Latent-to-4D — Beyond Pixels¶

Development line: project:latent-to-4d · thread beyond-pixels
Last event: 2026-08-13 · 1 dated since 2026-08-13 · Researched: 2026-09-05 · confidence: high

What it is¶

Latent-to-4D is a research method from Beyond Pixels for text/image-to-4D. L4AR takes the final latent from a compatible video DiT and outputs cameras with dynamic world-space point maps. Checkpoints transfer across DiTs only within a shared VAE. Code, weights, and an inference workflow are not out yet. We track and evaluate the result; we cannot deploy it reproducibly today.

Development line¶

• 2026-08-13 — Beyond Pixels project and source repository were recorded. On 2026-08-13, the Latent-to-4D development line recorded links to the Beyond Pixels project page and its source repository. This establishes a public project artifact together with an implementation location.

What changed¶

2026-08-13 — The project page and repository establish Latent-to-4D as a direct latent-to-4D generation method. The source notes that the paper, page, and repository became available on 2026-08-11, while inference code, training code, and pretrained weights remain in preparation. The model trains on about 1K 81-frame clips and claims +2.88–5.81 projection-based DINO-F1 over Wan+4RC. That gain does not prove metric accuracy for 4D geometry.

How to use this¶

From 2026-08-13, treat Beyond Pixels as a Latent-to-4D project reference with both a public project page and source repository. Validate methods and usage directly against those linked artifacts before adopting them.

1. Open the interactive gallery to check whether cameras and dynamic point maps fit the task; demos cover text-, image-, and controlled-4D examples. — https://hayd-zju.github.io/Beyond-Pixels/
2. Do not schedule reproducible runs before inference code and weights release: the repository contains only a README and marks both as in preparation. — https://github.com/hayd-zju/Beyond-Pixels

Best practices¶

• Treat compatibility as limited: verify that the source video DiT shares a VAE with Latent-to-4D before integrating. — https://arxiv.org/abs/2608.10744
• Do not turn motion control, manipulation, or navigation demos into claims of physical success: the authors describe them only as interface compatibility checks. — https://hayd-zju.github.io/Beyond-Pixels/

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• No released inference code, training code, pretrained weights, package instructions, license, hardware requirements, or reproducible runtime measurements were available on the official repository when read.
• The dated addition for 2026-08-13 is retained in what_changed; the separately dated arXiv submission was 2026-08-11.

Sources¶

Agent brief¶

• Subject: project:latent-to-4d, thread beyond-pixels, 1 dated events 2026-08-13 → 2026-08-13.
• Practical note: From 2026-08-13, practitioners should treat Beyond Pixels as a Latent-to-4D project reference with both a public project page and source repository, and validate methods and usage directly against those linked artifacts before adopting it.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.