Pika H3 Flex Technical Deep Dive
Post-Training H3 to Six-NFE Omni-Generator
October 29, 2026Research
Today, we’re excited to launch Pika H3 Flex, an open-source omni video model family that ranks #1 among open-source text-to-video models and delivers real-time video generation. Post-trained by the Pika team on MiniMax H3, Pika H3 Flex combines advanced fine-tuning and reward-based optimization on an internally curated dataset selected for aesthetic quality. It delivers state-of-the-art results across visual and audio quality metrics and human preference evaluations, outperforming leading competitors by up to 91% while remaining competitively priced (see Table 1).
In our human evaluation study, Pika H3 Flex ranks #1 for human anatomy and artistic and aesthetic styles. Pika H3 Flex also supports a range of downstream editing tasks, including but not limited to style transfer, movement transfer, object replacement, character swap and targeted edits, opening up new possibilities for interactive content creation and rapid creative iteration.
Pika H3 Flex supports output resolutions from 480p to 4K and offers five separate, editable audio stems—dialogue, score, SFX, foley, and ambience—alongside the generated soundtrack, a first for a video model. Pika H3 Flex is available through our API, ComfyUI, and Diffusers, with model weights available on Hugging Face and code on GitHub. Pika also offers customized training services. To learn more, contact enterprise@pika.art.
Table 1. Pricing and capabilities of text/image-to-video models. Cost-cell colors progress from green (lowest) to red (highest). Real-time refers to generation latency below the output clip duration. 4K denotes output resolution availability.
Model | Starting cost (USD/output-s) | Real-time | 4K output | Open weights | Audio stems |
Pika H3 Flex | $0.01 | ✅ | ✅ | ✅ | ✅ |
HeyGen Speed | $0.01 | Not disclosed | ❌ | ❌ | ❌ |
HeyGen Quality | $0.01 | Not disclosed | ❌ | ❌ | ❌ |
HeyGen Ultra | $0.01 | Not disclosed | ❌ | ❌ | ❌ |
MiniMax H3 | $0.08 | ❌ | ❌ | ✅ | ❌ |
Fal H3 Max | $0.0325 | ✅ | ❌ | ❌ | ❌ |
Fal H3 Turbo | $0.01625 | ✅ | ❌ | ❌ | ❌ |
Seedance 2.5 | $0.103 | ❌ | ❌ | ❌ | ❌ |
FLUX 3 Video | $0.06 | ❌ | ✅ | ❌ | ❌ |
Gemini Omni 1.1 Flash | $0.101 | ❌ | ✅ | ❌ | ❌ |
Wan 3.0 | $0.03 | ❌ | ❌ | ❌ | ❌ |
Table 2. Reference-to-video pricing and capabilities. Pika H3 Flex offers competitive pricing while supporting all listed applications. ‡ Gemini Omni 1.1 Flash supports image and limited video references, but does not support uploaded audio references; audio embedded in reference videos is ignored (API documentation). * Seedance 2.5 and Wan3.0 adds video reference charge.
Model | Starting cost (USD/s) | 4K output | Open weights | Multiple image, audio, video references | Uploaded-audio lip sync/redub |
Pika H3 Flex | 0.03 | ✅ | ✅ | ✅ | ✅ |
HeyGen Speed | 0.035 | ❌ | ❌ | ✅ | ✅ |
HeyGen Quality | 0.035 | ❌ | ❌ | ✅ | ✅ |
HeyGen Ultra | 0.035 | ❌ | ❌ | ✅ | ✅ |
MiniMax H3 | 0.08 | ❌ | ✅ | ✅ | ✅ |
Fal H3 Max | 0.05 | ❌ | ❌ | ✅ | ✅ |
Seedance 2.5 | 0.103 • 0.0615* | ❌ | ❌ | ✅ | ✅ |
Gemini Omni 1.1 Flash | 0.101 | ✅ | ❌ | Limited; no uploaded audio‡ | ❌ |
Wan 3.0 | 0.03 • 0.03* | ❌ | ❌ | ✅ | ✅ |
Showcases
Our showcases demonstrate Pika H3 Flex’s capabilities across text-, image-, and reference-conditioned video generation, including controlled camera motion, multi-shot generation, and composition from multiple references. Our model further demonstrates strong capabilities in downstream applications, including but not limited to motion transfer, object replacement, character swapping, and selective video editing. Its low latency and competitive pricing support rapid iteration from initial concept to high-resolution output, while its open-source availability enables flexible integration into creative workflows and experimentation with new applications.
Complex camera choreography
Table-tennis rally
Five scripted camera moves in a 15-second take: overhead, crane, whip pans, a 180° orbit, and a low dolly.
Input
Generated output
Full prompt
Image 1 is the exact two players, outfits, frog mask, table-tennis table and drained pool: keep them exactly as in Image 1.
A fast table-tennis rally, shot with exactly these camera moves:
00:00–00:03 — top-down from directly overhead, as in Image 1, as the rally begins and the ball flicks back and forth.
00:03–00:06 — a smooth crane straight down from overhead to table height, ending right beside the net as the ball zips across.
00:06–00:09 — fast whip pans left and right following the ball across the net, from one player to the other.
00:09–00:12 — a 180-degree orbit around the frog-masked player as she leaps and smashes the ball.
00:12–00:15 — a low-angle dolly backward along the pool floor as the woman in the red tracksuit dives and misses; the frog raises both arms and does a little victory dance; the camera settles and holds.
The players, table and pool stay consistent and in the same place throughout. Photoreal 35mm film look, true colour, light natural grain, matching Image 1. SPEED IS CRITICAL: everything moves at normal real-time speed, like 24fps footage played back at 24fps - natural motion, NO slow motion.
Sound: the quick tock-tock of the ball, sneakers squeaking on concrete, the smash, a thud as she dives, echo in the empty pool, the frog's muffled cheer. Music: a funky, upbeat drum-and-bass groove.
Balcony stretches
Three independently moving subjects with a continuous sequence of orbit, push-in, and vertical camera moves.
Input
Text prompt · excerpt
A white modernist tower with three stacked cantilevered balconies fills the right half of the frame against a pale-blue sky, in a minimal, pastel 1970s film aesthetic.
