RECAP-Forcing

REtaining Content APpearances for Long Video Generation

Haiyang Xu   Zheng Ding†   Zhuowen Tu

UC San Diego · †Project Lead

2-minute overview — the problem, the idea, and the results.

Every generated video on this page is 60 seconds, generated in a single pass with no post-processing — RECAP-Forcing applied training-free to a 1.3B causal autoregressive checkpoint.

Long videos should remember what appeared, not when.

RECAP-Forcing is a memory for causal video generators. It stores the KV cache of newly appearing content once, at first sight, so subjects and scenes stay consistent across a full minute while the video keeps moving.

27.5→58.1Dynamic Degree on Self-Forcing — more than double the motion
79.5best Total among training-free memory methods (next best: 78.9)
0.61→0.81subject identity after 5 minutes (DINOv2 similarity to frame 0)
~1.6×constant inference overhead, regardless of video length
Read the abstract

Long autoregressive video generation faces a fundamental memory challenge: with a finite attention window, a model must decide which information from an ever-expanding history to retain. Existing methods organize memory temporally, preserving recent frames while compressing or discarding older ones. We instead propose RECAP-Forcing, organizing memory by appearance novelty. A long video is not merely a sequence of frames, but an evolving cast of subjects, objects, and scenes whose identities must remain consistent over time. We organize memory by retaining the KV cache associated with newly appearing content — such as entering subjects, disoccluded regions, and newly introduced scenes — at the moment it first becomes visible, prioritizing novelty over recency. Memory should scale with the amount of newly introduced content, rather than with video length. This appearance-indexed memory makes long-range consistency an explicit property of the memory structure. Our framework unifies two mechanisms under this single principle. At the beginning of a video, when all visible content is novel, an attention sink preserves the initial scene. As the video evolves, an optical-flow-based novelty bank extends the same principle by selectively retaining newly revealed content. As a training-free inference method with no additional learnable parameters, RECAP-Forcing consistently improves visual quality and semantic fidelity across multiple strong baselines and outperforms existing memory methods.

01Memory Should Grow with New Content, Not with Time

A causal video model cannot attend to everything it has generated, so something has to be forgotten. Existing methods index memory by time: keep the recent past sharp, compress or drop the rest. We index it by appearance novelty instead — record content once, when it first appears, and keep it.

The same rollout indexed two ways: frame-indexed memory re-stores content at every step and evicts it over time, while appearance-indexed memory stores content once, when it newly appears.
The same rollout, indexed two ways. Frame-indexed memory (left) re-stores old and new content at every step, then compresses or evicts it, so memory pressure grows with video length. Appearance-indexed memory (right) stores each new appearance once — the man at t1, a passing car at t2, a pedestrian at t3, another passer-by at t4 — at canonical temporal position 0.

Prior · indexed by time

Caught between forgetting and anchoring

  • Drift. As old content is evicted or compressed, the model loses its visual evidence and subject identity slowly wanders.
  • Freeze. Pinning a fixed reference pulls later frames toward the same appearance and suppresses motion.

Ours · indexed by appearance

Remember what is new, once

  • Store at first sight. Entering subjects, newly revealed regions and new scenes are recorded the moment they appear.
  • Retain by novelty, not recency. Memory grows with the amount of new content, not with video length.
  • Content from second three is still directly accessible at minute five.

02How RECAP-Forcing Works

A training-free, parameter-free inference plug-in for causal autoregressive video diffusion (Wan2.1-T2V-1.3B distilled with Self-Forcing). One principle — record what appears and retain it — applied at two stages of generation.

Every query attends over

t = 0 · the first frame

Reinforced Sink

Nothing has been seen yet, so the opening frame is new in its entirety. We turn it into a persistent first-impression memory by boosting its attention weight by λ = 5, so the identity of the opening subjects is held, not only the overall color and style.

softmaxj ( q⊤kj / √d + βj ),  βj = ln λ for sink keys, else 0

t > 0 · every frame after

Novelty Bank

Later, “new” means content not observed before — a subject entering, a region disoccluded by motion. Optical flow flags it, and its per-layer key/value vectors go into a fixed-size bank that keeps the K most novel entries, whatever their age.

RECAP-Forcing memory bank: candidate selection, KV gather, and selected patches
(a) Each frame pair contributes newly appearing patches; their per-layer keys and values are copied from the KV cache into the bank. (b) Green cells mark stored patches — entering subjects and newly revealed parts, never content already seen.

Inside the novelty bank

  1. DetectRAFT optical flow between consecutive frames. Pixels with no reliable antecedent — they entered the frame or were disoccluded — receive a novelty score.
  2. StoreCopy the key/value vectors of each novel patch from every attention layer into the bank, frozen together with its score.
  3. RetainKeep only the top-K entries (one frame’s worth of tokens). Bank keys use temporal RoPE phase 0, so a patch survives by novelty, never by recency.

One setting for every backbone, video length and prompt set. Per-block cost stays constant.

Technical details — preliminaries, novelty scoring, hyperparameters

Preliminaries

We build on causal, block-wise video generators such as Self-Forcing, which distill a bidirectional video diffusion transformer into a few-step causal generator. A video is produced as a stream of 3-frame blocks; each latent frame is a 30×52 grid of 1,560 tokens, and a block is denoised conditioned only on a bounded, sliding KV cache covering 6 latent frames — one persistent sink frame, two prior frames, and the current block. The bound keeps per-block compute constant and enables arbitrary-length generation, but it also forgets by construction: once a token scrolls past the window it is gone, and the opening frame is its sole survivor. As more of the early visual context is discarded, the model retains progressively less information about the original appearance; small inconsistencies accumulate over time, gradually causing the generated content to drift from its earlier identity.

