. |
| 2 | Foreground cobble normal map is at the wrong scale | It makes the ground read synthetic immediately. |
| 3 | Vegetation is effectively a monoculture | One repeated silhouette depresses ecology, production-model, and detail scores together. |
| 4 | Shadow crush in navigable views | The fly-through holds roughly 12% near-black pixels against 1.7% in the reference. |
| 5 | Erosion leaves no depositional record | Channels exist, but alluvial fans and sediment behavior do not. |
The biggest lesson was procedural, not artistic.
“Cannot leave the Built-in Render Pipeline” had been inherited through handoffs as fact. URP was already installed locally. Rechecking that assumption reopened roughly twenty points of ceiling that had been blocked for about thirty iterations.