H3-Base Task Family

MiniMax H3 FL2VA

First-Last-frame-to-Video-and-Audio — an H3-Base task family for text-driven and frame-controlled video generation.

FL2VA is one of two task families in MiniMax H3-Base, the open-weight release of MiniMax H3. It covers text-to-video, first-frame-anchored video, and first + last frame transition generation in a single checkpoint. The companion task family, Ref2VA, handles multimodal reference-to-video with image, video, and audio references.

  • H3-Base Task Family
  • Text-to-Video
  • First-Frame Anchor
  • First + Last Frame
  • Native Audio
  • Open Weights on Hugging Face
FL2VA vs Ref2VA10 Free Credits Daily

FL2VA workflows run through the Image-to-Video tab below. Upload a first frame (required for frame-anchored FL2VA) and, optionally, a last frame. Leave both empty and switch to Text-to-Video to run FL2VA as pure text-to-video.

First frame (required)0/1
Last frame (optional)0/1
0/2500
768pRecommended

Recommended for testing prompts and first drafts.

2K output is available on the 2K tier only.

From input image

Output follows the uploaded frame.

5s
Est. cost: ~30 creditsBalance: — credits

Checking your account…

OutputVideo
History
MiniMax H3 2K image-to-video: family dinner with ramen bowlMiniMax H3 Image to Video sample

Pick a result, make it yours

Create a Video Like This

Choose an example and your prompt, model, and settings load instantly.

Vintage Binocular Brand Film AI video previewProduct Ads

Vintage Binocular Brand Film

Product Ads
Model
MiniMax H3
Duration
5 seconds
Resolution
768p
Aspect ratio
16:9

“Use Images 1–4 as sequential keyframes, seen through a vintage binocular viewfinder searching for the MINIMAX installation. Open out of focus with subtle handheld shake, then push in quickly and rack focus onto Image 1. Between keyframes, use fast binocular-scan transitions with whip movement, motion blur, optical smearing, and brief exposure flicker. Cut at peak blur, then settle and snap back into focus. Keep the twin circular lens mask absolutely fixed throughout: identical position, scale, feathered black vignette, and edge softness, with no warping or drift. Only the image inside the mask may move. In Image 2, let the fabric move gently in the wind while the MINIMAX lettering follows the folds and remains legible. In Image 3, the subject should feel like a stylish passerby caught by chance, walking, turning, and swinging their arms naturally. In Image 4, the subject adjusts their glasses or lifts their chin slightly with a cool, effortless fashion-campaign attitude. Red typography should resolve with the focus: begin slightly blurred and at low opacity, then fade into clarity over 0.3–0.5 seconds. A subtle vertical slide or slight tracking expansion is allowed. Fade it out before the next transition or let motion blur carry it away. No spins, bounces, or large fly-ins/outs. Visual language: a voyeuristic, Wes Anderson-inspired 35 mm film look with fine grain, soft highlight halation, restrained color, and red typographic accents. Minimal, premium, lightly playful. Do not add people, vehicles, buildings, or logos. Preserve the core composition and the MINIMAX installation exactly. PRODUCTION CONTROL MODE: Reference to Video. Assign Images 1–4 only as sequential visual anchors. TIMELINE: 0–3s search and rack focus; 3–10s two fast scan transitions; 10–15s settle on the final mark. CONTINUITY LOCKS: the twin circular mask, vignette softness and installation lettering never change geometry. CAMERA: handheld micro-shake inside the mask only; cut at peak motion blur. AUDIO: wind, cloth movement and two short optical whooshes; restrained low pulse, no dominant score. END FRAME: hold the final mark sharp for one second. AVOID: mask drift, melted text, cross-dissolves and extra subjects.”