What counts as new?

A long video does not need its timeline; it needs its cast. An entity's appearance is established by a sparse set of novel appearance events — its entry and the later sightings of newly revealed parts — and nothing about the intervening forty seconds is required to render it faithfully to keep consistency. This raises a natural question: what counts as new? The answer has two exhaustive cases: at t = 0 nothing has been seen, so the opening frame is new in its entirety; afterwards, content is new only relative to what has already appeared — an object entering the frame, a region disoccluded by motion, a style introduced by a change of scene.

The first frame: a reinforced sink

The base model already keeps frame 0 as an attention sink, but under the default attention distribution later queries assign insufficient weight to it: the sink mainly anchors global appearance characteristics — color tone, background, overall style — and helps prevent quality degradation, while the identity of subjects introduced in the opening scene can still drift. We therefore strengthen the sink attention, promoting it from a stability anchor to a persistent first-impression memory that is more strongly attended to by queries, by adding a fixed positive bias to the pre-softmax logits of the sink keys:

αj = softmaxj ( q⊤kj / √d + βj ),  βj = ln λ for sink keys, 0 otherwise.

Adding ln λ multiplies each sink key's unnormalized attention weight by λ, uniformly raising the sink's prior importance while preserving content-based selection through query–key similarity. Because the bias is applied before the softmax, the distribution stays normalized — sink keys still compete with all other keys — and it is directly compatible with fused attention kernels. Its magnitude is a measurement rather than a choice: we set λ = 5, with its effect analyzed in the λ sweep in Results.

Every frame after: an appearance novelty bank

The reinforced sink covers the opening scene, but a long video continually introduces new content: a second subject enters the frame, a camera pan reveals a storefront, or an occluded region becomes visible. After t = 0, “new” therefore means content that has not been observed before, making novelty detection a property of the video stream rather than of the model's current query. Optical flow provides a direct cue, and admission is query-agnostic: patches are retained by visual appearance novelty rather than query–key affinity, without entity grouping or re-identification. The bank capacity K is a configurable memory budget rather than an architectural constant: one frame's worth of tokens (K = 1,560) works well for the one-minute videos in our experiments, and longer videos can naturally allocate additional slots. For each newly decoded frame, RAFT forward–backward flow runs against its predecessor: a pixel u is newly appeared when its backtrace finds no reliable antecedent — the cycle residual ecyc exceeds 2 px, the photometric residual epho exceeds 0.2, or the backtrace leaves the frame. Such pixels score

ν(u) = ecyc(u) + α epho(u)(α = 0.5; all other pixels score 0)

and a patch takes the maximum novelty over its pixels, normalized by its block's typical firing strength and frozen from then on:

s(p) = maxu ∈ p ν(u),  ŝ(p) = s(p) / max(ρb, 1)(ρb = 90th percentile of positive scores in block b)

This block-wise normalization prevents brief bursts of large motion from dominating the fixed budget, so admissions depend on relative novelty rather than absolute motion magnitude. The map of already-claimed novelty is propagated with the flow, so continuously traceable content fires only once, when it first appears; if a region returns after its flow correspondence has been broken, it can register as a new appearance event and receive a new anchor.

Each newly decoded frame is compared with its predecessor using RAFT to obtain a novelty map, from which newly appearing patches are selected (Figure 2a). For each selected patch, we copy the corresponding key/value vectors from every attention layer of the KV cache and store them with their frozen novelty scores. At retrieval, bank keys use temporal rotary phase 0 while retaining their original spatial phase, avoiding unseen long-range temporal RoPE offsets and treating the bank as a time-neutral appearance memory. The bank is updated with a simple top-K rule: among all candidates observed so far, only the K highest-scoring patches are retained. Memory retention is therefore time-neutral and parameter-free: a patch remains in the bank according to its appearance novelty rather than how recently it was observed, without a learned controller or an explicit recency term.

Hyperparameters

HyperparameterValue
Sink size1 latent frame
Attention window6 latent frames
Block size3 latent frames
Bank capacity K1,560 tokens
Sink reinforcement λ5
Flow cycle-error threshold2 px
Photometric-error threshold0.2
Photometric weight α0.5
Score normalizationper-block 90th pct.

03Results

VBench-Long on a frozen set of 128 prompts × 5 seeds (640 videos per configuration), 60-second videos, fixed checkpoints.

Gains on every backbone

One inference hook, four backbones: Total improves on all of them, with the largest gains in motion (Dynamic Degree) and semantics — without sacrificing visual quality.

