A memory for what comes next

FrameMorrow

Future-Guided Frame Selection with Prospective Tokens
for Long-Horizon Video Generation

Bo Yin¹ Xiaobin Hu¹ Jiaqi Zhao^1,2 Shuicheng Yan¹

¹ National University of Singapore² Harbin Institute of Technology, Shenzhen

Read the paper Code Watch results

See the difference

Same backbone. Better memory.

Native generation alongside
generation with FrameMorrow.

Native backboneCausal Forcing

+ FrameMorrowCausal Forcing

0:00 / 0:59

Long video generation

Follow the subject and scene as the video progresses. Both videos use Causal Forcing, with historical frame selection added on the right.

Synchronized playback

Enable JavaScript to use synchronized playback, or watch the native video and FrameMorrow video directly.

The intuition

Relevant now ≠ useful next.

Select history for future needs.

A corridor may match the current view. But returning to a room requires remembering its earlier appearance. FrameMorrow learns to retrieve historical evidence for the upcoming generation.

Conceptual illustration: context matching retrieves a corridor, while future-guided selection retrieves the earlier room and armchair needed for the next scene. — Conceptual illustration of current similarity versus future usefulness.

A compact future proxy

Look ahead. Select history.

No full future rollout required.

Anticipate information needs

Prospective tokens summarize what may matter next, using available history, recent context, and the generation condition.

Retrieve explicit frames

The tokens score historical candidates. A small set of selected frames carries relevant visual evidence forward.

Reuse across generators

Selected frames enter each backbone through its own conditioning interface. The selector stays outside the generator.

FrameMorrow overview: inference uses prospective tokens to select historical frames, while training distills future-derived teacher rankings into the selector. — Realized future observations supervise the selector during training only. Inference uses available history and the current generation condition.

Quantitative results

Measured across generators.

Orange rows: + FrameMorrow
↑ Higher is better · ↓ Lower is better

01Long video generationMovieGenBench · 60 seconds

VBench-Long on 128 MovieGenBench prompts
Model / method	Subject ↑	Background ↑	Motion ↑	Dynamic ↑	Aesthetic ↑	Imaging ↑	Avg. rank ↓
Self-Forcing
Native	95.84	95.27	98.20	51.72	56.05	62.22	4.33
+∞-RoPE	97.24	96.24	98.58	46.64	56.09	63.28	3.17
+Deep Forcing	96.08	95.38	98.24	41.44	56.68	60.81	4.17
+LongLive-RAG	97.60	96.51	98.70	44.69	57.19	64.97	2.00
+FrameMorrow	97.71	96.45	98.60	64.56	57.27	68.47	1.33
LongLive 1.0
Native	97.13	95.89	98.61	44.56	58.17	67.56	3.83
+∞-RoPE	97.00	95.85	98.53	53.36	57.48	66.94	4.25
+Deep Forcing	97.17	96.04	98.73	45.13	57.48	67.27	3.42
+LongLive-RAG	97.32	96.08	98.62	49.90	58.30	67.79	2.25
+FrameMorrow	97.32	96.17	98.75	50.42	58.75	67.82	1.25
Causal Forcing
Native	93.52	94.12	95.74	72.32	51.24	62.30	4.83
+∞-RoPE	93.81	93.78	96.09	92.47	54.42	67.50	3.33
+Deep Forcing	94.27	94.18	96.62	78.59	52.12	64.25	3.17
+LongLive-RAG	94.29	94.24	96.48	88.20	54.95	68.16	2.33
+FrameMorrow	94.54	94.28	96.59	89.74	56.27	68.76	1.33

Bold marks the best score within each backbone. Average rank is computed across the six metrics; lower is better.

02Interactive video generationSingle-shot & multi-shot · 60 seconds

Interactive video generation
Model / method	Quality ↑	Consistency ↑	Aesthetic ↑	Avg. CLIP ↑
Single-shot
LongLive 1.0	83.49	92.62	64.21	28.75
+FrameMorrow	84.06	93.21	64.56	29.08
Self-Forcing	78.71	84.95	58.16	26.39
+FrameMorrow	81.74	89.30	60.62	28.05
Causal Forcing	76.00	80.46	55.45	24.53
+FrameMorrow	78.66	83.48	58.75	25.24
CausVid	81.94	89.77	63.53	28.63
+FrameMorrow	82.26	90.52	63.64	29.03
Multi-shot
LongLive 2.0	82.77	93.14	61.32	28.35
+FrameMorrow	83.18	93.49	61.68	28.44
ShotStream	83.12	92.75	62.30	29.09
+FrameMorrow	84.85	96.76	61.74	29.16

Bold marks the better score within each backbone pair. Average CLIP score summarizes six consecutive 10-second segments.

03Interactive world modelsAction-conditioned generation

Action-conditioned interactive world generation
Model / method	Subject ↑	Background ↑	Imaging ↑	Anti-flicker ↑	Motion ↑	Action alignment ↑
Matrix-Game 3.0	0.801	0.894	68.20	0.936	0.950	0.842
+FrameMorrow	0.816	0.910	68.35	0.939	0.949	0.861
WorldMem	0.782	0.924	70.23	0.910	0.916	0.866
+FrameMorrow	0.793	0.932	70.47	0.913	0.918	0.889
YuMe 1.5	0.765	0.872	50.98	0.944	0.965	0.883
+FrameMorrow	0.798	0.886	51.80	0.947	0.967	0.912

Bold marks the better score within each backbone pair. Imaging uses a 0–100 scale; the other metrics lie in [0, 1].

One idea, multiple settings

Memory beyond a single generator.

Evaluated across long-video generation, interactive generation, and action-conditioned world models.

5benchmarks

11generative models