View full prompt

Use Images 1–4 as sequential keyframes, seen through a vintage binocular viewfinder searching for the MINIMAX installation. Open out of focus with subtle handheld shake, then push in quickly and rack focus onto Image 1. Between keyframes, use fast binocular-scan transitions with whip movement, motion blur, optical smearing, and brief exposure flicker. Cut at peak blur, then settle and snap back into focus. Keep the twin circular lens mask absolutely fixed throughout: identical position, scale, feathered black vignette, and edge softness, with no warping or drift. Only the image inside the mask may move. In Image 2, let the fabric move gently in the wind while the MINIMAX lettering follows the folds and remains legible. In Image 3, the subject should feel like a stylish passerby caught by chance, walking, turning, and swinging their arms naturally. In Image 4, the subject adjusts their glasses or lifts their chin slightly with a cool, effortless fashion-campaign attitude. Red typography should resolve with the focus: begin slightly blurred and at low opacity, then fade into clarity over 0.3–0.5 seconds. A subtle vertical slide or slight tracking expansion is allowed. Fade it out before the next transition or let motion blur carry it away. No spins, bounces, or large fly-ins/outs. Visual language: a voyeuristic, Wes Anderson-inspired 35 mm film look with fine grain, soft highlight halation, restrained color, and red typographic accents. Minimal, premium, lightly playful. Do not add people, vehicles, buildings, or logos. Preserve the core composition and the MINIMAX installation exactly. PRODUCTION CONTROL MODE: Reference to Video. Assign Images 1–4 only as sequential visual anchors. TIMELINE: 0–3s search and rack focus; 3–10s two fast scan transitions; 10–15s settle on the final mark. CONTINUITY LOCKS: the twin circular mask, vignette softness and installation lettering never change geometry. CAMERA: handheld micro-shake inside the mask only; cut at peak motion blur. AUDIO: wind, cloth movement and two short optical whooshes; restrained low pulse, no dominant score. END FRAME: hold the final mark sharp for one second. AVOID: mask drift, melted text, cross-dissolves and extra subjects.

Epic Space Opera Teaser AI video previewCinematic

Epic Space Opera Teaser

Cinematic
Model
MiniMax H3
Duration
5 seconds
Resolution
768p
Aspect ratio
16:9

“Epic theatrical space-opera teaser Keep the pace fast and the scale enormous without letting the edit drag. Use sharp hard cuts, a shaking command deck, white-hot flashes, split-second black frames, and a violent jump-to-warp impact. Title cards should use wide-tracked cinematic typography—not pure white—with restrained material texture, subtle illumination, and a faint edge glow. Animate the titles by emerging from deep-space shadow, catching a sweep of starlight, opening their letter spacing, leaving a slight afterimage, and flashing briefly against black. PRODUCTION CONTROL MODE: First & Last Frame. SHOT BUDGET: maximum four shots. TIMELINE: 0–3s command-deck disturbance; 3–7s gateway reveal; 7–11s jump-to-warp impact; 11–15s title on black. CAMERA: one primary movement per shot. TEXT LOCK: exact title copy, centered, wide-tracked, fully readable for at least 24 frames. AUDIO: sub-bass rise, one metallic impact at the warp cut, then a short vacuum drop. AVOID: extra title cards, soft dissolves, pure-white typography and constant camera shake.”

View full prompt

Epic theatrical space-opera teaser Keep the pace fast and the scale enormous without letting the edit drag. Use sharp hard cuts, a shaking command deck, white-hot flashes, split-second black frames, and a violent jump-to-warp impact. Title cards should use wide-tracked cinematic typography—not pure white—with restrained material texture, subtle illumination, and a faint edge glow. Animate the titles by emerging from deep-space shadow, catching a sweep of starlight, opening their letter spacing, leaving a slight afterimage, and flashing briefly against black. PRODUCTION CONTROL MODE: First & Last Frame. SHOT BUDGET: maximum four shots. TIMELINE: 0–3s command-deck disturbance; 3–7s gateway reveal; 7–11s jump-to-warp impact; 11–15s title on black. CAMERA: one primary movement per shot. TEXT LOCK: exact title copy, centered, wide-tracked, fully readable for at least 24 frames. AUDIO: sub-bass rise, one metallic impact at the warp cut, then a short vacuum drop. AVOID: extra title cards, soft dissolves, pure-white typography and constant camera shake.