Baseline → + RECAP-Forcing on VBench-Long (128 prompts × 5 seeds, 60-second videos; higher is better). The same training-free hook is attached to every backbone; Helios is 14B, the others 1.3B. Quality is a 7-dimension score, Semantic a 9-dimension score, and Total = (4 Quality + Semantic) / 5.
BackboneDynamic DegreeQualitySemanticTotal
Self-Forcing27.5→58.1+30.680.4→82.9+2.558.0→65.5+7.575.9→79.5+3.6
Infinite-Forcing61.3→71.3+10.082.2→83.3+1.162.9→65.4+2.578.3→79.7+1.4
LongLive20.3→26.6+6.381.2→81.5+0.365.9→66.8+0.978.1→78.5+0.4
Helios (14B)36.2→47.8+11.676.6→80.2+3.645.4→60.5+15.170.3→76.3+6.0
Full VBench-Long results — all 16 dimensions
Main results on VBench-Long across all 16 dimensions (128 prompts, 5 seeds, 60-second videos; all values are percentages, higher is better). RECAP-Forcing — an optical-flow memory bank plus a reinforced sink — is applied as the same training-free inference hook to every base model (its "+ Ours" row); for Self-Forcing we additionally report a plain, unreinforced sink (λ = 1). Helios is 14B, all other models are 1.3B.
Method S.C. B.C. T.F. M.S. D.D. A.Q. I.Q. O.C. M.O. H.A. C. S.R. S. A.S. T.S. Ov.C. Q. Sem. Total
Self-Forcing 98.096.898.398.927.553.869.279.532.177.072.862.612.120.620.919.680.458.075.9
+ Sink (λ = 1) 98.297.098.398.930.453.367.989.241.175.178.273.614.620.420.321.280.462.376.8
+ Ours 97.796.397.498.558.158.470.991.847.377.194.668.417.718.722.822.182.965.579.5
Infinite-Forcing 97.495.997.198.461.355.868.893.649.974.181.964.012.318.622.921.982.262.978.3
+ Ours 97.496.097.298.471.357.568.897.148.975.796.163.415.118.523.022.283.365.479.7
LongLive 98.797.098.599.220.357.171.698.252.275.980.480.514.818.122.722.681.265.978.1
+ Ours 98.597.098.599.126.657.371.098.450.776.782.780.419.517.923.022.881.566.878.5
Helios 95.897.498.699.136.245.249.036.18.966.753.649.04.621.220.321.376.645.470.3
+ Ours 96.497.398.198.847.853.660.083.728.580.072.263.418.619.322.223.780.260.576.3

A plain, unreinforced sink on Self-Forcing already helps (Total 75.9 → 76.8) but recovers little motion (Dynamic Degree 27.5 → 30.4); RECAP-Forcing more than doubles Dynamic Degree to 58.1 while improving aesthetic and imaging quality. On Infinite-Forcing, whose sink already raises Dynamic Degree to 61.3, it adds another +10.0 (71.3).

Comparison of RECAP-Forcing with other training-free methods, each applied on top of the Self-Forcing base model.
MethodS.C.B.C.T.F.M.S.D.D.A.Q.I.Q.O.C.M.O.H.A.C.S.R.S.A.S.T.S.Ov.C.Q.Sem.Total
Self-Forcing98.096.898.398.927.553.869.279.532.177.072.862.612.120.620.919.680.458.075.9
+ Deep Forcing98.493.798.294.832.055.971.898.348.576.682.973.121.818.323.022.378.766.176.2
+ MemRoPE98.496.798.099.041.256.673.194.244.176.684.868.415.919.121.821.182.463.678.7
+ Rolling Sink98.997.398.799.223.058.372.697.948.280.790.376.517.618.223.222.581.867.178.9
+ Ours97.796.397.498.558.158.470.991.847.377.194.668.417.718.722.822.182.965.579.5

Best among training-free memory methods

Deep Forcing, MemRoPE and Rolling Sink improve selected metrics but recover much less motion (Dynamic Degree 23.0–41.2 versus 58.1), while RECAP-Forcing reaches the best overall score and stays competitive in temporal consistency.

Training-free memory methods, each applied on top of the same Self-Forcing base model (best value per column in bold).
MethodSubject ConsistencyDynamic DegreeQualitySemanticTotal
Self-Forcing98.027.580.458.075.9
+ Deep Forcing98.432.078.766.176.2
+ MemRoPE98.441.282.463.678.7
+ Rolling Sink98.923.081.867.178.9
+ Ours97.758.182.965.579.5

Both components matter

The two mechanisms split the work. The reinforced sink anchors what is established early (better consistency and semantics, but less motion); the memory bank lets later content develop without drifting (more motion). Together they give the best Total. We ablate on Infinite-Forcing, because vanilla Self-Forcing has no sink frame to reinforce.

Ablation of the reinforced sink and memory bank (VBench-Long, on Infinite-Forcing). Check marks show which component is enabled in each row; the last row, with both enabled, is the full model.
Reinf.
Sink
Memory
Bank
S.C.B.C.M.S.D.D.A.Q.I.Q.Q.Sem.Total
✗✗97.495.998.461.355.868.882.262.978.3
✓✗98.296.798.841.657.868.981.965.278.6
✗✓96.895.598.178.255.268.582.863.078.8
✓✓97.496.098.471.357.568.883.365.479.7
Sensitivity to λ and bank size

Both knobs are stable around their defaults (bank of 1 frame, λ = 5). A larger bank buys only marginal Total gains, and 8 frames drops it to 79.3. Raising λ from 1 to 5 lifts Total from 78.8 to 79.7, while 8, 10 and 12 over-constrain the dynamics and degrade it to 79.2, 78.9 and 78.6.

Lambda sweep
λ sweep. Sink attention grows smoothly with λ (~24% at λ = 5); Subject Consistency is already saturated while Dynamic Degree erodes past λ ≈ 8 — hence λ = 5.
Ablation of RECAP-Forcing's memory bank capacity and sink-reinforcement strength λ, sweeping each around its default (bank capacity of 1 frame, λ = 5).
Bank Capacity (frame) Sink-Reinforcement Strength (λ)
Metric 1/4 1/2 1 2 4 8 1 3 5 8 10 12
Q. 83.083.283.183.483.382.9 82.883.083.182.782.582.2
Sem. 66.065.865.765.465.965.1 63.064.565.765.264.464.1
Total 79.679.779.779.879.879.3 78.879.379.779.278.978.6

Preferred by people

In a blind pairwise study (9 participants, 216 judgments), RECAP-Forcing was preferred over the baseline and three competing memory methods on every axis.

Win rate of RECAP-Forcing against Self-Forcing, Deep Forcing, MemRoPE and Rolling Sink; the dashed line marks 50% (chance). “Hard to say” judgments count as ties with half a win.