Generated output
Full prompt
A tall white modernist apartment tower against a soft pale-blue sky with faint streaky cirrus cloud. The building sits on the right half of the frame; the left half is open sky. Smooth off-white render facade with small square windows in thick white frames. Three identical cantilevered concrete balconies are stacked vertically on the building's corner, each a thick white slab with an angled, wedge-shaped underside and a thin white metal railing of vertical balusters. Crisp, graphic, minimal, soft 1970s film palette: chalky whites, powder blue, muted pastels.
Three people are doing a slow, calm morning stretch routine, one per balcony. They are always gently in motion, never frozen, but every movement is slow, smooth and controlled, like yoga or tai chi: each stretch takes three to four seconds to flow into, eases gently at the end, and flows unhurried into the next. No fast, jerky, snapping or flailing movements. Graceful, deliberate, meditative.
Top balcony: a young woman with dark hair pulled back, in a lilac fluffy knit jumper and matching lilac wide-leg knit trousers. She slowly reaches both arms high overhead, then gradually leans her upper body into a long side bend to the left, eases back to centre, and slowly leans to the right.
Middle balcony: a man with short dark hair in a rust-red tracksuit. He hangs in a deep forward fold with palms on the floor, then slowly rolls up through his spine to standing, and lifts his arms in one wide, slow circle overhead.
Bottom balcony: an older woman with short silver-grey hair in a cobalt blue short-sleeved polo and loose cream trousers. She holds a wide warrior lunge with arms stretched out to each side, slowly shifts her weight from one bent knee to the other, and gently sweeps one arm up and over her head into a side bend.
Each person moves at their own unhurried rhythm, slightly out of sync with each other, with full-body range of motion: knitwear and trousers stretching and settling, real balance and weight.
Camera, one continuous shot with no cuts:
0-3s: Wide establishing frame. The camera begins a slow, smooth orbit around the corner of the tower, arcing about 40 degrees, with strong parallax between the balconies, building face and sky.
3-5s: Still moving, the camera pushes in hard on the middle balcony until the man in the red tracksuit fills the frame alone, waist-up, mid-stretch, railing in the foreground.
5-7s: A steady vertical boom up the face of the building, past the concrete underside of the slab above, revealing the lilac woman mid-side-bend.
7-10s: A fast, controlled boom straight down past the middle balcony to the bottom one, revealing the older woman shifting through her lunge.
10-13s: Smooth pull back and reverse orbit, landing on the opening wide composition, all three still actively stretching.
Soft diffused daylight, gentle soft shadows under each slab, no hard sun. 35mm film look with fine grain. Stabilized crane/drone motion, never jittery.
Sound: natural diegetic location sound only — light wind gusting past the building, distant city hum, faint birds, soft fabric rustle. No music, no score, no singing, no melody, no rhythm track. No speech, no voices, no talking, no dialogue, no murmuring, no vocal sounds of any kind.
Farm tracking shot
A rapid lateral truck transitions into a forward dolly, followed by a close-up cut.
Input
Text prompt · excerpt
A vegetable farm with parallel raised beds, bamboo A-frame trellises, young tomato and bean plants, and terracotta pots under a soft overcast sky.
Generated output
Full prompt
A large working vegetable farm under a soft overcast sky, pale grey-white cloud with patches of soft blue. Endless parallel rows of raised beds of dark brown tilled soil, each lined with tall bamboo cane A-frame teepees tied with twine, young leafy green tomato and bean plants at their base, terracotta pots along low weathered timber bed edges.
The video opens with NO person visible. The first four seconds show only rows of plants and bamboo.
Shot 1, one continuous camera move:
0-4s: The camera is low, at waist height, and trucks sideways to the right at high speed, travelling perpendicular to the rows. Row after row after row of bamboo teepees and green plants rushes across the frame from right to left, each row flashing past one after another: canes, plants, gap of soil, canes, plants, gap of soil. Strong motion blur on the foreground canes, the horizon steady. Like a camera car racing along the end of the field.
4-5s: The camera brakes hard and stops dead, perfectly lined up looking straight down the length of one single row, the row's soil path running to a vanishing point in the centre of frame. A tiny dark figure is just visible far away at the end of the row. A brief still beat.
5-8s: The camera launches forward and dollies fast and low straight down that row between the bamboo canes, the canes streaming past either side. It decelerates as the figure grows into a woman kneeling in the soil and stops in a medium-wide side-on shot of her: early 30s, pale skin, dark brown hair in a low loose chignon, wearing a dramatic Victorian-silhouette ballgown made of crinkled glossy black plastic tarpaulin, with a deep V neckline, enormous puffed short sleeves and a huge gathered skirt pooled across the soil behind her. She leans forward over the bed, planting a small seedling with a steel hand trowel.
Hard cut.
Shot 2:
8-12s: Tight close-up at soil level on her hands. She presses dark damp soil firmly around the base of the seedling and pats it down twice, sets the trowel aside, then brushes her palms together to dust them off. Crumbs of soil fall from her fingers. The glossy crinkled black plastic of her skirt fills the background.
The plastic gown has real stiffness: it crackles, creases and catches soft highlights. Soil is clumpy and leaves dirt on her fingers. Soft diffused overcast light, no hard sun or hard-edged shadows, but rich contrast and dimension. Surreal fashion-editorial tone played straight. Photographic natural colour, subtle film grain.
Sound: natural diegetic location sound only — rushing wind and whoosh of canes flashing past, then quiet wind through leaves, the crinkle of the plastic gown, trowel scraping soil, soft pats, dry hands brushing together. No music, no score, no singing, no melody, no rhythm track.
Multi-shot storytelling
Wedding nibbler
A four-shot story directed from one image, with camera changes, dialogue, and synchronized sound.
Input
Generated output
Full prompt
Start exactly on the first frame: a fluffy white lop-eared rabbit nibbling the side of a tall three-tier buttercream cake on a glass cake stand, on a cream silk-draped plinth in front of a dusty-rose velvet curtain. Soft, warm, filmic light, true colour, fine 35mm grain.
[Shot 1, 00:00–00:04] A slow push-in on the rabbit as it keeps nibbling a hole in the bottom tier, cake crumbs falling onto the cloth, one floppy ear twitching.
[Shot 2, at 00:04 cut] Extreme macro close-up: the rabbit's twitching pink nose and whiskers dusted with buttercream as it chews fast.
