MiniMax H3 Max guides document start-end frames and motion-control timelines

• H3 Max work is splitting into three practical paths, text-to-video, image-to-video, and start-end-frame generation, as the three-mode H3 Max test makes explicit.
• Start and end stills are being used as scene anchors, with the start-end-frame workflow reporting all of its scene changes from one generation instead of an edit.
• Motion-control timelines now put multiple H3 shot types in one project, according to the H3 Motion Director demo.
• The strongest prompts read like timed shot lists with continuity constraints, a pattern visible in the fish-market storyboard.
• Fast previews are enabling that level of specification: a 15-second timing test reported a 15-second video rendered in under 15 seconds.

fal’s launch post calls H3 Max a post-trained MiniMax H3 optimized for speed, while ComfyUI’s H3 tutorial lays out the three input modes. A community-built Motion Director adds a mixed timeline, continuity controls at segment boundaries, and selective reruns, which is the interesting part for anyone making more than one shot.

The three-mode H3 Max test labels its three approaches directly: start-end frame, image-to-video, and text-to-video.

• Text-to-video: a prompt supplies the subject, camera, action, sound, and pacing.
• Image-to-video: a supplied still establishes a visual starting point, with optional first-and-last-frame control in ComfyUI’s H3 tutorial.
• Reference-to-video: images, video, and audio become conditioning material for identity, style, motion, camera, or voice.

fal’s image-to-video API treats image_url as optional, so a text-only request can route through the same surface to text-to-video.

Speed remains the launch pitch. an H3 Max launch post described the model in terms of quality, latency, and price.

Creator reports span different clips and settings: the 15-second timing test reported a sub-15-second render, a 30-second generation post put its result at one minute, and a detailed H3 Max video review collected additional launch tests.

The start-end-frame workflow set both endpoint stills before generation and said every scene change came from one pass.

That turns the prompt into a description of the journey between two fixed compositions. a four-clip start-end test generated four 15-second segments, then joined them with Premiere Pro Morph Cut before adding color, halation, grain, and a Topaz 1080p, 60fps upscale.

The same continuity idea scales beyond endpoints. the Meme City run asked an agent to place multiple meme characters on the same New York street under the same lighting, then transition between them without cuts in one continuous shot.

The H3 Motion Director demo points to a timeline-based control layer for character work.

The Motion Director project describes three mechanics:

• Mixed T2V, I2V, FL2V, R2V, V2V, and RV2V segments in one production timeline.
• Visual and generated-audio continuity controls at each segment boundary.
• Selective reruns, so only chosen segments are regenerated.

the Visionary workspace takes a related route: its H3 interface replaces the conventional prompt box with a document containing shots, cut times, speaker IDs, and per-subject retention.

The fish-market storyboard uses five timed three-second shots to move from a vendor’s hand to an upside-down flooded boulevard, while keeping its action escalation inside one generation.

Magnific’s timing notes reduce that construction to four rules: number and timestamp the shots, hold one element constant, escalate one variable, and reserve the reveal for the final two seconds.

Three published prompt structures show how far that idea can stretch:

• the Chibi Commander prompt maps a 15-second music-video passage in six 2.5-second beats and locks character proportions, outfit, color, and illustration style.
• the CRT attract-mode template uses five reference images, five three-second beats, alternating formation direction, and a fixed RGB-scanline treatment for every morph.
• the robot-fighter prompt uses five three-second shots, hard cuts, named camera moves, and specific sound effects.

a five-prompt H3 Max thread adds dialogue, shot counts, sound design, and negative constraints to its short-form scenes. the detailed H3 Max prompting guide is part of the same growing pile of creator documentation.

The paper-character workflow builds characters in Nano Banana 2, places those image references into scenes, then animates them in H3; its final stage uploads and trims audio through a media extractor.

The associated paper-character post calls the stills-to-animation handoff “super simple,” but its thread shows the asset preparation behind it. A separate stop-motion recipe in the felt-animation prompt specifies felt, wool, wire armature, 12fps judder, fingerprints, seams, loose fibres, a small set, and practical lights.

The long beginner H3 node guide breaks the local workflow into the parts hidden behind the creator-facing templates.

• A MiniMax H3 Reference to Video or Image to Video node prepares conditioning and the empty video latent.
• The VAE compresses image, video, and audio inputs into a latent representation, while the text encoder turns prompt language into conditioning.
• SamplerCustomAdvanced starts from noise and walks the latent through scheduled denoising steps; the guide describes small early changes as the phase that establishes scene layout and motion.
• The guide’s author uses a beta57 scheduler at 33 steps and CFG 1, framing the latter as appropriate for H3’s distilled setup and avoiding a negative-prompt pass.