Per-competitor breakdown
Human study results. Win rates of RECAP-Forcing in blind pairwise comparisons against each competitor. "T" denotes the number of hard-to-say judgments, treated as ties with 0.5 credit.
CompetitorConsistencyMotionOveralln
Self-Forcing85.2% (6T)76.9% (7T)84.3% (3T)54
Deep Forcing73.1% (15T)76.9% (5T)77.8% (4T)54
MemRoPE84.3% (7T)67.6% (5T)75.0% (3T)54
Rolling Sink82.4% (7T)76.9% (5T)79.6% (6T)54
All81.3% (35T)74.5% (22T)79.2% (16T)216

Identity holds through the full minute

Recency memory fails by mid-video: subjects lose identity or duplicate, a second subject that enters later never stabilizes, and the style drifts from the prompt. With RECAP-Forcing the same base model holds subject, scene and style — including subjects introduced long after the opening frame — while motion stays natural.

Frame strips of four one-minute prompts, baseline vs RECAP-Forcing
Qualitative results on four one-minute prompts spanning photographic, claymation, and cartoon styles. Across all scenes, the baseline loses subject identity, style, scale, or background consistency; RECAP-Forcing preserves the subject(s) and the full scene throughout the minute.

05Comparison with Training-Free Memory Methods

The same 32 prompts, now against three prior training-free memory methods, all applied at inference only. Prior memories stabilize the rollout by freezing it; RECAP-Forcing restores motion while holding identity.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A full-body shot of a man crafted entirely from rocks, walking through a dense forest. His rocky form is rugged and textured, with various shades of gray and brown. He strides confidently, his steps creating small ripples in the forest floor.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A delicate watercolor illustration of a red fox trotting through a quiet birch forest blanketed in snow. The fox's coat is a rich rust-orange with a white chest and black-stockinged legs, its thick tail streaming behind as it moves in a light, even gait. Slender white birch trunks repeat into the pale distance, their bark flecked with soft grey strokes, and loose washes of blue shadow pool across the drifts. Small puffs of snow lift from its paws with each step. The palette is muted and wintry, with soft blooms of pigment bleeding into the paper's texture. Snow lies untouched in every direction as the fox continues its steady trot. A low tracking shot follows at the fox's shoulder height, keeping its bright fur the focus.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A breathtaking nature documentary shot of a humpback whale gliding through deep, open blue ocean. The whale's massive body is a mottled slate-grey, barnacles crusting its long white-edged pectoral fins, which sweep slowly like wings as it swims. Shafts of sunlight pierce the surface far above, breaking into shimmering curtains that ripple across its back. A steady trail of small silver bubbles streams behind it. The surrounding water is a uniform, endless gradient of sapphire fading to darkness below, with drifting motes of plankton catching the light. The whale maintains an unhurried, continuous glide through the vast blue. The camera holds a slow parallel tracking shot at the whale's flank, conveying its serene, colossal grace.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