Desert Fashion Campaign AI video previewFashion

Desert Fashion Campaign

Fashion
Model
MiniMax H3
Duration
5 seconds
Resolution
768p
Aspect ratio
16:9

“Create a premium 16:9 landscape fashion film. Use Image 1 for the overall mood, location, and film texture; Image 2 for the talent; Image 3 for the bag; and Image 4 for the closing brand mark. This is a fashion campaign for the clothing and bag. The tone is elevated, cool, and restrained, but the edit should still feel lively and fashion-forward—not like a conventional narrative film or an e-commerce ad. Keep the story simple: beside a vintage car on a desert highway, a woman walks to the rear of the car, opens the trunk, takes out a black bag, shares a quiet beat with the man standing nearby, then leaves carrying the bag. Integrate the clothing and bag naturally into the performance so they feel like part of the characters’ identity. PRODUCTION CONTROL MODE: Reference to Video. Image 1 controls desert location and film texture; Image 2 the woman; Image 3 the black bag; Image 4 the closing mark. ACTION ORDER: approach trunk → open trunk → lift bag by its handle → brief eye contact → walk away. CAMERA: three restrained editorial setups, never more than one move per setup. PRODUCT LOCK: preserve bag silhouette, material, hardware and color. AUDIO: desert wind, trunk latch, footsteps on gravel; minimal music. END FRAME: clean brand mark after the walk-away. AVOID: extra products, wardrobe swaps and e-commerce turntable shots.”

View full prompt

Create a premium 16:9 landscape fashion film. Use Image 1 for the overall mood, location, and film texture; Image 2 for the talent; Image 3 for the bag; and Image 4 for the closing brand mark. This is a fashion campaign for the clothing and bag. The tone is elevated, cool, and restrained, but the edit should still feel lively and fashion-forward—not like a conventional narrative film or an e-commerce ad. Keep the story simple: beside a vintage car on a desert highway, a woman walks to the rear of the car, opens the trunk, takes out a black bag, shares a quiet beat with the man standing nearby, then leaves carrying the bag. Integrate the clothing and bag naturally into the performance so they feel like part of the characters’ identity. PRODUCTION CONTROL MODE: Reference to Video. Image 1 controls desert location and film texture; Image 2 the woman; Image 3 the black bag; Image 4 the closing mark. ACTION ORDER: approach trunk → open trunk → lift bag by its handle → brief eye contact → walk away. CAMERA: three restrained editorial setups, never more than one move per setup. PRODUCT LOCK: preserve bag silhouette, material, hardware and color. AUDIO: desert wind, trunk latch, footsteps on gravel; minimal music. END FRAME: clean brand mark after the walk-away. AVOID: extra products, wardrobe swaps and e-commerce turntable shots.

Cyber-Grunge Fashion Film AI video previewCinematic

Cyber-Grunge Fashion Film

Cinematic
Model
MiniMax H3
Duration
5 seconds
Resolution
768p
Aspect ratio
16:9

“Use Image 1 as the reference for texture and mood, and Image 2 for the subject’s appearance. Generate a 15-second, 16:9 fashion short. Preserve the subject’s identity: long platinum-blonde hair, narrow black vintage sunglasses, a glossy black patent-leather trench coat, a cool, self-assured expression, and orange firelight reflected across the coat. Style: fast-cut fashion film on analog stock, set against a nighttime blaze with black smoke and orange-red flames. Layer in VHS glitches, CCTV signal interruptions, 1990s film grain, scanlines, chromatic aberration, light leaks, flash-to-white transitions, and subtle frame jitter. PRODUCTION CONTROL MODE: Reference to Video. Image 1 controls texture and fire-lit mood; Image 2 controls identity, hair, sunglasses and coat. TIMELINE: 0–4s close-up identity lock; 4–10s two fashion poses with hard cuts; 10–15s walking hero shot. EFFECT RULE: glitches may affect signal and background, not facial geometry or garment construction. CAMERA: handheld 35mm language with two deliberate flash cuts. AUDIO: fire crackle, low industrial beat and one tape-stop accent. AVOID: face replacement, coat color changes, continuous glitch overlays and unreadable silhouettes.”