[Shot 3, at 00:07 cut] Medium-wide from the side: the velvet curtain parts behind the cake and a pastry chef in her fifties, in a white chef's jacket, steps through holding a piping bag, freezes, and says in a shocked, quiet voice: "That was for the wedding." The rabbit stops chewing and turns its head to look at her.
[Shot 4, at 00:11 cut] Low angle at tabletop height, a slow dolly to the side: the rabbit looks straight into the lens, slowly resumes chewing, then hops once behind the cake out of sight, leaving a perfect bite-shaped hole.
Keep every person, animal, object, outfit and the location exactly as in the first frame; nobody transforms or morphs into anyone else. Real people speak with natural lip-sync. SPEED IS CRITICAL: everything moves at normal real-time speed, like 24fps footage played back at 24fps - natural motion, NO slow motion.
Sound: soft crunchy nibbling, crumbs pattering on silk, the heavy swish of velvet, the chef's one line, a tiny thump as the rabbit hops. Music: a delicate, slightly mischievous harpsichord waltz.
Birthday-cake sitcom
A five-shot sitcom sequence directed from one image, with dialogue, audience reactions, and sound design.
Input
Generated output
Full prompt
A 1990s American multi-camera sitcom scene exactly like Image 1: bright, flat studio lighting on a living-room set, slight videotape softness, the same woman in red. A live studio audience laughs throughout. Hard cuts between different camera angles.
[Shot 1] 0-3 s: wide shot. She carries a pretty birthday cake into the living room and sets it on the coffee table; a small dog in a party hat sits by the sofa, staring at the cake. (Audience: "Aww.")
[Shot 2] 3-5 s: medium shot. A sudden knock at the door. She points at the dog: "Wait here. Don't touch it." and walks out of frame to answer the door.
[Shot 3] 5-9 s: close-up, locked on the coffee table: the dog jumps up and wolfs down the cake, cream flying, party hat askew, tail wagging wildly. (Audience bursts out laughing.)
[Shot 4] 9-12 s: wide shot. She comes back and freezes: only a small piece of cake is left, the dog has cream all over its face and sits back down as if nothing happened, licking its lips. She throws her hands up, baffled: "…What happened?!"
[Shot 5] 12-15 s: close-up. The dog lets out a creamy little burp and tilts its head at her. She looks at the camera and sighs. (Big laugh and applause, sitcom transition music.)
Sound: studio audience laughter and applause, a knock, the dog gobbling and a tiny burp, an upbeat 90s sitcom theme. Only these two lines. No subtitles or other text.
Motion graphics
Move
A geometric design animated with beat-synced transitions and typography.
Input
Generated output
Full prompt
A bold full-bleed motion graphic, NO frame, NO border, NO paper edges — the design fills the whole 16:9 screen. Flat Swiss/geometric style, exact colours and shapes of Image 1, crisp edges. BIG motion, never static, beat-synced to a punchy 120 BPM track.
0-3 s: the screen starts as one flat colour; geometric blocks slam in from all four sides on each beat, building the layout of Image 1.
3-7 s: circles roll across, triangles flip, stripes slide; a figure made of shapes runs through the composition, each step stamping a new colour block.
7-10 s: the letters of MOVE drop in one by one with heavy squash-and-stretch, then hold perfectly still; spell it exactly.
10-13 s: everything except MOVE explodes outward and snaps back on the beat, rearranged into a new layout.
13-15 s: fast camera push-in and a 15-degree rotation into the O of MOVE, ending on a solid colour.
Sound: punchy electronic beat with a hit on every shape. No other text.
Sunday
A risograph poster brought to life with animated lettering, a moving character, and sound.
Input
Generated output
Full prompt
A bold full-bleed risograph motion poster, NO frame, NO border, NO paper edges: the artwork fills the whole 16:9 screen edge to edge, exactly the riso grain, flat ink colours and illustration style of Image 1. Start almost empty and assemble the poster of Image 1 with BIG, playful motion, ending exactly on its composition.
0-3 s: a flat blue sky with the city skyline; the giant black-and-white cat drifts down from the top of frame like a fluffy cloud and settles onto the rooftops, eyes closed.
3-6 s: the letters of SUNDAY slam up from behind the skyline one by one with a squash-and-stretch bounce; spell it exactly. The cat stays IN FRONT of the letters, its head poking out between them exactly like Image 1.
6-9 s: the cat yawns hugely, stretches one paw across the letters, and the letters wobble; then it snuggles back down.
9-12 s: tiny details come alive: windows light up, a small plane loops behind the letters, the cat's tail flicks and curls over the skyline.
12-15 s: confident slow camera push-in toward the cat's sleepy face as it opens one eye, winks, and closes it again; SUNDAY stays perfectly still and sharp behind it.
Sound: a lazy lo-fi Sunday-morning piano loop, a big yawn, soft bouncy thumps on each letter, gentle purring. No other text.
Product films
Lip gloss
A single product reference guides a multi-shot beauty commercial.
Input
Generated output
Full prompt
A 15-second multi-shot product commercial for the chunky clear lip gloss tube with crimson jelly gloss from reference image 1 (exact same faceted clear cap, ridged neck and deep crimson gloss, no label), served like a luxury dessert.
SHOT 1 (0-3s): Overhead crane move descending onto a duck-egg blue table as a polished silver cloche lifts off a white plate, revealing the pink panna cotta with the gloss tube resting on it. The cloche's mirror surface reflects only the pale blue room and soft light — no people, no camera, no crew reflected anywhere.
SHOT 2 (3-6s): Extreme macro side angle: the panna cotta wobbles gently, light rolling across its glossy surface and through the crimson gloss inside the clear tube.
SHOT 3 (6-10s): Hard cut to a clean, empty, seamless pale duck-egg blue background — no plate, no dessert, nothing else in frame. The gloss tube floats upright in soft light and opens by itself: the faceted clear cap slowly unscrews and lifts away, drawing the applicator wand up out of the tube, a thick glossy string of crimson gloss stretching between the doe-foot tip and the tube neck and catching the light. Slow-motion macro, camera arcing smoothly around it. No hands, no people.
SHOT 4 (10-12s): Same clean duck-egg blue background: macro on the glossy crimson doe-foot applicator tip, a single drop of gloss gathering and glistening, light refracting through the faceted cap.
SHOT 5 (12-15s): Hard cut back to the table: low dolly gliding across and settling on the final hero frame exactly like reference image 1.