An extreme macro shot in a naturalistic style of a honeybee hovering at a cluster of lavender blossoms, drinking nectar in slow, deliberate sips. The bee's fuzzy amber-and-black body glistens with fine golden hairs, pollen dusting its hind legs in bright saffron clumps. Its wings beat in a shimmering blur while it works between the small purple petals. The background dissolves into a soft, uniform bokeh of mauve and green, with other stalks swaying as gentle out-of-focus shapes. Warm summer light saturates the frame, catching the translucent veins of the wings. The bee shifts calmly from blossom to blossom along the same swaying stalk. A razor-thin depth of field keeps the bee crisp in a steady, intimate close-up.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A cinematic wide shot of a lone rider swathed in flowing indigo robes and a loose cream headscarf, crossing an endless sea of golden desert dunes on a tall, sand-colored camel. The camel plants each padded foot with a slow, swaying gait, its tasseled saddle blanket woven in deep reds and oranges rocking gently with every step. Ripples of windblown sand texture the dunes, which repeat in soft identical waves to the horizon under a cloudless amber sky. Fine grains drift from the crests in thin veils, and long late-afternoon shadows stretch across the slopes. The rider sits relaxed and upright, at ease with the animal's rhythm. The camera tracks smoothly alongside, keeping rider and camel framed against the vast, unbroken dunes.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A vibrant cel-animated scene of a single hot air balloon drifting serenely above endless rolling green hills. The balloon's envelope is striped in cream, coral, and deep teal, with a small woven wicker basket swaying softly beneath it. The hills below repeat in gentle emerald waves, dotted with tiny wildflowers and threaded by the shadows of slow-moving clouds. The balloon rides a steady breeze, bobbing almost imperceptibly as its burner glows now and then with a soft orange flicker. The sky is a clear morning gradient of powder blue fading to pale gold near the horizon. Hand-painted clouds drift lazily past at a constant, unhurried pace. A wide, slowly drifting shot keeps the balloon centered in the tranquil sky.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A soft watercolor illustration of a young girl in a buttercup-yellow raincoat and red rubber boots walking slowly through a steady, gentle rain along a tree-lined park path. She holds a clear umbrella speckled with painted droplets, tilting it now and then as she strolls at an easy, wandering pace. Loose washes of green and grey suggest wet hedges and lawns on either side, with fine white lines of rain streaking down the paper. Puddles bloom with soft rippling rings around her boots. The palette is fresh and calm, the light silvery and even beneath a uniformly overcast sky. The rain falls at the same soothing rate throughout. A medium tracking shot follows a few steps behind, capturing her small, cheerful stride.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A whimsical 3D animated scene of a small plum-purple dragon curled contentedly atop a mound of glittering gold coins inside a warm candlelit cavern. The dragon's scales shimmer with iridescent highlights, tiny wings folded on its back, and its tail flicks in a slow, happy rhythm. It puffs perfect little smoke rings from its snout at regular intervals, each ring wobbling upward and dissolving into the amber gloom. Softly flickering torch flames line the rough stone walls, making the treasure sparkle with shifting points of light. Its half-closed emerald eyes convey deep, drowsy satisfaction. The cavern remains still and cozy around it. A close medium shot at coin level holds the dragon centered, savoring each looping ring of smoke.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A luminous oil painting in a romantic style of a white swan gliding across a still, dusky lake. The swan's plumage is rendered in creamy impasto strokes touched with lavender shadow, its slender neck curved in a perfect arc as it drifts with slow, dignified poise. Behind it trails a soft V of ripples that catches the last apricot light of evening. Water lilies float in loose clusters across the glassy surface, and a hazy treeline dissolves into violet mist at the far shore. The sky glows in blended bands of rose and amber, scattering broken shimmers of color around the bird. The swan paddles onward in an unbroken, tranquil line. A serene wide shot balances the swan against the glowing lake.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A fantasy 3D rendered scene of a gentle forest spirit shaped entirely from soft moss, tiny ferns, and dew-lit clover, walking slowly through an endless glade of towering ancient trees. Its rounded body glows faintly green from within, and miniature white flowers bloom along its shoulders, trembling with each unhurried step. Fireflies of golden light drift around it in slow orbits, and its footprints briefly glow before fading into the undergrowth. Shafts of hazy emerald light fall between the uniform mossy trunks that repeat into the misted depths. The spirit's pace is calm and constant, its small head turning gently to admire the ferns. A smooth low tracking shot follows through the undergrowth, wrapped in the glade's soft enchanted glow.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A delightful claymation scene of a rosy-cheeked garden gnome with a tall crimson felt hat and a snowy clay beard watering a bed of oversized daisies outside his mushroom-capped cottage. He tips a dented tin watering can in slow, looping passes, sparkling bead-like droplets pattering onto the petals, which bounce softly in response. His patched blue coat and tiny leather boots show charming fingerprint textures. The garden around him is a tidy, continuous border of clay flowers and pebble paths beneath a felt-cloth sky of even powder blue. Warm, diffused light gives every surface a handmade glow. He shuffles contentedly along the same flower bed, watering in rhythm. A fixed wide shot frames gnome and cottage like a storybook diorama.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A dynamic aerial drone shot of a windsurfer skimming across an open turquoise sea, leaning back against a taut lime-and-white sail as the board slices a clean white wake through the sparkling swell. Spray bursts rhythmically from the nose of the board, and the surfer's coral rash guard flashes in the bright sun. The sea rolls in long, even waves to a sharp horizon in every direction beneath a cloudless sky. Far off on that horizon, a second sail begins as a tiny bright sliver, growing steadily larger and clearer as it carves closer on the same wind, until it shares the frame as an equally clear subject. The scene settles into a balanced two-windsurfer composition, both riders slicing parallel wakes in a continuous, spray-flinging glide, tracked smoothly from above in glittering light.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