View full prompt

Use Image 1 as the reference for texture and mood, and Image 2 for the subject’s appearance. Generate a 15-second, 16:9 fashion short. Preserve the subject’s identity: long platinum-blonde hair, narrow black vintage sunglasses, a glossy black patent-leather trench coat, a cool, self-assured expression, and orange firelight reflected across the coat. Style: fast-cut fashion film on analog stock, set against a nighttime blaze with black smoke and orange-red flames. Layer in VHS glitches, CCTV signal interruptions, 1990s film grain, scanlines, chromatic aberration, light leaks, flash-to-white transitions, and subtle frame jitter. PRODUCTION CONTROL MODE: Reference to Video. Image 1 controls texture and fire-lit mood; Image 2 controls identity, hair, sunglasses and coat. TIMELINE: 0–4s close-up identity lock; 4–10s two fashion poses with hard cuts; 10–15s walking hero shot. EFFECT RULE: glitches may affect signal and background, not facial geometry or garment construction. CAMERA: handheld 35mm language with two deliberate flash cuts. AUDIO: fire crackle, low industrial beat and one tape-stop accent. AVOID: face replacement, coat color changes, continuous glitch overlays and unreadable silhouettes.

Retro Anime Crime Title Sequence AI video previewCinematic

Retro Anime Crime Title Sequence

Cinematic
Model
MiniMax H3
Duration
5 seconds
Resolution
768p
Aspect ratio
16:9

“Create a 15-second, 16:9 opening-title sequence for a stylish crime mystery. Draw from retro Japanese animation titles, hard-edged silhouettes, comic-book collage, asymmetric split screens, bold geometric color fields, English credit typography, sparse katakana accents, and a noir-jazz sensibility. Aim for 60% suspense and 40% jazz: mysterious, cool, agile, and urban—never horror, heavy drama, or a cheerful jazz music video. Build the sequence as motion-graphics collage: linework appears over black, split-screen boundaries snap into place, then color blocks and frames assemble piece by piece. Character silhouettes, prop close-ups, and English credits slide, pop, and reveal through masks on the beat. Keep it designed and graphic, not like conventional narrative animation. Credits must be clean and legible. Animate thin frames drawing on, names sliding in, letters appearing one at a time, or color blocks revealing type, then hold briefly. Do not introduce Chinese text, garbled characters, or misspellings. Each role and each name appears once only; do not assign multiple roles to the same name. Vary the transitions: circular vinyl-record wipes, vertical car-door cuts, long character shadows wiping frame, red-line slices, oversized letter masks, split-screen reconfigurations, hard color-block cuts, and tiled frames snapping into place. Keep every transition crisp, rhythmic, suspenseful, and collage-driven. No soft dissolves or fluid morphs. BGM: an original 15-second title cue, 60% suspense and 40% jazz. Use a low drone, tense string pizzicato, cool synth pulses, low kick, sparse brushwork, fragments of walking bass, short baritone-sax phrases, and clipped brass accents. Build suspense with low frequencies and hi-hat in the first 2 seconds; bring in the low beat at 3 seconds, jazz-bass movement at 6 seconds, and a short sax/brass riff at 10 seconds. Freeze the final 2 seconds on a tense chord and drum hit. Keep it mysterious and criminally cool, never upbeat, and do not imitate an existing melody. PRODUCTION CONTROL MODE: Reference to Video. Images 1–5 define graphic language, character silhouettes, palette and credit layout; do not combine identities across references. SHOT PLAN: black-and-white linework → geometric split-screen assembly → prop close-ups → final ensemble title. TYPOGRAPHY LOCK: exact English credit strings supplied in the prompt, one credit per card, minimum one-second hold. MOTION: masks, slides and snap cuts on beat; no fluid morphing. AUDIO: brushed drums, upright bass and clipped brass accents. AVOID: narrative lip sync, 3D rendering, horror imagery and decorative fake text.”

