Papers ยท watchlist
Monitored papers
Recent papers that could improve our 3D reconstruction pipeline
(3d-data-reconstruction โ 3d-viewer). One entry per paper:
what it does, why it matters to us, what its limits are, and โ most importantly โ
what event would change our verdict.
P00x and
never deleted: a discarded paper keeps its entry (with the reason), so we don't re-evaluate
it from scratch six months later. Status vocabulary:
watching (interesting, no action possible yet) ยท
testable (code/weights available, experiment possible) ยท
tested (we ran it on our data โ link the experiment entry) ยท
adopted / discarded (with reason).
A mirror index lives in Obsidian (EC Vault โ Researches โ Monitored Papers - 3D
Reconstruction) โ when a status changes, update both.
Watchlist
| id | paper | date | relevance for the pipeline | status |
|---|---|---|---|---|
| P001 | PixWorld โ unifying 3D scene generation and reconstruction in pixel space | Jul 2026 | feed-forward gaussians from sparse views + generative completion of uncaptured areas | watching (no code/weights yet) |
P001 โ PixWorld ยท added Jul 13, 2026
- Paper
- PixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space โ Gao, Wang, Cao, Yu, Wang, Bian (NTU + AISphere) ยท arXiv 2607.05373
- One line
- A single ~1B diffusion transformer that both reconstructs and generates pixel-aligned 3D gaussians from a handful of views, feed-forward.
- Status
- Watching. No code or weights released as of Jul 13, 2026 โ no repository linked in the paper, none found by search.
What it does
Takes a set of views, splits them into "clean" (conditioning, reconstruction) and "noisy" (generative target), and outputs pixel-aligned 3D gaussians in a single forward pass. The real novelty is where the diffusion lives: not in a VAE latent (the Gen3R route) but directly in pixel space, supervised through the differentiable rendering of the predicted gaussians โ the diffusion objective optimizes the 3D representation itself. Two extra anchors: a geometry perception loss that aligns rendered views with ground truth in the feature space of frozen 3D foundation models (ฯยณ, VGGT), and pseudo-depth supervision from DA3.
Numbers: beats MVSplat/DepthSplat on sparse-view reconstruction (RealEstate10K 4-view: PSNR 26.21) and LVSM on 1โ2-view generation; first overall on WorldScore (71.04), with the top score on camera control (91.08). Trained at 336ร448 on 32 A800s.
Why it matters to us
1 โ The economics of reconstruction. Our current pipeline (LingBot-MAP poses/depth โ per-scene 3DGS optimization) needs dense coverage and ~30 min of GPU per scene. A feed-forward model that emits gaussians from 2โ4 views changes what fraction of a gameplay capture is reconstructable at all.
2 โ The direct answer to E001's wall. E001 established that our 3DGS quality collapses ~20 cm outside the capture corridor, and that it is a limit of the capture, not of the optimization โ no regularizer moves it. The generative half of PixWorld is exactly the missing tool for that class of problem: plausibly completing what the player never looked at, instead of leaving holes.
3 โ Zero integration cost downstream. The output is standard 3D gaussians:
whatever it produces is already renderable in 3d-viewer and the demos.
4 โ The strategic direction. Winning WorldScore's camera control means camera-consistent, navigable scene generation from very few images โ the bridge from "reconstruct what the player captured" to "generate the explorable game world". For a pipeline whose input is gameplay recordings, that is the trajectory that matters.
test-gta-demo-1 frames,
then fine-tune on our captures with LingBot-MAP poses/depth replacing DA3. Independently
of the release, the pattern is worth tracking: pixel-space diffusion supervised
through gaussian rendering + geometry loss on frozen 3D foundation models is where the
feed-forward splatting thread is converging โ expect follow-ups with code.