An enchanting macro shot of a ladybug trundling along a dew-beaded blade of grass, its lacquered scarlet shell gleaming with crisp black spots as morning light flares through the droplets. Its tiny legs work in a busy, constant rhythm, antennae tapping gently ahead, while the blade sways softly under its miniature weight. The background is a dreamy, uniform bokeh of glowing greens and golds. From that soft blur, a second ladybug gradually sharpens into focus as it climbs a neighboring blade, its red shell gaining clarity and sparkle as it draws level, until it shares the frame as an equally clear subject. The scene settles into a balanced two-ladybug composition, both beetles trundling along their parallel blades in a continuous, gentle rhythm, held in an extreme close-up with jewel-like dew light.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A rich oil painting in a chiaroscuro style of a cellist in a deep emerald gown performing in a vast candlelit chamber with dark velvet drapery. Her bow draws long, unhurried strokes across the strings, candle flames trembling in golden clusters around her, light pooling on the cello's glowing amber varnish. The rest of the chamber recedes into warm, even darkness. Within that darkness, a second cellist gradually resolves from the shadows, the candlelight slowly finding the curve of another cello and the folds of a wine-red gown, gaining presence stroke by stroke, until she shares the frame as an equally clear subject. The scene settles into a balanced two-cellist composition, both musicians bowing in a slow, continuous duet, their arms rising and falling in rhythm, framed in a stately medium-wide shot wrapped in candle glow.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A tender watercolor illustration of a silver-haired grandmother in a rose-pink cardigan knitting in a high-backed armchair beside a glowing hearth. Her wooden needles click in a steady, practiced rhythm, a ball of sky-blue yarn slowly turning in the basket at her feet, firelight blooming in warm washes across her gentle face. The cozy room holds shelves of faded books and a plump teapot on a side table, painted in soft amber tones. In the lamplit corner, a second grandmother gradually becomes distinct from the muted background wash, her lilac shawl and needles gaining color and definition as she settles into the opposite armchair, until she shares the frame as an equally clear subject. The scene settles into a balanced two-knitter composition, both women knitting in a continuous, clicking rhythm, framed in a warm, close domestic shot.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A graceful cinematic shot of a figure skater in a wine-red coat and white gloves gliding across a vast frozen lake at dusk, carving long, looping arcs into the pale blue ice. Her blades leave fine white curves that trail behind her steady, flowing strokes, breath drifting in soft clouds. The lake stretches smooth and identical to a treelined horizon silhouetted against a lavender-and-peach sky. From the far end of the ice, a second skater begins as a small dark glide of motion, growing steadily in size and clarity as he sweeps nearer stroke by stroke, until he shares the frame as an equally clear subject circling in tandem. The scene settles into a balanced two-skater composition, both tracing continuous, interweaving loops across the ice, filmed in a wide, slowly drifting shot in cold pastel light.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A child in a cozy winter outfit, blowing gently on a steaming mug of hot cocoa to cool it down. The child has rosy cheeks and warm mittens, with a look of anticipation and slight concentration on their face. They stand in a living room with a fireplace, the warm glow casting a soft, golden light on the scene. A few books and toys are scattered around, adding to the cozy atmosphere. The background shows a blurred view of a fireplace mantel with a wooden clock and some ornaments. The photo has a warm, homey feel. A close-up shot from a slightly lower angle.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A vibrant tropical fish swimming gracefully among colorful coral reefs in a clear, turquoise ocean. The fish has bright blue and yellow scales with a small, distinctive orange spot on its side, its fins moving fluidly. The coral reefs are alive with a variety of marine life, including small schools of colorful fish and sea turtles gliding by. The water is crystal clear, allowing for a view of the sandy ocean floor below. The reef itself is adorned with a mix of hard and soft corals in shades of red, orange, and green. The photo captures the fish from a slightly elevated angle, emphasizing its lively movements and the vivid colors of its surroundings. A close-up shot with dynamic movement.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A nighttime scene from a vintage film-style photograph, depicting a giant, otherworldly creature slowly walking down a desolate, rundown city street. Only one dim streetlamp casts flickering shadows, illuminating the creature's massive, imposing form. Its skin is rough and covered in peculiar growths, with glowing eyes that reflect the dim light. The creature's steps echo in the empty alleyways, creating a sense of eerie quiet. The background features crumbling buildings, broken windows, and trash-strewn sidewalks. The photo has a grainy texture and a muted color palette, capturing the haunting atmosphere of the scene. A medium shot with a slight tilt to the camera, emphasizing the creature's movement and presence.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A realistic photograph in a gritty urban style of a white and orange tabby alley cat dashing across a narrow back street alley during a heavy downpour. The cat is drenched, its fur matted and slick, and it looks determinedly for shelter. Its green eyes are wide and alert, focused intently on finding a safe place. The background is blurred, revealing a dimly lit alley with wet cobblestones and a few dilapidated buildings. The photo has a sharp focus on the cat, capturing its natural movements and the dynamic environment. A medium shot from a low angle, emphasizing the cat's urgency and the wet, urban setting.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A single white sheep bending down to drink water from a calm river. The sheep has fluffy wool, long curved horns, and soft brown eyes. It is positioned near the riverbank, with its front legs partially submerged in the clear water. The river flows gently, reflecting the surrounding greenery and blue sky. The background shows lush grass and trees along the riverbank, creating a serene pastoral landscape. The sheep's body is slightly tilted as it bends down to drink, emphasizing a natural and tranquil motion. Medium close-up shot focusing on the sheep and the river.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A Samoyed and a Golden Retriever dog are playfully romping through a futuristic neon city at night. The neon lights emitted from the nearby buildings glistens off of their fur.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

An extreme close-up of a gray-haired man with a beard in his 60s, he is deep in thought pondering the history of the universe as he sits at a cafe in Paris, his eyes focus on people offscreen as they walk as he sits mostly motionless, he is dressed in a wool coat suit coat with a button-down shirt, he wears a brown beret and glasses and has a very professorial appearance, and the end he offers a subtle closed-mouth smile as if he found the answer to the mystery of life, the lighting is very cinematic with the golden light and the Parisian streets and city in the background, depth of field, cinematic 35mm film.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A llama sits in a cozy reading nook, surrounded by plush pillows and soft blankets. Warm, golden lighting from a floor lamp creates a welcoming atmosphere. The llama reads a picture book aloud, using expressive voices for the characters. The camera captures the llama's animated face and the illustrations in the book.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A charming 3D animated scene of a small brass-and-copper robot with round aquamarine eyes polishing the enormous fresnel lens inside a lighthouse lamp room. It glides slowly around the glittering glass on a circular rail, sweeping a soft cloth in wide, even arcs that leave the prisms gleaming. Warm light refracts through the lens in shifting rainbow slivers that drift across its polished shell. Beyond the encircling windowpanes, a uniform veil of pale sea mist glows softly, holding the little room in its own quiet world. The robot's antenna bobs contentedly with each pass of the cloth. A slow orbiting shot circles with it around the great lens, savoring the looping, careful ritual.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A 35mm film photograph style scene of a silver-haired man in loose slate-grey cotton practicing tai chi on the smooth flagstones before a long vermilion temple wall. His arms sweep through slow, circular forms, weight shifting evenly from foot to foot in an unbroken, meditative flow. The wall runs on identically in both directions, its red paint softly weathered, and thin morning light lays a gentle sheen across the stones. Faint film grain and muted warm tones give the scene a timeless calm, a single sparrow flickering past far behind as an incidental speck. His breathing and movement keep one continuous tempo. The camera holds a steady lateral medium shot, drifting almost imperceptibly with his flowing forms.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A delicate watercolor illustration of a brown hare loping easily across an endless moor of blooming purple heather. Its long black-tipped ears sway with each springy stride, powerful hind legs folding and stretching in a light, even rhythm. Loose washes of violet and umber roll away in identical low swells to a soft pearly horizon, and dry grass stems flick past its flanks in fine brushline strokes. The overcast light is silvery and even, letting soft blooms of pigment bleed into the paper around the hare's warm fur. Its unhurried lope carries on without pause across the open ground. A low tracking shot follows at the hare's shoulder height, keeping its bright amber eye and flowing stride in focus.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A cozy claymation scene of a lilac clay octopus seated at a tiny upright piano in a snug parlor, all eight arms rippling across the keys in a rolling, continuous melody. Its big pearl-button eyes half close with pleasure as the arms rise and fall in overlapping waves, suction cups pressing the keys with soft, precise taps. The parlor wall behind repeats a hand-painted stripe of rose and cream wallpaper, and a small gramophone horn and potted fern sit in warm lamplight. Gentle thumbprints texture every clay surface. The flowing arpeggios loop on without pause, the octopus swaying gently on its stool. A fixed, slightly low medium shot frames performer and piano, delighting in the rippling eight-armed rhythm.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A vibrant cel-animated scene of an antique carousel turning slowly in an empty fairground, its painted horses rising and falling on twisted brass poles in a gentle, endless wave. The canopy is striped in ivory and ruby, hung with rows of warm golden bulbs that glow steadily against a calm dusk-blue sky. Hand-painted mirrors on the central column catch soft flashes of light with every rotation, and faint music-box notes seem to hang in the air. The surrounding fairground lies quiet and uniform, a smooth expanse of pale trodden grass fading into painted haze. The carousel keeps its slow, dreamlike spin without pause. A fixed wide shot holds the whole glowing wheel centered, letting its looping motion soothe the frame.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A naturalistic wildlife photography shot of a massive American bison plodding through steady, light snowfall across an open sagebrush plain. Frost clings to the dense chocolate wool of its hump and beard, and each deliberate step swings its great lowered head gently from side to side, breath pluming in rhythmic clouds. The plain stretches identically in every direction, low silver-green shrubs dusted white beneath a flat pearl-grey sky. Snowflakes drift at a constant, gentle rate, gathering on the animal's broad back as it walks. Its slow, unstoppable pace never varies. A long telephoto tracking shot keeps the bison large and sharp against the softly blurred plain, honoring its patient winter march.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