View full prompt

Create a 15-second, 16:9 opening-title sequence for a stylish crime mystery. Draw from retro Japanese animation titles, hard-edged silhouettes, comic-book collage, asymmetric split screens, bold geometric color fields, English credit typography, sparse katakana accents, and a noir-jazz sensibility. Aim for 60% suspense and 40% jazz: mysterious, cool, agile, and urban—never horror, heavy drama, or a cheerful jazz music video. Build the sequence as motion-graphics collage: linework appears over black, split-screen boundaries snap into place, then color blocks and frames assemble piece by piece. Character silhouettes, prop close-ups, and English credits slide, pop, and reveal through masks on the beat. Keep it designed and graphic, not like conventional narrative animation. Credits must be clean and legible. Animate thin frames drawing on, names sliding in, letters appearing one at a time, or color blocks revealing type, then hold briefly. Do not introduce Chinese text, garbled characters, or misspellings. Each role and each name appears once only; do not assign multiple roles to the same name. Vary the transitions: circular vinyl-record wipes, vertical car-door cuts, long character shadows wiping frame, red-line slices, oversized letter masks, split-screen reconfigurations, hard color-block cuts, and tiled frames snapping into place. Keep every transition crisp, rhythmic, suspenseful, and collage-driven. No soft dissolves or fluid morphs. BGM: an original 15-second title cue, 60% suspense and 40% jazz. Use a low drone, tense string pizzicato, cool synth pulses, low kick, sparse brushwork, fragments of walking bass, short baritone-sax phrases, and clipped brass accents. Build suspense with low frequencies and hi-hat in the first 2 seconds; bring in the low beat at 3 seconds, jazz-bass movement at 6 seconds, and a short sax/brass riff at 10 seconds. Freeze the final 2 seconds on a tense chord and drum hit. Keep it mysterious and criminally cool, never upbeat, and do not imitate an existing melody. PRODUCTION CONTROL MODE: Reference to Video. Images 1–5 define graphic language, character silhouettes, palette and credit layout; do not combine identities across references. SHOT PLAN: black-and-white linework → geometric split-screen assembly → prop close-ups → final ensemble title. TYPOGRAPHY LOCK: exact English credit strings supplied in the prompt, one credit per card, minimum one-second hold. MOTION: masks, slides and snap cuts on beat; no fluid morphing. AUDIO: brushed drums, upright bass and clipped brass accents. AVOID: narrative lip sync, 3D rendering, horror imagery and decorative fake text.

Hand-Drawn Kitchen Creature AI video previewCinematic

Hand-Drawn Kitchen Creature

Cinematic
Model
MiniMax H3
Duration
5 seconds
Resolution
768p
Aspect ratio
16:9

“15 seconds, 16:9 landscape. Blend live-action footage of a small kitchen at dusk with hand-drawn luminous animation. The last sunset light lingers at the window. The lived-in kitchen contains an old wooden table, a half-washed mug, a lightly fogged glass bottle, and a hanging dish towel. Shoot as if someone is filming one-handed on a phone: subtle hand tremor, hesitant close-focus pulls, backlit exposure breathing, and slightly coarse noise in the shadows. It should feel like an astonishing event captured in a rush at home, not a carefully dressed commercial. Do not show giant eyes, split mouths, fangs, threatening behavior, lunges, sudden black frames, or jump scares. Use only room tone, cloth friction, a soft mug clink, faucet drips, the camera operator’s footsteps and quiet breathing, plus gentle electronic tones and tiny vocalizations from the drawn creatures. PRODUCTION CONTROL MODE: Text to Video. TIMELINE: 0–4s hesitant kitchen discovery; 4–9s luminous creature crosses the table; 9–13s it reacts to a mug clink; 13–15s operator steadies the phone. CAMERA: one-handed phone, delayed autofocus and exposure breathing. SOUND: room tone, faucet drip, cloth friction, mug clink, footsteps and breathing; only tiny electronic creature tones, no score. CONTINUITY: kitchen layout and sunset direction remain stable. AVOID: giant eyes, fangs, lunges, jump cuts to black and studio lighting.”

