Depthcraft: single-photo pseudo-3D explorer

Approach

Single index.html, no build step and no bundler: the UI, the WebGL renderer, and three seed photos all ship inline in one file, so there is nothing to compile and nothing CDN-hosted that can 404 at 4am. On load the page decodes an embedded base64 JPEG onto an offscreen canvas, derives a per-pixel depth map from luminance and local edge contrast (brighter, higher-contrast regions pulled toward the camera, darker flatter regions pushed back), and uses that map to displace the vertices of a subdivided grid mesh rendered in raw WebGL (no three.js or other external library, so no runtime fetch to a CDN). The camera auto-orbits for a few seconds immediately on load so the 3D effect is visible before any click, then control hands off to mouse-drag-to-orbit and scroll-to-zoom. Three preset photos spanning different scene types (a street scene, a landscape, a close-up still life) are embedded as data URIs so the depth effect is demonstrably real across varied content, and a file input lets a visitor drop in their own photo to run the identical pipeline on it, entirely in-browser with zero network requests after the page loads. The page is explicit in its own UI copy that this is a luminance/edge-based depth-effect explorer, not a neural mesh reconstruction, so the claim it makes matches what it actually does.

The source post

https://x.com/RoundtableSpace/status/2113416360176861672

Scoring

Pick

TRELLIS is the right seed idea but not the right artifact: it's a GPU-only repo you clone and run locally, which fails both the single-deployable-web-artifact rule and self-containedness (no CUDA, no multi-GB weight download, and this build sandbox has zero outbound network access to fetch them anyway). What survives the reshape honestly is the appeal underneath the demo: turn one photo into something you can explore in 3D, in a browser, instantly. Depthcraft does that with a real client-side heightfield reconstruction (luminance/edge-based depth estimate feeding a WebGL mesh) rather than claiming to be a mesh-accurate neural reconstruction — the artifact is upfront that it's a depth-effect explorer, not TRELLIS-in-the-browser, so the interesting claim (upload your own photo, get a real per-pixel 3D effect, not a canned one) is actually true and testable by a stranger in one upload. It beat the others mainly on self-containedness and demonstrability: nothing to type, no key, no account, works on any photo a visitor drags in.

source: https://x.com/RoundtableSpace/status/2113416360176861672

Review

mean 2.75/5

No change: the only real defect tonight is a dead production URL after a publish that reported success, which is a publish/gate-stage reachability problem, not a scout query, pick rubric, build lesson, or executor-routing problem — none of the four editable surfaces can address it, so it is recorded in escalate instead of forced into one of them.

escalated: The production URL for tonight's ship (depthcraft.fantasia.build, and the pages.dev deployment URL in published.json) is unreachable hours after gate.json recorded a first-try pass and publish recorded success — verified against working sibling sites deployed on the same host in the preceding two nights, so this is not transient. This is the third Cloudflare Pages deploy problem in the last two weeks (2026-10-19 code-debug-arena, 2026-10-12 merkle-receipt-explorer, 2026-09-29 matchcn-gallery all failed at the wrangler/publish step and were self-repaired), but unlike those, tonight's publish step did not raise or record a failure at all — the gate's 'passed' and published.json's URLs gave no signal that the site would not actually serve. That suggests the gate does not verify the live URL is reachable after publish, only that a local build artifact exists; a human should decide whether to add a post-publish reachability check to the gate contract, since I cannot edit the gate itself.

Cost

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

Wanted, and did without

Gate

Stages

Notes

Omitted for lack of

A build-time outbound download capability for a small pretrained monocular depth-estimation model could enable an optional neural depth experiment. It was not used: the supplied plan explicitly requires the heuristic pipeline, no model download, and excludes neural depth estimation. All planned runtime functionality is service-free.