A delightful claymation scene of a wrinkly clay walrus lounging on a small ice floe, strumming a tiny ukulele with one broad flipper in a mellow, looping rhythm. Its magnificent yarn-textured whiskers bounce as its head sways, long ivory tusks gleaming, while the other flipper taps the ice in time. The surrounding sea is molded in smooth teal clay waves that rise and fall in a gentle, repeating swell beneath an even periwinkle felt sky. Little sculpted sparkles dot the floe's edge, and thumbprints texture every surface with handmade charm. The unhurried island tune strums on and on. A fixed, warmly lit wide shot frames walrus and floe like a tabletop postcard, swaying to the beat.

Baseline + Deep Forcing
Baseline + MemRoPE
Baseline + Rolling Sink
Baseline + Ours (RECAP-Forcing)
Prompt

An enchanting macro shot of a swallowtail butterfly perched on a spiky purple thistle bloom, fanning its wings open and closed in a slow, hypnotic rhythm while its proboscis sips in delicate pulses. Sunlight glows through the wings' stained-glass panels of cream, black, and blue, and fine hairs on the thistle sparkle at the frame's edge. The background dissolves into a uniform dreamy bokeh of mauve and gold. Within that soft blur, a second swallowtail gradually sharpens from a pale fluttering shape as it settles on a neighboring bloom, growing steadily in size and clarity, until it shares the frame as an equally clear subject fanning in counterpoint. The scene settles into a balanced two-butterfly composition, both fanning and sipping in a continuous, gentle rhythm, held in a razor-focused close-up.

1 / 32

06Minute-Long Generations

The full one-minute RECAP-Forcing rollouts of the same 32 prompts, uninterrupted. Click the caption to read the full prompt.

07Supplementary Analysis

Implementation details and extra studies. Each item opens to its full analysis.

AImplementation Details No learned parameters, no fine-tuning, one setting for every backbone.

The primary baseline is Self-Forcing, a DMD-distilled causal video diffusion model built on Wan2.1-T2V-1.3B. It generates 832×480 video in 3-frame blocks with a sliding KV cache of 6 latent frames (30×52 = 1,560 tokens each). A 60-second video is 240 latent frames, or 80 blocks. The other backbones use their public checkpoints and inference code unchanged, with RECAP-Forcing attached as the same inference hook.

After every block, per-pixel novelty is computed with RAFT-small at half resolution (12 refinement iterations) and max-pooled onto the patch grid. Refills are incremental — only newly admitted positions are copied, and only activations of still-novel tokens are kept — so per-block cost is constant. The hyperparameter table in Section 02 (Technical details) is shared across all backbones, lengths and prompt sets.

BLength Generalization Dynamic Degree is higher at every length from 5 to 60 s (+10.1 at 60 s); consistency stays comparable.
Length sweep
Length sweep (Baseline vs RECAP-Forcing, 5–60 s; error bars show ±std over 5 seeds). (a) Higher Dynamic Degree across all lengths. (b) Subject Consistency remains comparable to the baseline.

Evaluating durations from 5 to 60 seconds, RECAP-Forcing consistently achieves higher Dynamic Degree, with the gain reaching +10.1 points at 60 s, while Subject Consistency remains comparable to the baseline — strong dynamics in longer videos without sacrificing consistency.

CAppearance Memory and Identity Preservation Identity holds over 5 minutes (DINOv2 similarity to frame 0: 0.81 vs 0.61) and across a 7.5 s absence.

Does the stored appearance stay useful over long horizons? Three increasingly demanding settings: identity under continuous visibility, retention over much longer rollouts, and reappearance after a long absence.

C.1Long-horizon subject identity

20 prompts keep a distinctive subject on screen for the full minute. We compare an early window (2–8 s) with the final 10 s using DINOv2 similarity, on the full frame and on the subject crop (Qwen2.5-VL detector, 97% detection rate). RECAP-Forcing clearly improves both.