View full prompt

15 seconds, 16:9 landscape. Blend live-action footage of a small kitchen at dusk with hand-drawn luminous animation. The last sunset light lingers at the window. The lived-in kitchen contains an old wooden table, a half-washed mug, a lightly fogged glass bottle, and a hanging dish towel. Shoot as if someone is filming one-handed on a phone: subtle hand tremor, hesitant close-focus pulls, backlit exposure breathing, and slightly coarse noise in the shadows. It should feel like an astonishing event captured in a rush at home, not a carefully dressed commercial. Do not show giant eyes, split mouths, fangs, threatening behavior, lunges, sudden black frames, or jump scares. Use only room tone, cloth friction, a soft mug clink, faucet drips, the camera operator’s footsteps and quiet breathing, plus gentle electronic tones and tiny vocalizations from the drawn creatures. PRODUCTION CONTROL MODE: Text to Video. TIMELINE: 0–4s hesitant kitchen discovery; 4–9s luminous creature crosses the table; 9–13s it reacts to a mug clink; 13–15s operator steadies the phone. CAMERA: one-handed phone, delayed autofocus and exposure breathing. SOUND: room tone, faucet drip, cloth friction, mug clink, footsteps and breathing; only tiny electronic creature tones, no score. CONTINUITY: kitchen layout and sunset direction remain stable. AVOID: giant eyes, fangs, lunges, jump cuts to black and studio lighting.

What Is FL2VA?

FL2VA (First-Last-frame-to-Video-and-Audio) is one of two task families in MiniMax H3-Base, the open-weight release of MiniMax's H3 video generation model. FL2VA covers three related workflows in a single checkpoint: text-to-video generation, first-frame-anchored video, and first + last frame transition generation. The companion task family, Ref2VA, handles multimodal reference-to-video with image, video, and audio references.
  • What "and Audio" means: FL2VA can generate synchronized audio in the same pass as video — no separate audio model required.
  • What FL2VA is not: FL2VA is not a separate product from H3. It is a checkpoint family inside H3-Base.
  • Where to get it: Open weights on Hugging Face (MiniMaxAI/MiniMax-H3).
  • When to pick FL2VA over Ref2VA: When your inputs are text and/or one or two anchor images — no reference video, no reference audio.

FL2VA vs Ref2VA — MiniMax H3 Task Families

MiniMax H3-Base ships as two separate checkpoints, each specialized for a different family of video generation tasks. Choosing the right checkpoint before downloading is the single most important decision for local users.

FL2VA versus Ref2VA comparison
DimensionFL2VARef2VA
Full nameFirst-Last-frame-to-Video-and-AudioReference-to-Video-and-Audio
Supported inputsText · First frame image · Last frame imageText · Multiple images · Multiple videos · Audio
Sub-workflowsT2V, First-frame-to-Video, First+Last-frame-to-VideoReference-to-Video, Video-to-Video, Multi-modal reference
Audio inputNo (generates audio, does not accept audio references)Yes (audio reference for voice, music, sound)
Multi-image supportMax 2 (first + last frame, structural role)Up to 9 (identity, style, scene, etc.)
Best forText-driven and frame-anchored generationIdentity- / style- / motion-controlled generation
CheckpointH3-Base FL2VA weightsH3-Base Ref2VA weights
GuideThis pageMiniMax H3 Reference Video
Same base model, two task-specialized checkpoints — not interchangeable at load time.

FL2VA Sub-Workflows

One FL2VA checkpoint, three ways to use it. The distinction is which images (if any) you provide as conditioning.

T2V

Text-to-Video

Only a text prompt, no images. FL2VA generates a full clip from text alone. Best for creative, unconstrained ideation where no visual anchor is needed.

First Frame

First-Frame-Anchored

Upload one image as the opening frame. H3 continues the motion forward from that anchor. Best for animating a portrait, product shot, or concept art.

First + Last

First + Last Frame Transition

Upload both opening and ending frames. H3 generates the visual transition between them. Best for controlled shot transitions, morph-style transformations, and storyboard scenes.