SOUND: refined restaurant ambience, the metallic lift of the cloche, a soft wobble, the satisfying twist-and-pull sound of the cap opening, a glossy stretch, then a chic playful French-pop instrumental. LOOK: high-end cinematic beauty/tech commercial, editorial and refined, Jacquemus / Loewe campaign energy. Soft pale true-colour grounds against deep saturated hero colours, large soft diffused light, shadows soft but present, low-to-moderate contrast, creamy highlights, subtle 35mm film grain, shallow depth of field, anamorphic feel. CAMERA is always moving with bold confident motion — dollies, macro orbits, crane moves, motion blur on fast moves — never static. Product always tack-sharp and recognisable. Hard cuts between shots, each shot a different angle and scale, one clear story from start to end. Correct real-world physics. NO text, NO titles, NO captions, NO logos, NO subtitles anywhere. Audio: no voiceover, no dialogue, only a polished commercial soundtrack plus crisp tactile sound design.
Headphones
A single product reference guides beat-synced cuts, macro details, and changing camera angles.
Input
Generated output
Full prompt
A 15-second fast-cut, high-energy cinematic product commercial for the cobalt-blue over-ear headphones from reference image 1 (exact same shape, colour, mesh headband, knit cushions and metal hinges, no logos) about deep bass. Edited like a premium tech launch film: rapid hard cuts on the beat, every shot a bold, different cinematic angle and lens. The film is about the HEADPHONES. NO glass, NO water, NO plants, NO other props. The whole film stays on a clean seamless powder-blue studio set; the background is always clean powder blue, never a wall edge or dark studio.
SHOT 1 (0-2.5s): OPENS on exactly the composition of reference image 1 — same framing and angle, the cobalt headphones holding the ripe peach between their ear cups on powder blue. The peach stays perfectly still and seated in place — it does NOT jump, bounce or move position — it only PULSES: its surface subtly swells and contracts like a heartbeat in time with each deep bass hit, the ear cups gently tightening with it. Slow push-in.
SHOT 2 (2.5-4s): Hard cut on the beat, the peach is gone for the rest of the film: snap-zoom extreme macro across the cobalt headband mesh, the fabric trembling with the bass.
SHOT 3 (4-5.5s): Extreme low worm's-eye angle, wide lens, looking up at the empty headphones towering against powder blue, fast dolly-in.
SHOT 4 (5.5-7s): Top-down overhead shot spinning slowly, the headphones laid open in a perfect graphic shape on powder blue.
SHOT 5 (7-8.5s): Dutch-angle macro sliding along the polished metal hinge and telescoping arm, light streaking across the metal.
SHOT 6 (8.5-10s): Rack focus from the soft knit ear cushion in the foreground to the second ear cup behind, shallow depth of field.
SHOT 7 (10-12s): Fast whip-pan into a low tracking shot as the headphones glide across the powder-blue floor toward camera, heavy motion blur, on the drop.
SHOT 8 (12-13.5s): Profile silhouette-style side angle, soft edge light outlining the curve of the headband, slow-motion.
SHOT 9 (13.5-15s): Crane down to the final hero shot of the empty headphones alone, three-quarter angle on powder blue, held on the last beat.
SOUND: a deep sub-bass heartbeat pulse building into a punchy, stylish minimal electronic track with hard hits on every cut and a big drop. LOOK: high-end cinematic beauty/tech commercial, editorial and refined, Jacquemus / Loewe campaign energy. Soft pale true-colour grounds against deep saturated hero colours, large soft diffused light, shadows soft but present, low-to-moderate contrast, creamy highlights, subtle 35mm film grain, shallow depth of field, anamorphic feel. CAMERA is always moving with bold confident motion — dollies, macro orbits, crane moves, motion blur on fast moves — never static. Product always tack-sharp and recognisable. Hard cuts between shots, each shot a different angle and scale, one clear story from start to end. Correct real-world physics. NO text, NO titles, NO captions, NO logos, NO subtitles anywhere. Audio: no voiceover, no dialogue, only a polished commercial soundtrack plus crisp tactile sound design.
Animated shorts
Gothic wardrobe
A character and style reference guides a multi-shot animated wardrobe sequence with interface sound effects.
Input
Generated output
Full prompt
Image 1 is the exact character, room and hand-painted style: a pale gothic girl with a black bob and cat-ear hair, a dark teal dressing room, a red vanity mirror, racks of black and red dresses and skull ornaments, thick hand-painted 2D animation. She keeps the cat-ear hair the whole time; her face stays clean and pretty, never distorted.
Game UI: top-left tab "WARDROBE"; on the right exactly three options, no empty slots: "BLACK DRESS" / "RED LOLITA" / "WHITE GOWN"; a "CONFIRM" button at the bottom. All text spelled exactly, sharp and stable; the UI stays at the edges and never covers her face. The selected option has a glowing frame.
Rule: every outfit change happens only AFTER the cursor clicks the option; before the click she keeps wearing the previous outfit.
[Shot 1] 0-3 s: wide shot as in Image 1, she wears the black dress in front of the mirror; the UI fades in with "BLACK DRESS" selected. The cursor slowly moves to "RED LOLITA".
[Shot 2] 3-6 s: the cursor clicks "RED LOLITA" and the selection jumps to it; after the click, a quick sparkle flash and she is in a red lolita dress; she twirls once in front of the mirror and winks playfully.
[Shot 3] 6-10 s: the cursor moves to "WHITE GOWN", pauses half a second, clicks; after the click, the same sparkle flash and she is in a long white gown with a veil; she twirls, the veil lifts and settles. A small push-in: she lifts her eyes to the camera for a moment of eye contact, the corner of her mouth rising slightly.
[Shot 4] 10-12 s: the cursor clicks "CONFIRM", the UI panels fold away, a small "READY" flashes in the bottom-right corner without covering her.
[Shot 5] 12-15 s: the camera follows behind her as she pushes open the doors into a moonlit gothic ballroom full of dancing ghosts; her veil floats behind her and the doors close.
Sound: UI clicks, the same sparkle chime and dress swish on each change, a creaky door, a music-box gothic waltz.
NOVA space courier
A character reference guides an eight-cut retro-anime trailer with title cards and music.
