柔軟なビデオ拡散モデル（3 分読了）

ICML 2026 · Spotlight

One model, two generation regimes

Xinyin Ma1,2

Julius Berner2

Chao Liu2

Arash Vahdat2

Weili Nie2,†

Xinchao Wang1,†

1NUS

2NVIDIA

† Equal advising

A detailed illustration in a realistic style depicting a raccoon wearing a classic detective's hat, holding a magnifying glass and a notebook. The raccoon has expressive brown fur, large round eyes, and a curious expression, as it examines a small, mysterious object. It stands on two legs, slightly tilted to one side, looking intently at something in front of it. The background features a cluttered detective's office with old books, maps, and various tools scattered around, giving the scene a cozy, vintage feel. The room is dimly lit, with a soft glow coming from a nearby lamp. A medium shot capturing the raccoon in action.

A dramatic close-up shot of a man's face during a stormy sea voyage, capturing his intense fear and desperation. His face is illuminated by the flickering light of the ship's lanterns, casting shadows that accentuate his worried expression. Dark, stormy clouds loom overhead, and waves crash against the ship, adding to the chaotic environment. The man's eyes are wide with panic, and his lips are tightly pressed together. His tousled hair and slightly wet clothes add to the sense of urgency. The background is a blur of movement, with the ship's deck and the turbulent sea visible in the periphery. The photo has a gritty, realistic texture, emphasizing the raw emotion and struggle of the moment. A dynamic close-up from a slightly tilted angle.

A dynamic action shot of a turtle wearing a sleek racing suit, riding a colorful skateboard down a steep hill. The turtle has a determined expression, its small legs pumping vigorously as it skates with speed and agility. The skateboard wheels spin rapidly, leaving a slight blur in the background. The hillside is lined with tall grass and wildflowers, and the sky is a bright blue with fluffy clouds. The scene captures the turtle's momentum and excitement as it races down the hill, with a slight tilt to the camera angle to enhance the sense of speed and movement.

A vibrant and dynamic illustration in the style of a nature documentary, featuring a dragon-toucan walking gracefully through the vast grasslands of the Serengeti. The dragon-toucan has iridescent green and blue feathers, with a long, curved beak and large, expressive eyes. It strides confidently across the savannah, its wings slightly spread for balance. The background showcases a rich tapestry of African wildlife, with zebras, gazelles, and elephants in the distance. The sun is setting, casting a warm golden glow over the landscape. The camera angle is from a low, ground-level perspective, capturing the dragon-toucan in motion as it moves through the grass.

A dynamic tracking shot in the style of a classic Hollywood film, capturing a steam locomotive chugging through a snowy landscape. The train moves forward with a sense of urgency, the camera following closely behind to highlight the speed and power of the journey. Snowflakes swirl around the train, creating a sense of movement and cold. The scenery changes rapidly, revealing dense forests, winding tracks, and distant mountains covered in snow. The background features blurred snow-covered trees and distant hills, with patches of sunlight breaking through the clouds. The train’s smokestack releases billowing steam, adding to the dramatic effect. A wide-angle lens captures the expansive view, emphasizing the vastness of the snowy wilderness.

A handheld shot following a young child running through a field of tall grass, capturing the spontaneity and playfulness of their movements. The child has curly brown hair and a mischievous smile, arms swinging freely as they sprint across the green expanse. Their small feet kick up bits of grass and dirt, creating a trail behind them. The background features a blurred landscape with rolling hills and scattered wildflowers, bathed in warm sunlight. The photo has a natural, documentary-style quality, emphasizing the dynamic motion and joy of the moment. A dynamic handheld shot from a slightly elevated angle, following the child's energetic run.

A macro shot focusing on the face of a young woman with freckles, her expression intense as she looks intently for something. Her freckles are scattered across her cheeks and nose, adding a playful charm to her face. Her eyes are wide and slightly squinted, peering closely at the object of her search. Her hair is loose, framing her face gently, with strands falling over her forehead. The background is blurred, but you can make out the faint outline of a table or desk where she is searching. The texture of her skin is smooth and细腻，带有淡淡的红润。A close-up shot from a very close angle, capturing the natural and focused expression of the young woman.

A macro shot in realistic style of a man wearing an antique diving helmet with dark glass and a jetpack, standing on a molten lava surface. He strides confidently, his body slightly bent forward, with a determined expression. Behind him, a majestic dragon soars through the sky, its wings spreading wide and scales glistening in the flickering light. The background is a dramatic landscape with smoldering volcanic peaks and swirling clouds, creating a sense of otherworldly danger and adventure. The man’s muscles are flexed, and his arms are outstretched as he walks, adding a dynamic quality to the scene. A medium shot with a slight tilt upwards, emphasizing both the man and the flying dragon.

