EV-JEPA: Elastic V-JEPA for Any Device and Cross-Device Inference

简体中文

Deferred because of cost; not reproduced or validated. This is a research reference. The review logged the deferral without downloading models or data, training, registering an executable adapter, or integrating Evolve. Paper figures below are not local measurements.

Paper information

Field	Value
Paper	arXiv 2610.14324; pinned v1 full text
Organizations	Meta Reality Labs Research / FAIR at Meta
First author's affiliation	Meta Reality Labs Research
First public date	2026-10-09 (arXiv v1)
Authors' open-source code	Not found: v1 did not link official implementation for this method as checked on 2026-10-14. This is not a claim that no code exists anywhere.
Adapter	ev-jepa (planned, not implemented or registered)
Local code	No algorithm implementation. Deferred specification is a plan, not executable code.
Validation status	Not implemented or GPU-validated. Not a capability comparison or executable Evolve operator.

Original paper

Background and contribution

A fixed-capacity video model cannot readily accommodate both edge-device budgets and collaboration across devices. Separately trained widths also need not have compatible internal representations. EV-JEPA shares parameters across nested widths, jointly training narrow, wide, and narrow-to-wide paths at every step. A frozen video teacher constrains their common representation space, and adjustable token compression supports transfer.

Architecture and data flow

flowchart TD
  V[Video and spatiotemporal masks] --> N[Narrow path]
  V --> H[Narrow-to-wide hybrid]
  V --> W[Wide path]
  F[Full video] --> T[Frozen V-JEPA2 teacher]
  N --> L[Three-path masked-token L1]
  H --> L
  W --> L
  T --> L
  H --> C[Optional elastic token compression and cross-device transfer]

This is a local redraw; see original Figure 1.

Key figure from the paper

On the left, nested width slices share weights. On the right, narrow, wide, and narrow-to-wide paths are trained together under the same frozen video teacher. This is the paper's T3 architecture, not a local model.

Source: Figure 2 in v1, S3.F2. figure-source.json records the original and local PNG hashes, dimensions, and processing. Copyright belongs to the authors; the local PNG conversion preserves aspect ratio and is neither a placeholder, AI redraw, nor a local measurement.

Core equations

Sections 3.3–3.4 jointly optimize the same width pair \((w_s,w_l)\) and split layer \(k\):

\[ \mathcal L(c)=|M|^{-1}\sum_{i\in M}\|z_{S,i}^{(c)}-T(X)_i\|_1,\qquad \mathcal L_{\rm T3}=\mathcal L(w_s)+\mathcal L(w_s,w_l,k)+\mathcal L(w_l). \]

Shared parameters are sliced by channel. A learned projection increases width at layer \(k\) on the hybrid path. The teacher is frozen, not EMA. Compressor post-training adds a reconstruction loss at the split layer; ordinary pruning is not a substitute for training compatibility.

Paper results

On SSv2, K400, and EK100, the elastic model was within about 1 percentage point of independently trained references, with some intermediate widths beating existing elastic training by up to 11 points. The narrow-to-wide hybrid was within about 1.1 points of the high-capacity reference. These are offline video results, with no reported production A/B. Transfer compression compares with uncompressed 8-bit RGB, not a hardware video codec.

Local reproduction

Cost and deferral

Section 4.1 specifies 128 GPUs, 300 epochs, Kinetics-710/SSv2/HowTo100M video, and a frozen V-JEPA2 ViT-G teacher, training all three paths in each step. The compressor separately requires 100 epochs × 300 iterations with global batch 4096. This paper budget, not a local timing measurement, is well beyond one A100, one A30, or a local Mac. No full video download or multi-arm training was attempted.

The estimate assumes only one available accelerator or a Mac, not simultaneous multi-node capacity. A smaller experiment is not automatically authorized. Reopening requires an explicit budget decision, pinned public resources, and preservation of the defining mechanisms.

Data protocol, code, and results

No algorithm run, local metrics, model download, or passing GPU receipt exists. The deferred specification records components and future acceptance requirements but supplies no paper.yaml, adapter.py, or executable entry point. Paper tables, zero-training fixtures, and empty registry entries do not constitute reproduction.

Any future run must isolate train, validation, and test and select only on validation. Real train/inference paths, cost accounting, multiple seeds, and matched-budget controls must be executed. An advertised CUDA path still requires a sanitized A100/A30 receipt.

Reopening criteria

This remains deferred for cost and does not block other delivery or research review. Reopening requires a fresh budget assessment and necessary resources, not replacing defining mechanisms with heuristics or gold-answer access. The pinned source and specification support the review, not an implementation claim.