Input
Generated output
Full prompt
A 15-second 1980s anime game-character trailer with 8 hard cuts, 24 fps. Image 1 is the only authority for the courier girl's identity, outfit and retro cel style. Palette: hot pink, cyan, purple and sunset orange synthwave. On-screen text appears ONLY on the three flat text cards below, each one line in a huge clean bold sans-serif, perfectly spelled, centred, held still for at least one second, nothing overlapping the letters. She is cute and playful: real winks and a peace sign.
CUT 1 (0-1.5 s): extreme close-up, her eyes open under the visor, a glowing parcel reflected in it.
CUT 2 (1.5-3 s): text card, white words on hot pink: "SPACE COURIER".
CUT 3 (3-5.5 s): she guns her hover bike out of a neon hangar, pink speed trail, parcel strapped behind her.
CUT 4 (5.5-8 s): she swerves between asteroids, one hand on the parcel, and laughs.
CUT 5 (8-9.5 s): text card, white words on deep purple with a small ringed-planet icon: "NEXT STOP: SATURN".
CUT 6 (9.5-11.5 s): she drifts the bike to a stop on a moon porch with Saturn's rings in the sky and rings the doorbell.
CUT 7 (11.5-13 s): a tiny green alien opens the door; she hands over the parcel and winks.
CUT 8 (13-15 s): final card, huge white word on cyan: "NOVA"; she leans into frame beside it flashing a peace sign.
Music: upbeat 80s synth-pop with a hit on every cut. No other text, letters or numbers anywhere.
Composition from multiple references
Knitted game world
A game-asset overview guides characters, environments, and objects across a multi-shot sequence.
Input: 7 reference images
Generated output
Full prompt
"Level 1" — a short, clear story.
References: Image 1 is the HERO, a real boy in a sky-blue knitted power-suit with pom-pom antennae — he is always this boy. Image 2 is his SIDEKICK, a felted green caterpillar — always this caterpillar. Image 3 is the BOSS, a giant crocheted crow with mismatched button eyes and a tiny crown. Image 4 is the knitted meadow level. Image 5 is the knitted boss arena with the yarn nest. Image 6 shows the collectible knitted heart, golden button and felt star. Image 7 shows the sewing-tin treasure chest. Keep every character, place and object exactly as in its reference.
The world: a REAL handmade knitted video-game world filmed for real like a stop-motion feature — matte wool, felt, crochet and impasto paint, bright toy colours, soft daylight, slight film grain. Side-scrolling camera. No on-screen text, no numbers.
[Shot 1, 00:00–00:06] The meadow of Image 4, camera tracking sideways: the hero runs right, bounces off a red felt toadstool springboard and hops across the floating wool-cushion platforms, snatching golden buttons from Image 6 out of the air — each one pops into a little puff of yarn fluff; the caterpillar sidekick wriggles along behind him. He grabs the knitted heart from Image 6 and it bursts into a shower of tiny felt hearts around him — the boy himself does not change at all. At the end of the meadow he and the caterpillar dive head-first into the open end of a giant striped knitted tube lying on the grass.
[HARD CUT at 00:06] The boss arena of Image 5: the hero shoots out of the other end of the knitted tube, tumbles, and lands on his feet; the caterpillar plops out after him.
[Shot 3, 00:07–00:12] The giant crow boss from Image 3 rises out of the yarn nest, spreads its felt wings and swoops down at the hero; the hero jumps high and bounces once on the crow's head with both boots; the crow wobbles, its tiny crown flies off, and it flops back into the nest, dizzy.
[Shot 4, 00:12–00:15] The sewing-tin treasure chest from Image 7 pops open beside the flag, golden buttons spilling out; the hero and the caterpillar jump up and down cheering.
NO MORPHING — CRITICAL: every character keeps exactly its own body, species, face, size, costume and colours from its reference in every frame. Nobody and nothing ever transforms, melts, dissolves, merges, shape-shifts or turns into another character, animal or object. Scene changes happen ONLY on a clean hard cut, never as a morph or a dissolve. SPEED IS CRITICAL: everything moves at normal real-time speed, like 24fps footage played back at 24fps - natural motion, NO slow motion.
Sound: soft woolly boings, plinky button pops, a sparkly twinkle for the heart, a whoosh through the tube, flapping felt wings, a muffled crow squawk, a cheerful chest pop. Music: a bouncy chiptune melody played on real toy piano, ukulele and glockenspiel. Mix the sound effects and music clearly and fully audible.
Brutalist moodboard
A moodboard guides characters and architecture in a three-shot cinematic sequence.
Input: 6 reference images
Generated output
Full prompt
A quiet, cinematic three-shot sequence, 12 seconds, 16:9, shot on medium-format film with soft overcast light, using the attached references.
The woman is from image 3: same face, hair and outfit. The man is from image 5: same face, coat and silhouette, with everything he wears in the same pale cream as the city. The world is built from the architecture in images 1, 2 and 6: the curved stacked ramp tower, the vast concrete stairways and the interlocking monolithic blocks, all raw board-marked concrete in monochrome cream and beige under a pale sky. She is the only colour in the world.
SHOT 1, CLOSE-UP (0-3s): Tight on her face and shoulders against a board-marked concrete wall. She looks down, then slowly lifts her eyes and turns her head to look off to the right, as if she has noticed something.
SHOT 2, LOW-ANGLE MEDIUM-WIDE, THEN ORBIT (3-10s): On a broad concrete terrace in front of the vast stairways and the curved ramp tower, the man stands completely still, head tilted back, looking up at the sky. She walks up beside him and lightly touches his coat sleeve. In that instant his coat becomes a rich deep plum and the trousers underneath a deep maroon. He lowers his head and turns to look at her as she walks on past him and out of frame to the right. The camera slowly orbits around him, low and close, the ramp tower and concrete blocks wheeling behind him, as he stands watching where she went.
SHOT 3, WIDE (10-12s): A vast, still, futuristic plaza of monumental cream concrete blocks, cantilevered slabs and stepped terraces, the curved ramp tower at the left edge of frame. Two tiny figures stand far apart on the empty paving: him, a dark spot of plum, and her, a small spot of baby blue. Nothing else moves.
Fabric moves with weight, steps land with weight. No one speaks. No music; only soft footsteps on concrete, the brush of her hand on his sleeve, and a light wind.
Selective video editing
Bowling-alley transformation
Replace the environment, a ball, and background figures while preserving the foreground actor, camera, and timing.
