SceneTether
Activity-Gated Value Anchoring for
Long-Horizon Autoregressive Image Animation

Matiur Rahman Minar Seunghun Oh Ganghyeon Jeong Unsang Park

One photograph, one minute of motion

SceneTether
Spatial Echo-Forcing
DyMoS port
ALG-inspired
Input photograph of the rapids and shoreline

SceneTether is a training-free method for long-horizon image animation. It anchors low-activity regions to the input photograph while letting water, smoke, and fire continue to move.

Inspect the 58-second detail crops and motion maps
SceneTether shoreline detail at 58 secondsSceneTether late motion heatmap
SceneTether
Spatial Echo-Forcing shoreline detail at 58 secondsSpatial Echo-Forcing late motion heatmap
Spatial Echo-Forcing
DyMoS port shoreline detail at 58 secondsDyMoS port late motion heatmap
DyMoS port
ALG-inspired shoreline detail at 58 secondsALG-inspired late motion heatmap
ALG-inspired

Abstract

Long-horizon image animation has to preserve the scene as moving materials evolve. SceneTether estimates a spatial activity map during warm-up and uses it to anchor low-activity attention values to the source image during autoregressive generation. Active regions retain their recent values. On a Causal Forcing++ evaluation of fixed-camera natural scenes, SceneTether improves source-support PSNR over a spatial Echo-Forcing port by 5.88 dB at one minute and by 7.31 dB at four minutes on the pre-registered subset.

Observation

Autoregressive generators build each new segment from recent output. Over minutes, appearance errors can accumulate even in parts of the scene that should stay still. A flash-flood example shows why persistent scenery and localized motion need different treatment.

Frozen activity map for the flash-flood scene
Warm-up activity map
Untethered
SceneTether

Method

SceneTether uses a frozen activity map to decide where the attention read should retain recent values and where it should blend them toward the input image. It runs on a frozen Causal Forcing++ backbone and adds zero denoising evaluations.

SceneTether pipeline showing source prefill, activity-map estimation, and gated value anchoring during autoregressive generation
SceneTether estimates activity once from the first 1.5 seconds, then reuses that map throughout the rollout.

Store the source

Keep the input image’s clean values as a spatial reference.

Estimate activity

Identify moving regions during warm-up and freeze the spatial weights.

Blend at read time

Read co-located source values in stable regions and recent values in active regions.

Activity-gated value anchoring

Ṽᵣ(p) = (1 − β(p)) Vᵣ(p) + β(p) V₀(p)

At spatial position p, β controls how much of the source value enters the attention read. The weight is zero in active regions and increases in static regions.

See the attention read in detail Attention diagram showing interpolation of recent and source values while using all newest-frame keys

Baseline comparisons

Compare three rollouts generated from the same input on Causal Forcing++.

Rural river

Untethered
Spatial Echo-Forcing
SceneTether

Night rain

Untethered
Spatial Echo-Forcing
SceneTether

Flood at sunset

Untethered
Spatial Echo-Forcing
SceneTether

Firefighting

Untethered
Spatial Echo-Forcing
SceneTether

Long-horizon animation

Explore SceneTether clips at one, two, and four minutes from six fixed-camera scenes.

Tropical shore

1 minute
2 minutes
4 minutes

Steam vent

1 minute
2 minutes
4 minutes

Fire pit

1 minute
2 minutes
4 minutes

Rain on a city window

1 minute
2 minutes
4 minutes

Farmland sluice

1 minute
2 minutes
4 minutes

Dust storm

1 minute
2 minutes
4 minutes
Inspect seawall stills from 60 to 240 seconds
SceneTether seawall at 60 secondsSpatial Echo-Forcing seawall at 60 seconds
60 s
SceneTether seawall at 120 secondsSpatial Echo-Forcing seawall at 120 seconds
120 s
SceneTether seawall at 180 secondsSpatial Echo-Forcing seawall at 180 seconds
180 s
SceneTether seawall at 240 secondsSpatial Echo-Forcing seawall at 240 seconds
240 s

SceneTether appears in the upper row of each pair. Spatial Echo-Forcing appears below.

Articulated motion

Exploratory one-minute rollouts on animal scenes beyond the fixed-camera natural-phenomena evaluation.

Blue tit
Giant panda
Lioness
Retriever

BibTeX

Download
@misc{minar2026scenetether,
  title         = {SceneTether: Activity-Gated Value Anchoring for Long-Horizon Autoregressive Image Animation},
  author        = {Matiur Rahman Minar and Seunghun Oh and Ganghyeon Jeong and Unsang Park},
  year          = {2026},
}