Stream3D

Sequential Multi-View 3D Generation via Evidential Memory

CVPR 2026 Workshops

A training-free wrapper that turns frozen single-view 3D generators into coherent streaming generators with bounded cross-chunk memory.

Kaichen Zhou^* Zeyang Bai^* Xinhai Chang^* Mengyu Wang^† Paul Pu Liang^† Fangneng Zhan^†

World Mind Lab, HKUST Media Lab and EECS, MIT Kempner Institute, Harvard University

^* Equal contribution ^† Joint supervision

Stream3D output Drag to orbit

Long streams.Constant memory.

Stream3D treats streaming 3D generation as an evidence selection problem, not a latent transport problem.

View-conditioned 3D generators create compelling assets from a single image, yet real observations arrive as long monocular streams. Processing every frame independently creates inconsistent geometry and appearance.

Stream3D keeps the underlying generator frozen. A lightweight attention probe scores the evidence contributed by each view, then a compact token-wise memory retains only the most informative historical frames for the next generation step.

Training-freeNo retraining, architecture changes, or auxiliary losses.

Bounded stateCross-chunk memory remains \(\mathcal{O}(QD)\), independent of stream length.

Evidence-awareDifferent 3D tokens retain the historical views that support them best.

Stream3D framework showing attention probes, evidence scores, adaptive evidential memory, and evidence-based multi-generation — A lightweight warmup pass extracts token-wise evidence. Top-ranked views update a fixed-capacity memory and condition the frozen 3D generator.

Interactive results

Compare ground-truth geometry with SAM3D and Stream3D on selected GSO objects. Orbit any asset to inspect all three from the same viewpoint.

If 3D assets do not load in Chrome, open chrome://settings/system and enable Use graphics acceleration when available.

Loading selected result...

Quantitative evaluation

Measured gains

Stream3D improves appearance and geometry on both GSO and NAVI. Values are taken from the latest paper tables.

Dataset	Method	CD ↓	IoU ↑	PSNR ↑	LPIPS ↓
GSO	SAM3D	0.094	0.664	14.178	0.178
GSO	Stream3D	0.048	0.775	16.145	0.139
NAVI	SAM3D	0.138	0.721	16.159	0.132
NAVI	Stream3D	0.128	0.741	16.474	0.123

Side-by-side comparison

Follow each input stream alongside SAM3D and Stream3D. Orbit either reconstruction to inspect both outputs from the same viewpoint.

Loading selected comparison...

Downstream robotic manipulation

Each record pairs eye-in-hand input, the corresponding reconstructed geometry, and fixed third-person execution verification.

VGGT + SAM

Unsuccessful execution

Banana

First-person video

3D geometry · VGGT point cloud

Third-person video

Stream3D

Successful execution

Banana

First-person video

3D geometry · Stream3D mesh

Third-person video

VGGT + SAM

Unsuccessful execution

Red pepper

First-person video

3D geometry · VGGT point cloud

Third-person video

Stream3D

Successful execution

Red pepper

First-person video

3D geometry · Stream3D mesh

Third-person video

Memory that does not grow with time

Stream3D stores evidence scores and frame indices, not an ever-growing history of latent activations.

For each query token, two \(Q \times D\) matrices retain the top-\(D\) evidence values and their source frame indices. New views can replace weaker evidence, while frames that never enter any token list are discarded.

The cached evidence at every token and rank is monotonically non-decreasing. This is an evidence-space property, not a claim that every generated sample must improve monotonically.

Memory footprint

Cross-chunk state

Constant in \(T\)

Incoming views

Fixed-capacity evidence \(2QD\) scalars, constant in \(T\)

Evidence memory and generated 3D mesh improving from chunk 0 to chunk 12 — As stronger evidence arrives, retained views change and the reconstructed asset becomes more complete.

Method at a glance

How evidential memory works

Probe which views matter, retain the strongest token-wise evidence, then let the frozen generator use the selected bundle.

01Probe

Score evidence

Strong and selective cross-attention signals that an incoming view supports a particular 3D token.

Combine attention strength with selectivity.

02Retain

Keep the top evidence

For every token, retain only the strongest evidence scores and the source frames that provided them.

Memory capacity stays fixed across the stream.

03Select

Vote for views

Count how often each historical frame is retained, then select the most-supported frames for generation.

Only the selected views condition the frozen generator.

The exact evidence score, fusion rule, and streaming algorithm are detailed in the paper.

Citation

Please cite the CVPR 2026 Workshops paper if Stream3D supports your research.

@inproceedings{zhou2026stream3d,
  title     = {Stream3D: Sequential Multi-View 3D Generation via Evidential Memory},
  author    = {Zhou, Kaichen and Bai, Zeyang and Chang, Xinhai and
               Wang, Mengyu and Liang, Paul and Zhan, Fangneng},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and
               Pattern Recognition (CVPR) Workshops},
  year      = {2026}
}