Input
Generated output
Full prompt
Edit the reference video @Video1. Keep the same camera framing and moves, the same timing, the same hard cut to the wider shot, and every movement of the man in the foreground. Copy the man in the foreground from @Video1 exactly, frame for frame - same face, hair, clothes, bracelets, shoes, the milk bottle, the bench he sits on and the clear bag at his feet. Do not restyle or re-draw him.
Change only three things:
- ENVIRONMENT: the lawn-bowls green becomes a real retro bowling alley. Where the striped grass was, there are glossy honey-coloured maple bowling lanes running away from the camera along the same diagonal, with gutters and dotted lane arrows; the white rail becomes the lane divider and ball return; at the far end, white pins stand under a soft-lit pin deck. Bright, soft, even overhead light, warm wood against cream and tomato red - not dark, not neon.
- The single loose red ball on the floor beside the bag becomes a glossy cherry-red bowling ball with three finger holes. It rolls exactly as the red ball rolls in @Video1 - on its own, curving down the lane toward the far end.
- The two elderly men in white in the far background become two real deer standing on the far end of the lane, small and soft in the distance, heads lowered sniffing at the floor where the men were bending. When the ball arrives in the wide shot, both deer lift their heads and turn to look at the man, exactly when the men straighten up and turn in @Video1.
Wardrobe and scene edit
Replace knitwear with cotton T-shirts and add a distant airplane while preserving the original scene and performance.
Input
Generated output
Full prompt
Edit this video. Keep everything exactly as it is — the same location, framing, camera, timing, the three men (faces, hair, trousers, socks, shoes, every movement and gesture), the pink felt carpet and how it moves, the lighting, colour and sound. Change only the two things below.
- TOPS: Each man's knitted jumper becomes a plain short-sleeved crew-neck cotton t-shirt in the same colour as his jumper. No knit texture, no ribbing, no stripes, no logos or graphics, a smooth flat matte cotton surface. Each one is a relaxed regular fit, slightly loose, hanging naturally, with soft creases where the carpet presses down on the shoulder. His forearms are now bare.
- First man (ginger hair): plain forest-green t-shirt. The two cream chest stripes are gone.
- Second man (dark wavy hair): plain ecru cream t-shirt.
- Third man (short black curls): plain tomato-red t-shirt.
- AIRPLANE: Add one small two-seat propeller stunt plane far away in the sky. It flies from left to right, the opposite direction to the men. Classic single-engine aerobatic plane, low wing, two seats one behind the other, painted white with a red stripe. It is small in frame, roughly the size of a fingernail, high in the sky above the horizon, and never passes in front of the men or the carpet. Timing: the sky is empty until about 10 seconds in. Then the plane appears at the left edge of frame, high up, and flies slowly to the right. It moves slowly enough that by the last frame it is only about a third of the way across the frame, still high in the left side of the sky — which is exactly where the third man is looking when he stops and squints up. Same soft overcast light on it as the rest of the scene, slightly hazy with distance. No smoke trail. Add the faint drone of a distant propeller engine to the sound.
Nothing else changes.
Motion transfer
Record-shop dance
Combine a character sheet, environment image, and motion-reference video in a stylized dance sequence.
Input
Video reference 1
Image reference 1
Image reference 2
Generated output
Full prompt
References: @Image1 is the character sheet (front, three-quarter and back views of ONE character). @Image2 is the environment. @Video1 is the motion reference only. Make a new video of the character from @Image1, inside the environment from @Image2, performing exactly the body motion, choreography, timing and camera movement of the person in @Video1. Take NOTHING else from @Video1 - not the person's face, body, clothes, props, location, background, colours or photographic look. Keep the character exactly as designed in @Image1 - same face, proportions, outfit, colours - and render the whole video in the exact art style of @Image1 and @Image2, never photoreal. Only ONE character in frame. Style: vintage pop-art fashion illustration with black ink outlines, flat colour blocks, halftone and risograph grain, exactly as @Image1 and @Image2. Desmond dances on the terracotta checkerboard floor of his record shop from @Image2: the same whirling spins, flung arms and kicks as in @Video1, his mustard suit jacket flaring with each turn. Where the person in @Video1 holds the potted fern overhead, Desmond holds his vinyl record high above his head like a dance partner. He ends with arms wide and a little deadpan bow. Audio: an original upbeat 1960s-style soul groove playing from the turntable, loafers tapping and squeaking on the floor, a soft record-crackle. Normal real-time speed, NO slow motion. No text, no captions, no logos.
Station choreography
Combine an opening frame, character sheet, and motion reference to direct a multi-shot dance sequence.
Input
Video reference 1
Image reference 1
Image reference 2
Generated output
Full prompt
Image 1 is the opening frame and the world of this video: the lanky teenage boy with curly copper-red hair and freckles, in an oversized faded navy-and-white vintage track jacket, grey shorts, red-striped white tube socks and scuffed white sneakers, in the sunlit art-deco station concourse with the checkered floor. Start exactly on Image 1. Every shot takes place in this station, in this light and this real photographic look.
Image 2 is his character reference sheet: keep his face, hair, freckles, build and outfit exactly as in Image 2.
Video 1 is the MOTION and CAMERA reference: perform Video 1's full-body choreography exactly, move for move, in the same order and with the same timing, and use the same three shots with the same cuts at the same moments. Take only the movement and camera from Video 1 — never its person, clothing, car park, skyline or sunset. Every shot is re-staged inside the station.
[Shot 1, 00:00–00:04.7] Low camera at knee height, orbiting slowly around him, as in Image 1. He finishes the pivot on his left foot, steps his right leg far back into a deep low lunge, and holding the lunge sweeps his arms up until one arm reaches straight overhead.
[Shot 2, at 00:04.7 cut to a medium shot from the front, closer, keeping his whole head and body in frame] He rises out of the lunge and turns with both arms spread wide, his loose jacket flaring out, then lifts one knee high to his chest with both arms raised overhead, balancing on one foot; then he drops down to the tiles and rolls sideways across the checkered floor in one continuous roll.
[Shot 3, at 00:09.8 cut to a wide shot from behind him, low] He springs up from the floor and runs away from the camera across the long checkered floor toward the bright arched window, makes a small running leap, lands softly and stands still, chest rising, as the sun blazes through the window ahead of him.