FL2VA Model Card

FL2VA checkpoint specifications
AttributeValue
Checkpoint familyMiniMax H3-Base — FL2VA task family
Full nameFirst-Last-frame-to-Video-and-Audio
Base modelMiniMax H3-Base
InputsText prompt · First frame image (optional) · Last frame image (optional)
OutputsVideo (MP4) · Synchronized stereo audio
Base resolution768p (upscale to 2K via full H3 system)
Duration range4–15 seconds
Precision optionsFP16 · FP8 · INT8 · GGUF (community quantizations)
Hosted accessminimax3.org generator · MiniMax API
Open weightsMiniMaxAI/MiniMax-H3 on Hugging Face
Last verified2026-08-18

Specifications sourced from MiniMax official documentation and the Hugging Face model card. FL2VA weights are part of the MiniMaxAI/MiniMax-H3 repository. Verify the official repository and compare FL2VA with Ref2VA.

FL2VA Hardware Requirements

pikah3.org practical checklist

FL2VA hardware requirements by precision
SetupVRAM FloorTypical GPUNotes
FL2VA FP16≥ 24 GBRTX 4090 / 5090 / A100 / H100Baseline full-precision.
FL2VA FP8≥ 16 GBRTX 4080 / 4090 / 5090Common community setup.
FL2VA INT8 / GGUF≥ 12 GBRTX 4070 Ti / 4080Quantized for mid-range cards.
Apple Silicon≥ 32 GB unifiedM2 / M3 / M4 Pro / MaxCommunity MPS / GGUF builds; typically slower than a dedicated GPU.

VRAM is dominated by the H3-Base FL2VA weights and any additional LoRAs stacked on top. For detailed GPU configurations, see the MiniMax H3 ComfyUI Guide.

How to Run FL2VA in ComfyUI

Step 1 — Download FL2VA Weights

Grab the FL2VA task-family checkpoint from MiniMaxAI/MiniMax-H3 on Hugging Face. Place the weights under ComfyUI/models/checkpoints/. For quantized variants, see the MiniMax H3 Download guide.

Step 2 — Build the ComfyUI Graph

Connect: H3-Base FL2VA loader → (optional LoRA loader) → prompt encoder → sampler → video decoder. First-frame and last-frame image inputs plug into the two image conditioning ports. Leave the last-frame port empty for T2V or first-frame-only workflows.

Step 3 — Sampler and Steps

Use DPM++ 2M / UniPC at 25–40 steps for baseline quality. For faster iteration, stack a Turbo LoRA and switch to an LCM sampler at 4–8 steps. See the MiniMax H3 Turbo guide for acceleration details.

FL2VA and Ref2VA are separate checkpoints — loading the wrong one in your workflow will silently degrade quality or produce unexpected results.

FL2VA Prompt Templates

Each sub-workflow benefits from a different prompt structure. Copy a template below, then customize it for your scene.

Text-to-Video (T2V) Template

Text only

Cinematic wide shot of a coastal cliff at sunrise. Waves crash against the rocks below. The camera slowly pushes forward while morning light breaks through scattered clouds. Seagulls cross the frame from left to right. Add natural wind ambience, crashing waves, and distant seabird calls.

First-Frame-Anchored Template

First frame

Preserve the subject's facial identity, hairstyle, clothing and opening composition. The subject slowly looks toward the camera, blinks naturally and gives a subtle smile. A light breeze moves a few strands of hair while the camera makes a very slow push-in. Keep the facial proportions stable and the motion restrained. Soft room ambience and quiet fabric movement.

First + Last Frame Transition Template

First + Last

Begin precisely from the first image and preserve the main subject's identity and visual design. Move continuously toward the composition shown in the last image. Describe the intermediate transformation through observable changes in pose, environment, lighting and camera position. Avoid unnecessary cuts so the transition remains visually continuous and reaches the supplied final frame naturally.

For the complete prompt structure H3 expects across all workflows, see the MiniMax H3 Prompt Guide.

5 Common FL2VA Mistakes

  1. Downloading Ref2VA weights and expecting first-last-frame behavior — FL2VA and Ref2VA are separate checkpoints.
  2. Uploading only a last frame without a first frame — FL2VA needs a first frame if you want frame anchoring; the last frame is optional.
  3. Using long multi-reference chains with identity, motion, and camera assignments — that is Ref2VA territory, not FL2VA.
  4. Expecting audio input to condition the output — FL2VA generates audio but does not accept audio references as input.
  5. Judging FL2VA quality on the 768p base output alone without going through the H3-Regenerate-2K stage for final quality.

