CRePE: Curved Ray Expectation Positional Encoding (NeurIPS 2026)

FoV 180°, ξ = 2.0

FoV 190°, ξ = 2.1

FoV 200°, ξ = 2.3

TL;DR. CRePE tells a video diffusion transformer where scene content lies along each camera ray, and how that ray bends under the lens of every other view. With only a prompt, a camera trajectory and lens parameters, it generates videos that follow the camera motion under pinhole, wide-angle and fisheye lenses (up to 200° field of view).

Method

Each token predicts where its content lies along its camera ray. CRePE projects that range into every other view through its lens and encodes the expected RoPE phase along the curved path.

3D view needs WebGL.

Key frame · click to pick a token

Query frame · curved path, endpoint chord dashed

Expected phasor

μ σ query c ω

Curved ray expectation

𝔼[e^iωc] = 1/(K−1) · Σ_k (e^iωc_k+1 − e^iωc_k) / (iω (c_k+1 − c_k))

Radial anchor

Pseudo radial distances (UniK3D) supervise one layer during training. No depth input at inference.

Radial MixForcing

Training sometimes substitutes a supplied radial map, so external maps can steer generation.

Lens showcase

Each clip is generated from a prompt, a camera trajectory and a lens. Switch to the UCPE baseline (same inputs), other seeds, or an equirectangular view that undoes the lens.

Camera control comparison

Zero-shot camera control on novel, fast camera trajectories: real camera paths of held-out RealCam-Vid clips (metric scale, pinhole lens), unseen in training. Over 30 such trajectories (38–80° turn, 6–18 m travel), CRePE's median rotation error is 12.8° versus 18.6° for UCPE. Both models get the same prompt, camera and seed; cameras are recovered from the generated videos with ViPE. Selected examples.

Reference

CRePE

UCPE

frame 0

Recovered cameras, top view (dashed: requested). Drag to rotate.

Pan angle vs. frame

	Pan at end	Rot. err.	CamMC

Internal geometry

What CRePE's radial head predicts at layer 15 while generating each clip: a positional variable inside attention, not a depth output. near far

External radial-map control

With Radial MixForcing, the same pathway accepts a radial map: generate a new video from a source video's camera and geometry, or transfer its motion. Radial map: red near, blue far.

Source video

Supplied radial map

CRePE-Mix (generated)

Results

PanShot test set (340 clips). Full metrics and ablations are in the paper.

Subset	Method	CamMC ↓	CS-I ↑	CS-T ↑	k₁ ↓	k₂ ↓	Gravity ↓	Aes. Qual. ↑	BG Cons. ↑
Non-pinhole	ReCamMaster	54.56	98.69	24.60	0.186	0.154	11.19	0.5596	0.9484
	UCPE	20.60	98.80	24.78	0.183	0.154	7.08	0.5616	0.9537
	CRePE	19.02	98.81	24.80	0.157	0.116	6.27	0.5639	0.9542
Pinhole	ReCamMaster	57.83	98.65	25.58	0.179	0.152	10.98	0.5605	0.9441
	UCPE	18.50	98.59	25.86	0.175	0.151	7.10	0.5619	0.9457
	CRePE	15.33	98.65	25.93	0.153	0.113	6.13	0.5620	0.9471

Architecture

BibTeX

@inproceedings{jin2026crepe,
  title     = {CRePE: Curved Ray Expectation Positional Encoding for Unified-Camera-Controlled Video Generation},
  author    = {Jin, Seonghyun and Kim, Youngmin and Park, Sunwoo and Ye, Jong Chul},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}