He belongs to the station in every shot: the sunlight from the arched window rims his curls and shoulders, and his shadow and faint reflection move with him across the polished tiles; his sneakers press into and push off the floor with real weight.
PHOTOREALISM IS CRITICAL: real footage shot on 35mm film — a slight fine film grain, real skin with pores and freckles, real creased fabric, natural light; no gloss, never airbrushed, never CGI, never plastic. SPEED IS CRITICAL: everything moves at normal real-time speed, like 24fps footage played back at 24fps - natural motion, NO slow motion.
Sound: sneakers squeaking and sliding on stone tiles, the nylon jacket swishing, the echo of the huge empty hall, the station clock ticking. Music: an upbeat, bouncy indie-pop track with handclaps that lifts at the final leap. Mix the sound effects and music clearly and fully audible.
Resolution finishing
Horse catwalk
The same 768p generation finished at 2K and 4K.
Input · 768p
768p draft
Outputs · 2K and 4K
2K
4K
Drawing-room hurdle
The same 768p generation finished at 2K and 4K.
Input · 768p
768p draft
Outputs · 2K and 4K
2K
4
Methods
Pika H3 Flex was developed through three post-training stages: supervised fine-tuning, distribution matching distillation, and post-distillation reinforcement learning (Fig. 1).
We first fine-tuned the pretrained MiniMax H3 model (MiniMax, 2026) on an in-house, aesthetically curated audio-video dataset spanning multiple generation and editing tasks. We then applied Projected Distribution Matching Distillation (PDMD; Wang et al., 2026) to reduce sampling from 50 denoising steps to six. Following distillation, we trained separate adapters using Reward Feedback Learning (ReFL; Xu et al., 2023) and Direct Preference Optimization (DPO; Wallace et al., 2024) to improve visual aesthetics, clarity, instruction following, and realism. Human pairwise preferences calibrated the automated judgments used to construct the DPO training pairs. The adapters were merged for inference without introducing additional generation steps.
The resulting model led the compared systems in our human evaluation on human anatomy, complex composition, and audio synchronization for text- and image-conditioned generation, and on visual style and framing and camera motion for reference-conditioned generation.
We further accelerated inference through fused kernels, cached timestep-dependent modulation, FlashAttention-4, block-sparse attention, and compressed communication. The optimized inference stack is served through the Pika API.
Supervised Finetuning
We performed supervised fine-tuning on a proprietary dataset curated in-house, comprising audio-video clips spanning everyday activities, performance, narrative, technical, and stylized content. For inquiries about the dataset, please contact enterprise@pika.art.
Starting from the official, guidance-distilled MiniMax H3 model (MiniMax, 2026), we fine-tuned on this dataset while retaining its 50-step sampling schedule. The resulting model exhibited improved visual style and aesthetics relative to the original model and served as the teacher for subsequent distillation.
Step Distillation
Starting from the 50-step, guidance-distilled model obtained through supervised fine-tuning, we applied Distribution Matching Distillation (DMD; Yin et al., 2024) to reduce the number of denoising steps. In our experiments, standard DMD exhibited training instability. We adopted Projected Distribution Matching Distillation (PDMD; Wang et al., 2026), which substantially improved training stability and output quality. The resulting student generates joint audio-video outputs in six denoising steps, compared with 50 for the teacher.
Post-Distillation Reward Learning
Starting from the six denoising step model, we performed post-distillation reward learning to align the model output with human preference. Specifically, we applied Reward Feedback Learning (ReFL; Xu et al., 2023) and Direct Preference Optimization (DPO; Wallace et al., 2024). ReFL targets aesthetic quality and visual clarity through differentiable reward feedback on current student outputs, while DPO learns from offline preferences covering instruction following, realism, and visual quality. We trained the two as separate low-rank adapters (LoRA; Hu et al., 2022) and merged their weights into one adapter, combining the strengths of ReFL and DPO at no added inference cost.
Inference acceleration
Starting from the post-trained six-step generator model, we cached timestep-dependent AdaLN projections and applied three groups of inference optimizations: fused kernels and packed communication, FlashAttention-4, and block-sparse attention with compressed communication. These reduce kernel overhead, attention computation, and communication costs without additional training.
Together, these optimizations reduced full-inference latency from 3.409 to 2.129 seconds, delivering 1.60× faster full inference and 1.74× faster transformer forwards. Measurements used five-second 768p text-to-video generation with audio on eight B300 GPUs, with the checkpoint and six-step sampling schedule held fixed (Fig. 2).
Results
In this section, we benchmarked Pika H3 Flex against leading video models, including the official MiniMax H3 (model card), Fal H3 Max (fal, 2026), and HeyGen Video (quality; model documentation). Quantitative metrics and qualitative comparisons demonstrate strong performance in visual quality, aesthetics, and audio quality, with comparable or better audio-video alignment and leading human preference scores in several evaluated categories.
Quantitative Evaluation
We evaluated text- and image-to-video generation against the official MiniMax H3 (model card), Fal H3 Max (fal, 2026), and HeyGen Video (quality; model documentation). All models generated outputs at 768p using the same prompt set. We assessed visual quality, audio quality, and cross-modal alignment using seven metrics, with higher scores indicating better performance:
- VT Video Technical — low-level visual integrity and technical quality (DOVER++; Wu et al., 2023).
- VA Video Aesthetic — high-level visual aesthetic quality (Aesthetic Predictor V2.5).
- KVQ Kaleidoscope Video Quality — saliency-guided local video quality assessment (Qu et al., 2025).
- UVQ Universal Video Quality — combined detail, exposure, contrast, noise, temporal-stability and color terms (Wang et al., 2021).
- AA Audio Aesthetic — equal mean of AudioBox perceptual quality (PQ) and content usefulness (CU; Tjandra et al., 2025).
- A-V Audio-Video Alignment — text-independent semantic agreement between audio and video (ImageBind; Girdhar et al., 2023).
- T-V Text-Video Alignment — semantic similarity between the video prompt and generated video, normalized to the table scale (VideoCLIP-XL-v2 model card).
Pika H3 Flex achieves the highest scores on all five visual and audio quality metrics (Table 3), outperforming each comparator on a majority of prompts for these metrics (Fig. 3). Audio-video and text-video alignment remain comparable across the four models.