FL2VA Troubleshooting

Common FL2VA issues and fixes
ProblemTry This
Video ignores the last frameEnsure the last frame is loaded into the last-frame conditioning port, not the first-frame port.
Transition looks jarring or cutAdd explicit motion description in the prompt ("continuous camera push", "gradual color shift").
Face identity drifts between first and last frameKeep the two frames stylistically consistent — same lighting, same camera focal length, same subject angle.
Audio is silentConfirm audio output is enabled in the workflow — FL2VA generates audio but it can be muted at the decoder stage.
ComfyUI loads FL2VA but produces noiseVerify you are using the FL2VA checkpoint, not Ref2VA — they are not interchangeable at load time.
Output is blurry at short clip lengthsIncrease steps from 4–8 (Turbo) to 25–40 (standard) for final renders.
First frame composition drifts immediatelyAdd "Preserve the opening composition exactly" at the start of the prompt. Avoid requesting large off-screen actions in the first second.

MiniMax H3 FL2VA FAQ

What does FL2VA stand for?

FL2VA stands for First-Last-frame-to-Video-and-Audio. It is one of two task families in MiniMax H3-Base, the open-weight release of MiniMax H3.

Is FL2VA the same as image-to-video?

They overlap but are not identical. FL2VA is a checkpoint family that covers text-to-video and first/last-frame workflows. Image-to-video is a broader concept that can also be served by Ref2VA when reference images define identity or style rather than anchoring the first frame.

What is the difference between FL2VA and Ref2VA?

FL2VA handles text-driven and frame-anchored generation (text-to-video, first-frame, first+last-frame). Ref2VA handles multimodal reference-guided generation where images, video, and audio each carry separate roles like identity, motion, camera, or voice. They are separate H3-Base checkpoints and are not interchangeable.

Where can I download FL2VA weights?

FL2VA weights are available from the MiniMaxAI/MiniMax-H3 repository on Hugging Face. Look for the FL2VA task-family checkpoint within the repository.

Does FL2VA generate audio?

Yes. FL2VA can generate native, synchronized stereo audio alongside video — that is what the "A" in FL2VA means. However, FL2VA does not accept audio references as input; for audio-conditioned generation, use Ref2VA.

Can I run FL2VA in ComfyUI locally?

Yes. Load the FL2VA checkpoint through a standard H3-Base model loader in ComfyUI, connect image conditioning ports for first and/or last frame, and run the sampler. See the ComfyUI setup section on this page for details.

What VRAM do I need for FL2VA?

FL2VA FP16 needs at least 24 GB VRAM (RTX 4090, A100). FP8 runs on 16–20 GB cards. INT8 or GGUF quantized weights extend to 12 GB GPUs. Apple Silicon needs at least 32 GB unified memory.

Can I use a Turbo LoRA with FL2VA?

Yes, as long as the Turbo LoRA was trained against FL2VA weights specifically. Loading a Turbo LoRA trained on Ref2VA into an FL2VA workflow can silently degrade quality. Match the LoRA to the checkpoint family.

Can FL2VA output 2K video?

FL2VA itself generates at 768p. The complete MiniMax H3 system reaches 2K via a separate H3-Regenerate-2K stage. On pikah3.org, the hosted generator handles the full pipeline including the 2K upscaling stage.

How do I try FL2VA without downloading anything?

Use the generator embedded on this page. FL2VA workflows run through the Image-to-Video tab — upload a first frame and optionally a last frame, write a prompt, and generate. No download or API key required. New users get 10 free credits daily.

Try MiniMax H3 FL2VA

Upload a first frame and an optional last frame, or start with text alone. FL2VA generates video with native audio — all in your browser.

10 free credits per daily check-in · Up to 30 total

Data References

Published 2026-08-18. Last reviewed 2026-08-18. pikah3.org is an independent third-party platform and is not affiliated with, endorsed by, or operated by MiniMax. Specifications are sourced from official MiniMax documentation and the Hugging Face model card.