Aether R&D
isolated prototype environment
Diffusion engine / browser spatial workflow
Runnable research artifacts for character, scene, camera, keyframe, and generated-world workflows. These are evidence—not product features—and every known failure is retained on purpose.
01 / End-to-end workflow
Four FAL-generated rigged costume assets, a genuine metric workshop proxy, protected costume swaps, and constrained AI static-pose revisions with exact browser fine-tuning. No diffusion or video yet.
Generated characters, outfits, props, and environment assets instantiated in PlayCanvas with exact camera, lens, pose, attachment, light, and blocking controls.
Depth lift, projection geometry, camera envelopes, and reconstruction experiments. Browser-local controls work here; GPU and provider submissions remain local-only.
02 / Generated worlds
Provider-generated layered geometry with genuine parallax and an opt-in 542 MB dense room. It is an appearance proxy, not a metric set.
Two generated videos trained into forced splats with synthetic camera paths. The scene breaks, but the artifacts remain available so you can judge the failure directly.
The direction-aware four-view request and SPZ-plus-collider viewer are implemented. The retained receipt honestly shows the missing credits instead of fabricating a result.
03 / Source consistency
One 3072×3072 generation containing four explicit cardinal views. Visual identity holds much better, although rigid roof topology still fails.
Raw ReconViaGen GLB beside a deliberately authored 2.5D projection room, with camera controls that expose where image projection stops being spatial geometry.
Published evidence boundary
This site is isolated from the customer application. Checkpoint 006 alone can spend through two bounded server-side FAL verbs; every other generation control remains read-only. Large immutable Hunyuan artifacts and append-only attempt attribution use a same-origin private R2 binding.