Table 3. Quantitative evaluation of text- and image-to-video generation at 768p on the same prompt set. Higher is better for all metrics; bold indicates the highest score in each row. Results are from our evaluation.
Metric | Pika H3 Flex | MiniMax H3 | Fal H3 Max | HeyGen Video (quality) |
VT ↑ | 12.3 | 11.7 | 10.6 | 11.5 |
VA ↑ | 5.41 | 5.28 | 5.00 | 5.33 |
KVQ ↑ | 0.024 | 0.022 | 0.019 | 0.020 |
UVQ ↑ | 0.602 | 0.595 | 0.590 | 0.595 |
AA ↑ | 7.02 | 6.85 | 6.89 | 6.76 |
A-V ↑ | 0.212 | 0.208 | 0.216 | 0.201 |
T-V ↑ | 0.225 | 0.225 | 0.226 | 0.224 |
Qualitative Evaluation
We conducted blind pairwise human preference studies against the same three models on text-, image-, and reference-to-video generation. All models rendered the same prompts at 768p. We report per-category Bradley–Terry scores on an Elo scale centered on the four evaluated models.
As shown in Fig. 4, Pika H3 Flex leads on human anatomy, complex composition, and audio synchronization for text- and image-to-video, and on visual style and framing and camera motion for reference-to-video. Other systems remain preferred in physics, multi-scene narrative, lighting and materials, and action-heavy or control-video reference tasks.
Latency
With our optimization, Pika H3 Flex achieves transformer denoising time below the generated clip duration across all evaluated configurations. At 768p, it is approximately 1.8× faster than Fal H3 Max on average across text-to-video, image-to-video, and 10-second reference-to-video comparisons.
For text/image-to-video, Pika H3 Flex is 1.5× faster than Fal H3 Max for 5-second clips and approximately 2× faster for 10- and 15-second clips, while matching Fal H3 Max Turbo. For reference-to-video, it achieves 1.3–1.7× speedups with image or image-plus-audio references for clips of at least 10 seconds, and 1.6–1.9× speedups with video references across all tested durations. Fal H3 Max Turbo does not offer a reference-to-video endpoint.
Figure 5 reports transformer denoising latency against the fal-hosted models. Figures 6 and 7 report warm inference latency by output resolution and conditioning type, respectively. Transformer comparisons exclude conditioning, decoding, and serving overhead; warm inference includes conditioning and decoding but excludes serving overhead.
Beyond the Generator: 4K Video and Audio Stems
Pika H3 Flex generates natively at 480p and 768p with synchronized stereo audio. Two finishing models extend these outputs: a single-step video upsampler for 1080p, 2K, and 4K, and a prompted audio separator for editable stems. Both operate within the same request without modifying the generator.
Video Upsampling
Our in-house upsampler is a 1B-parameter transformer that operates directly in pixel space, producing upscaled frames in a single network evaluation. Causal processing with a frame cache keeps memory independent of clip length, while block-sparse attention scales linearly with pixel count.
We trained the model from scratch on high-quality footage, including native 4K video, using synthetic degradations to construct low-resolution inputs. Reconstruction, adversarial, and reward objectives improved detail and temporal consistency while preserving composition, identity, motion, and color. The model supports scale factors from 1× to 4×, with deterministic output independent of batching and request order (see Fig. 8).
On a single GPU, the first 1080p frame is available after 0.17 seconds (see Fig. 9). Processing time scales linearly with clip length. We will opensource the upsampler model.
Audio Stem Separation
The generator jointly produces speech, music, effects, and ambience as one synchronized soundtrack. We apply prompted source separation to recover five editable stems: dialogue, score, SFX, foley, and ambience.
This approach preserves the generator’s audio objective and avoids requiring stem-labeled video for training. The stems are returned alongside the unchanged video and mixed soundtrack, enabling independent audio editing in post-production.
Conclusion
We introduced Pika H3 Flex, an open-weight model family for joint video and audio generation built through supervised fine-tuning, six-step distribution matching distillation, and reward and preference alignment. Combined with an optimized inference stack, Flex reduces generation from 50 to six denoising steps while improving measured visual and audio quality relative to the evaluated models, with comparable cross-modal alignment.
Flex supports text-, frame-, and reference-conditioned generation and editing, with up to 4K finishing and five editable audio stems. These capabilities enable rapid creative iteration across generation, motion transfer, and targeted video edits. Physics, multi-scene consistency, and complex reference control remain areas for further improvement. We release the model weights and code to support further research and development.
Access
Pika H3 Flex is available through our API, ComfyUI, and Diffusers, with model weights available on Hugging Face and code on GitHub. Explore the Pika API. Pika also offers customized training services. For inquiries, contact enterprise@pika.art.
References
- MiniMax. MiniMax H3. Model card, 2026. Hugging Face.
- Yin, T., et al. One-step Diffusion with Distribution Matching Distillation. CVPR, 2024. Paper.
- Wang, Z., et al. PDMD: Projected Distribution Matching Distillation for Video Diffusion Models. 2026. Paper · Project page.
- Xu, J., et al. ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation. NeurIPS, 2023. Introduces Reward Feedback Learning (ReFL). Paper.
- Wallace, B., et al. Diffusion Model Alignment Using Direct Preference Optimization. CVPR, 2024. Paper.
- fal. Introducing H3 Max by fal. 2026. Model announcement.
- HeyGen. HeyGen Video. Official model and API documentation. Documentation.
- Wu, H., et al. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives. ICCV, 2023. DOVER and DOVER++. Paper · Code.
- discus0434. Aesthetic Predictor V2.5: SigLIP-based Aesthetic Score Predictor. Code and model information.
- Qu, Y., et al. KVQ: Boosting Video Quality Assessment via Saliency-guided Local Perception. CVPR, 2025. Paper.
- Wang, Y., et al. Rich Features for Perceptual Quality Assessment of UGC Videos. CVPR, 2021. Universal Video Quality (UVQ). Paper · Code and model checkpoints.
- Tjandra, A., et al. Meta Audiobox Aesthetics: Unified Automatic Quality Assessment for Speech, Music, and Sound. 2025. Paper.
- Girdhar, R., et al. ImageBind: One Embedding Space To Bind Them All. CVPR, 2023. Paper.
- Alibaba PAI. VideoCLIP-XL-v2. Model card and checkpoints. Hugging Face.
- Hu, E. J., et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR, 2022. Paper.