A miniature 3D render in an octane engine style depicting adorable wool and felt monsters dancing together in a dreamy, bokeh-filled setting. These soft, cuddly creatures, with big expressive eyes and fluffy bodies, are illuminated by gentle, diffused lighting that casts a warm, ethereal glow. The background features a soft, hazy backdrop with a dreamy bokeh effect, adding a cinematic quality to the scene. The monsters are shown from various angles, capturing their playful movements and expressions, creating a charming and enchanting atmosphere. A medium shot with a dynamic camera angle, highlighting the natural and joyful dance of these woolen monsters.

A panoramic view sweeping right across a field of tall grass swaying gently in the wind, with a setting sun casting a warm golden glow in the background. The grass blades catch the last rays of sunlight, creating a shimmering effect. The horizon is blurred, highlighting the contrast between the vibrant grass and the soft, fading sky. A low-angle shot capturing the dynamic movement of the grass and the serene beauty of the twilight.

A slow cinematic push-in on an ostrich standing in a 1980s kitchen, the camera gradually zooming in to reveal the bird's curious expression. The kitchen is adorned with vintage appliances and Formica countertops, with a muted color palette of pastel greens and yellows. The ostrich, with its distinctive long neck and feathered plumage, stands confidently, one foot slightly raised. Its large brown eyes peer curiously at the viewer, as if pondering the strange surroundings. The background features blurred details of old newspapers scattered on the floor and a faded floral wallpaper. The lighting is warm and soft, casting gentle shadows. A close-up shot from a slightly lower angle.

A dynamic snowboarding scene in the style of a high-energy action shot, featuring a young snowboarder accelerating down a powdery slope. The snowboarder, with a determined expression, weaves expertly between tall pine trees, their trunks partially obscured by the swirling snow. The snow is pristine and fluffy, with the sun casting soft shadows and highlighting the snowboarder's movements. The background showcases a breathtaking mountain vista, with peaks shrouded in mist and a few distant ski lifts visible. The camera angle captures the snowboarder from a slightly behind-the-action perspective, emphasizing their speed and agility.

A detailed illustration in a realistic style depicting a raccoon wearing a classic detective's hat, holding a magnifying glass and a notebook. The raccoon has expressive brown fur, large round eyes, and a curious expression, as it examines a small, mysterious object. It stands on two legs, slightly tilted to one side, looking intently at something in front of it. The background features a cluttered detective's office with old books, maps, and various tools scattered around, giving the scene a cozy, vintage feel. The room is dimly lit, with a soft glow coming from a nearby lamp. A medium shot capturing the raccoon in action.

A dramatic close-up shot of a man's face during a stormy sea voyage, capturing his intense fear and desperation. His face is illuminated by the flickering light of the ship's lanterns, casting shadows that accentuate his worried expression. Dark, stormy clouds loom overhead, and waves crash against the ship, adding to the chaotic environment. The man's eyes are wide with panic, and his lips are tightly pressed together. His tousled hair and slightly wet clothes add to the sense of urgency. The background is a blur of movement, with the ship's deck and the turbulent sea visible in the periphery. The photo has a gritty, realistic texture, emphasizing the raw emotion and struggle of the moment. A dynamic close-up from a slightly tilted angle.

A dynamic action shot of a turtle wearing a sleek racing suit, riding a colorful skateboard down a steep hill. The turtle has a determined expression, its small legs pumping vigorously as it skates with speed and agility. The skateboard wheels spin rapidly, leaving a slight blur in the background. The hillside is lined with tall grass and wildflowers, and the sky is a bright blue with fluffy clouds. The scene captures the turtle's momentum and excitement as it races down the hill, with a slight tilt to the camera angle to enhance the sense of speed and movement.

A vibrant and dynamic illustration in the style of a nature documentary, featuring a dragon-toucan walking gracefully through the vast grasslands of the Serengeti. The dragon-toucan has iridescent green and blue feathers, with a long, curved beak and large, expressive eyes. It strides confidently across the savannah, its wings slightly spread for balance. The background showcases a rich tapestry of African wildlife, with zebras, gazelles, and elephants in the distance. The sun is setting, casting a warm golden glow over the landscape. The camera angle is from a low, ground-level perspective, capturing the dragon-toucan in motion as it moves through the grass.

