STORY 01
JoyAI-Echo 1.0 Visual memory · Multi-shot video · Preprint
REELLearning to Retrieve, Evaluate, and Evolve Visual Memory
for Any Long-Horizon Multi-Shot Video Generation
1 University of Science and Technology of China2 JD Explore Academy
One planner learns what to remember and when to reuse it.
Different video backends bring the story to life.
01 / REEL in action
A reference.
A story that continues.
Explore selected moments from four stories in the paper. Click any sequence for a closer look.
STORY 02
JoyAI-Echo 1.5A Teenage Courier and a Cracked Lantern
STORY 03
MiniMax-H3The Cobalt Guardian and the Calico
STORY 04
Seedance 2.0The Weaver’s Ribbon
Selected frames from the paper’s teaser, not consecutive video playback. Original shot indices are retained.
02 / Main results
A closer look
at the gains.
REEL, the no-memory baseline, and DirectorAgent. Two benchmarks, two backends, and the primary metrics.
Loading results…
Core-100
Overall visual quality on REEL-Bench.
Zoomed score axisHigher is better ↑
The score axis is zoomed and does not start at zero. Marker positions show scores, not relative improvement percentages.
Paper-reported scores on complete, decodable stories; failed and incomplete stories are excluded from quality averages. DirectorAgent uses Claude Opus 4.5. Full comparisons and evaluation details in the paper ↗
03 / The idea
Memory planning,
independent of the backend.
REEL retrieves visual evidence, creates missing references, reviews candidates, and updates memory. Adapters translate its plans for each video backend.
Retrieve relevant evidenceEvaluate the candidateEvolve visual memory
04 / Explore & build
Take a closer look.
PaperHugging Face
Read the preprint
The method, full results, and supplementary material.
Open PDF ↗ ImplementationExplore the code
Memory-planning tools, backend interfaces, and training components.
xuhang07 / REEL ↗Planner weights
Three training stages. Multi-shot GRPO uses the step-50 checkpoint.
Citation
If REEL helps your research
@misc{xu2026reel,
title = {REEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory
for Any Long-Horizon Multi-Shot Video Generation},
author = {Hang Xu and Guohui Zhang and Xiaoxiao Ma and Hu Yu and
Jie Huang and Jiayin Xiao and Lin Song and Haoyang Huang and
Nan Duan and Feng Zhao},
year = {2026},
note = {Preprint},
url = {https://xuhang07.github.io/REEL/}
}