20 prompts × 5 seeds, 60 s; DINOv2 similarity between an early window and the final 10 s, on the full frame and on the detected subject crop.
ConfigurationFrame Sim.WinsCrop Sim.Wins
Baseline0.745–0.642–
+ Ours0.85076%0.69664%

C.2Identity retention over extended horizons

Identity retention is the DINOv2 similarity of each frame to frame 0, on the same prompts and frame times for both methods. Over five minutes the recency baseline decays to 0.61, while RECAP-Forcing stops decaying after the first minute and holds 0.81 to the end.

20 prompts × 5 seeds; DINOv2 cosine to frame 0; higher holds identity; "mid"/"end" are 1/5 min for the five-minute rows and 30/60 s for the 60-second rows.
HorizonMethodstartmidend
5 minBaseline1.000.630.61
+ Ours1.000.810.81
60 sBaseline1.000.700.62
+ Ours1.000.870.88

C.3Controlled reappearance

A subject leaves for 7.5 s — five times the 1.5 s attention window — and must return with the same identity. Prompts alone cannot stage this (current backbones cannot follow temporal instructions), so each video is one continuous rollout of four prompt segments: empty scene, subject, empty again, subject re-conditioned. The sliding window is cleared at every boundary while sink and bank persist, so the bank is the only channel that can carry the subject’s appearance across the gap. The base model re-invents the subject from the prompt; RECAP-Forcing renders it against its stored first-appearance states.

20 scenarios × 5 seeds; subject enters, leaves for a 7.5 s gap (5× the attention window) with the sliding window cleared at every shot boundary, then is re-conditioned; DINOv2 similarity between the first-appearance and return windows.
ConfigurationFrame Sim.WinsCrop Sim.Wins
Baseline0.699–0.648–
+ Ours0.75273%0.68866%
DMemory Bank Analysis Flow-novelty admission beats random, recent and semantically diverse selection, and retention is time-neutral.

D.1Admission rule: what should enter the bank?

We vary only the admission rule (128 prompts × 5 seeds). Random and recent history barely help — Dynamic Degree 62.4 and 62.0 versus 61.3, and random selection even lowers Total. DINOv2 Diversity lifts Total to 79.4 but does not recover motion (60.6). Flow Novelty raises Dynamic Degree to 71.3 (+10.0) and gets the best Total (79.7): what enters the bank matters, and appearance novelty is the signal.

Bank admission rules under the main protocol: 128 prompts × 5 seeds, on Infinite-Forcing.
Admission ruleS.C.B.C.M.S.D.D.A.Q.I.Q.Q.Sem.Total
No Bank (Baseline)97.495.998.461.355.868.882.262.978.3
Random History97.196.998.562.456.469.282.458.777.7
Recent History97.496.198.462.056.069.082.364.278.7
DINOv2 Diversity97.696.198.660.657.669.482.965.679.4
Flow Novelty (Ours)97.496.098.471.357.568.883.365.479.7

Rules. Random samples uniformly from past content; Recent keeps the latest entries; DINOv2 Diversity picks K-means representatives in DINOv2 feature space; Flow Novelty admits by optical-flow novelty (ours).

D.2Bank retention behavior

Over 220 rollouts, the bank closely matches the top-K of all items observed so far, ranked by their frozen scores (final-state overlap 0.99; the residual gap is score ties), so eviction removes almost only below-cutoff admissions. The source-frame ages of the final entries are spread roughly uniformly over the rollout, with a mean near the video midpoint — survival is set by novelty, not recency.

EInference Cost A constant ~1.6× overhead that does not grow with video length.

Wall-clock time for 1-, 2- and 5-minute videos on a single GH200 GPU. A naive bank refill would scale quadratically with duration and need hundreds of GB of cached history; instead we store only novel-token activations and refill slots incrementally, reproducing the naive path bit-identically at constant per-block cost.

Generation time for a single 832×480 video on an 80 GB GH200 GPU.
LengthFrames (blocks)BaselineOursOverhead
1 min240 (80)185 s288 s1.56×
2 min480 (160)268 s450 s1.68×
5 min1200 (400)1435 s2239 s1.56×
FIs the Gain Good Motion? The extra motion is not camera drift: camera flow falls (1.20 → 0.90) while object flow is largely preserved.

A natural concern is that Dynamic Degree rises through cheap global camera panning. It does the opposite. For each video we compute dense RAFT flow on frame pairs sampled every ~0.5 s, fit a global homography with RANSAC, and split the flow into a camera component and a residual object component. RECAP-Forcing reduces camera motion (1.20 → 0.90) while largely preserving object motion (0.71 → 0.66), raising the fraction of motion attributable to objects from 37% to 42%.

Dynamic Degree counts the videos whose motion exceeds a genuine-motion threshold, not mean flow, so it can rise while mean flow falls: near-frozen clips move above the threshold while runaway-drift outliers are suppressed (per-video total-flow deviation ±2.9 → ±1.9; drift tail 37 → 24 clips).

128 prompts; RAFT flow split by a RANSAC global homography; magnitudes in px per frame pair; drift tail counts videos whose mean total flow exceeds 2× the config median.
ConfigCamera FlowObject FlowObject FractionTotal FlowDrift Tail
Baseline1.2040.7100.3801.91437
+ Ours0.8970.6590.4081.55624

BibTeX

@article{recapforcing2026,
  title={RECAP-Forcing: REtaining Content APpearances for Long Video Generation},
  author={Xu, Haiyang and Ding, Zheng and Tu, Zhuowen},
  journal={arXiv preprint arXiv:2608.26671},
  year={2026}
}