A dynamic tracking shot in the style of a classic Hollywood film, capturing a steam locomotive chugging through a snowy landscape. The train moves forward with a sense of urgency, the camera following closely behind to highlight the speed and power of the journey. Snowflakes swirl around the train, creating a sense of movement and cold. The scenery changes rapidly, revealing dense forests, winding tracks, and distant mountains covered in snow. The background features blurred snow-covered trees and distant hills, with patches of sunlight breaking through the clouds. The train’s smokestack releases billowing steam, adding to the dramatic effect. A wide-angle lens captures the expansive view, emphasizing the vastness of the snowy wilderness.

A handheld shot following a young child running through a field of tall grass, capturing the spontaneity and playfulness of their movements. The child has curly brown hair and a mischievous smile, arms swinging freely as they sprint across the green expanse. Their small feet kick up bits of grass and dirt, creating a trail behind them. The background features a blurred landscape with rolling hills and scattered wildflowers, bathed in warm sunlight. The photo has a natural, documentary-style quality, emphasizing the dynamic motion and joy of the moment. A dynamic handheld shot from a slightly elevated angle, following the child's energetic run.

A macro shot focusing on the face of a young woman with freckles, her expression intense as she looks intently for something. Her freckles are scattered across her cheeks and nose, adding a playful charm to her face. Her eyes are wide and slightly squinted, peering closely at the object of her search. Her hair is loose, framing her face gently, with strands falling over her forehead. The background is blurred, but you can make out the faint outline of a table or desk where she is searching. The texture of her skin is smooth and细腻，带有淡淡的红润。A close-up shot from a very close angle, capturing the natural and focused expression of the young woman.

A macro shot in realistic style of a man wearing an antique diving helmet with dark glass and a jetpack, standing on a molten lava surface. He strides confidently, his body slightly bent forward, with a determined expression. Behind him, a majestic dragon soars through the sky, its wings spreading wide and scales glistening in the flickering light. The background is a dramatic landscape with smoldering volcanic peaks and swirling clouds, creating a sense of otherworldly danger and adventure. The man’s muscles are flexed, and his arms are outstretched as he walks, adding a dynamic quality to the scene. A medium shot with a slight tilt upwards, emphasizing both the man and the flying dragon.

A miniature 3D render in an octane engine style depicting adorable wool and felt monsters dancing together in a dreamy, bokeh-filled setting. These soft, cuddly creatures, with big expressive eyes and fluffy bodies, are illuminated by gentle, diffused lighting that casts a warm, ethereal glow. The background features a soft, hazy backdrop with a dreamy bokeh effect, adding a cinematic quality to the scene. The monsters are shown from various angles, capturing their playful movements and expressions, creating a charming and enchanting atmosphere. A medium shot with a dynamic camera angle, highlighting the natural and joyful dance of these woolen monsters.

A panoramic view sweeping right across a field of tall grass swaying gently in the wind, with a setting sun casting a warm golden glow in the background. The grass blades catch the last rays of sunlight, creating a shimmering effect. The horizon is blurred, highlighting the contrast between the vibrant grass and the soft, fading sky. A low-angle shot capturing the dynamic movement of the grass and the serene beauty of the twilight.

A slow cinematic push-in on an ostrich standing in a 1980s kitchen, the camera gradually zooming in to reveal the bird's curious expression. The kitchen is adorned with vintage appliances and Formica countertops, with a muted color palette of pastel greens and yellows. The ostrich, with its distinctive long neck and feathered plumage, stands confidently, one foot slightly raised. Its large brown eyes peer curiously at the viewer, as if pondering the strange surroundings. The background features blurred details of old newspapers scattered on the floor and a faded floral wallpaper. The lighting is warm and soft, casting gentle shadows. A close-up shot from a slightly lower angle.

A dynamic snowboarding scene in the style of a high-energy action shot, featuring a young snowboarder accelerating down a powdery slope. The snowboarder, with a determined expression, weaves expertly between tall pine trees, their trunks partially obscured by the swirling snow. The snow is pristine and fluffy, with the sun casting soft shadows and highlighting the snowboarder's movements. The background showcases a breathtaking mountain vista, with peaks shrouded in mist and a few distant ski lifts visible. The camera angle captures the snowboarder from a slightly behind-the-action perspective, emphasizing their speed and agility.

5s videos generated by Flex-Forcing — better speed with better quality. Hover any clip to pause & read its prompt.

One Model, Bidirectional & Autoregressive

Flex-Forcing unifies the generation spectrum of both autoregressive generation and bidirectional generation.

Drag the slider from one bidirectional chunk to autoregressive small chunks

and watch the same prompt generate under every regime in between.

chunk 21

1 chunk · fully bidirectional

chunk 21

1 chunk · fully bidirectional

Chunk configuration *(shared by both videos)*

21

2111 / 109 / 6 / 67 / 7 / 76 / 5 / 5 / 55 / 4 / 4 / 43 / 3 / 3 / 3 / 3 / 3 / 3

← Bidirectional (large chunk · global planning)

Autoregressive (small chunks · streaming) →

Pick a prompt for each panel, then drag the shared slider. Larger chunks plan more globally

(closer to bidirectional diffusion); smaller chunks generate more autoregressively — the same

model covers both ends, on every prompt.

Abstract

Recent progress in large-scale generative models has substantially advanced video generation,

yet existing methods remain constrained by a rigid inference paradigm. Bidirectional diffusion

models excel at global coherence and visual fidelity but suffer from slow inference, while

autoregressive models offer efficient and streaming generation at the cost of long-range

consistency and exposure bias. We introduce Flex-Forcing, a unified training and

inference framework that enables a video diffusion model to seamlessly operate under both

bidirectional and autoregressive generation regimes. The core idea is a **flexible chunking

mechanism** jointly defined over the temporal dimension and denoising steps. This design

allows the model to (1) flexibly chunk to support different device budgets, (2) perform

bidirectional inference across chunks for global structure planning while generating frames

autoregressively within each chunk for efficient, fine-grained synthesis, and (3) perform

any-order, any-timestep autoregressive generation without the strict causal constraint.

Extensive experiments on multiple benchmarks demonstrate that Flex-Forcing achieves

consistently better video quality and long-video stability than strong baselines while

offering faster inference.

Method

A single flexible-chunking mechanism, defined jointly over video frames and denoising

timesteps, unifies autoregressive, bidirectional, and hybrid generation in one model.

Flexible chunking over frames and timesteps. Fine-grained chunks recover

autoregressive generation (efficient, lower quality); a single coarse chunk recovers

bidirectional generation (high quality, less efficient). Flex-Forcing interpolates between

them: chunking over frames sets the granularity, while chunking over timesteps lets the

granularity evolve along the denoising trajectory (high → low noise).

Speed–Quality Trade-off

Across configurations, Flex-Forcing dominates Self-Forcing on both FPS and VBench score.

Brute-force search over chunk configurations (5s video split into three chunks over 21

latent frames). The Pareto-optimal hybrid [9, 6, 6], a coarse-to-fine pattern, beats

uniform [7, 7, 7] and Self-Forcing [3, 3, …] on both FPS and VBench, and can even

edge out fully bidirectional inference.

With different configurations, our model achieves superior performance over Self-Forcing on

both FPS (frames per second) and VBench score.

Qualitative Comparison

Flex-Forcing (ours) vs. Self-Forcing under same prompts.

"A captivating underwater video of a graceful jellyfish drifting through crystal-clear water."

Self-Forcing

Ours

"A cinematic zoom-out from a flickering candle flame, gradually revealing a dimly lit room adorned with antiques."

Self-Forcing

Ours

"A close-up shot of a person skillfully shaping pottery from clay that changes colors with each touch."

Self-Forcing

Ours

"A delicate layer of morning frost slowly melting off a single rose petal, the tiny droplets glistening in the light."

Self-Forcing

Ours

"A dramatic action shot in the style of a superhero film, depicting a man pushing away a massive stone."

Self-Forcing

Ours

"A realistic sketch-style illustration depicting a person, likely a young artist with a focused expression."

Self-Forcing

Ours

Long-Range Consistency and Dynamic Degree (30s Videos)

Autoregressive comparison over an extended horizon — Flex-Forcing maintains stability over time

while preserving a higher dynamic degree.

"A cyberpunk-style illustration depicting a lone robot navigating a neon-lit cityscape."

Infinity-RoPE

Ours

"A vibrant, lively vlog-style video of a corgi energetically exploring tropical Maui."

Infinity-RoPE

Ours

"A vibrant concert stage scene in the style of a music video, featuring a woman singing in the spotlight."

Infinity-RoPE

Ours

"A dramatic, epic fantasy-style illustration depicting a towering giant shaped like a man."

Infinity-RoPE

Ours

Videos are played at 3× speed.

Applications: Any-Order, Any-Timestep Editing

We restrict editing to low-level refinement timesteps while keeping high-level planning fixed,

decoupling global structure from local details for far better consistency than Self-Forcing, where

small edits propagate across frames. Attending to both past and future tokens, our model edits any

chunk regardless of its temporal order.

No Editing

Appearance Stylization

wear an oversized sunglasses

New Object

a cat runs across the garden path

Environment Interaction

a dust rises beneath its paws

New Action

leaps upward in a playful jump

Self-Forcing

Flex-Forcing (Ours)