MiniMax Hailuo Image-to-Video Tips: The 70/30 Prompt Rule and Motion Control

Why Short Drama Creators Rely on MiniMax H3

This article was written by the Say Action team. The recommendations and analysis reflect our workflow experience in short drama production, not an independent third-party benchmark.

When producing episodic vertical dramas under tight delivery schedules, creators frequently encounter a major obstacle with AI video generators: unnatural facial expressions and rubbery motion during dialogue. MiniMax H3 (often surfaced as Hailuo AI) has gained widespread adoption across drama teams specifically because of its nuanced micro-expressions, coherent cloth physics, and stable cinematic lighting. Subtle human reactions—such as hesitant eye shifts, controlled breathing, and micro-smiles—render with convincing anatomical weight rather than melting into uncanny digital noise.

This distinct strength separates MiniMax H3 from competing models like Kling or Seedance in practical shot design. Where Kling AI excels at expansive acrobatic stunts or rapid camera pans, and Seedance provides unmatched multi-frame identity tracking across extended sequences, MiniMax H3 dominates mid-shots, close-ups, and dramatic character exchanges. When fed a pristine starting image, its latent diffusion engine respects existing highlights, fabric textures, and background depth without aggressive hallucinations or plastic skin smoothing.

On the Say Action canvas, directors leverage this model within a modular node pipeline. Rather than expecting a single prompt to invent an entire dialogue scene out of thin air, creators generate keyframe portraits using dedicated image models like Nano Banana, lock the exact framing and costume details, and connect the resulting keyframe node directly into a MiniMax H3 video generation block. This separation of visual staging from temporal movement guarantees predictable shot delivery.

The 70/30 Prompt Blueprint for Image-to-Video

A frequent trap in image-to-video generation is prompt redundancy. Novice creators often write lengthy literary descriptions re-explaining the character's clothing, room architecture, and color palette. When the diffusion model attempts to synthesize instructions that already exist inside the reference pixels, it creates conflicting vectors in latent space, frequently triggering distorted limbs, dissolving backdrops, or sudden costume alterations. Conversely, writing only an isolated motion prompt like "she turns around" causes the camera to drift aimlessly.

As a recommended rule of thumb, the 70/30 prompt blueprint addresses this challenge by balancing prompt focus between two distinct functional zones. Roughly seventy percent of the prompt text helps anchor static reference elements—confirming the exact subject features, wardrobe silhouette, and lighting temperature inherited from the input image. The remaining thirty percent guides the kinetic envelope—dictating the actor's deliberate movement, gaze vector, and camera constraint. Under this guideline, avoid asking the video model to introduce unestablished narrative props or sudden architectural shifts.

Consider a practical comparison in prompt construction. An undisciplined prompt such as "Cinematic dramatic 4k beautiful young woman crying in the rain and camera spins around dramatically" invariably yields warped facial geometry and chaotic framing. By contrast, a structured 70/30 prompt reads: "Anchored to keyframe: young woman in emerald silk blouse, soft directional backlight, shallow depth of field. Kinetic action: slow single tear rolls down left cheek, gentle swallow, eyes glance downward toward the phone screen; locked-off camera holds static framing." This gives the diffusion network concrete spatial and temporal boundaries.

When writing the motion portion of the prompt, professional cinematographers employ micro-kinetic verbs rather than emotional adjectives. Words like "furious" or "heartbroken" produce random gesticulating and grimacing; specific kinetic directives like "subtle jaw clench", "measured breath intake", or "slight 5-degree head tilt to the left" provide reproducible physical coordinates. In vertical 9:16 drama production, these deliberate micro-actions maintain character believability without tearing the image canvas.

Crucial Constraint: No Mixed Reference Modes in API

Behind the Say Action interface, the MiniMax provider adapter enforces strict upstream API contracts. The single most common failure reported by technical creators is attempting to supply character reference sheets alongside starting and ending frame anchors. In the official MiniMax video generation engine (video-provider-minimax.ts), whenever frame mode is activated (first_frame or last_frame), passing additional generic reference images throws an immediate engine validation error (minimaxMixedFrames).

This operational constraint is rooted in mathematical diffusion mechanics. When an AI video generator operates in frame interpolation mode, the starting and ending images serve as immutable Dirichlet boundary conditions for the temporal denoising process. Introducing floating character reference sheets injects contradictory latent noise into intermediate frames, resulting in grotesque identity blending where faces dissolve between the reference sheet and the anchor frames.

To build a robust drama pipeline on Say Action, production teams must select one of two explicit operating modes per shot. In Mode A (Single Start Frame Image-to-Video), you supply exactly one master keyframe into the First Frame input slot, using prompt guidance to direct forward motion. In Mode B (Generative Frame Transition), you supply an established First Frame and a coherent Last Frame to bridge two camera states. In neither mode should you attempt to attach supplementary character turnarounds into the same generation task.

Maintaining clean node hygiene on the infinite canvas prevents accidental job failures. Before triggering batch generation across a sequence of storyboard rows, inspect the video node configuration drawer. Ensure that the Cast Reference connection is disconnected whenever dual-frame interpolation is active. By honoring this upstream architectural boundary, your rendering queue proceeds without unexpected job rejections or pipeline halts.

High-Resolution Keyframe Anchoring

In image-to-video pipelines, temporal coherence is strictly bounded by input clarity. If your initial keyframe contains fuzzy edges, compression artifacts, or poorly resolved facial contours, the MiniMax video diffusion model will amplify these imperfections into severe subject warping within the first twenty frames. Maintaining actor likeness across an entire episode requires establishing flawless static anchors before rendering motion.

Within the Say Action workspace, creators generate base keyframes using high-fidelity image models like Nano Banana or FLUX, ensuring the subject's gaze direction, costume textures, and background separation are locked down. Avoid using screen grabs from low-bitrate video clips or blurry mid-motion drafts as starting frames. A crisp, high-resolution portrait with balanced rim lighting provides the diffusion network with unambiguous geometric features to preserve throughout the shot.

Unintended camera drift represents another frequent source of shot destruction. MiniMax H3 is exceptionally sensitive to environmental descriptors; writing descriptive words like "wind rustling through trees" or "turbulent ocean waves" often causes the AI camera operator to pitch or sway uncontrollably. To enforce rock-solid stability in dialogue shots, always prepend explicit camera stabilization instructions such as static locked-off medium close-up, tripod mount, or camera holds stationary position.

Finally, incorporate a disciplined human quality assurance (QA) checkpoint into your canvas workflow. When a batch of video nodes completes rendering, review the initial half-second of playback to verify that facial symmetry remains intact. If an actor's likeness shifts during motion, make small adjustments to prompt kinetic intensity or replace the start frame on that isolated node, rather than regenerating the entire scene or altering surrounding storyboard rows.

Resolutions, Durations, and Credit Economics

Managing compute budgets while meeting release schedules is essential for commercial vertical drama production. In Say Action, the MiniMax H3 adapter provides two standardized resolution targets—768P for rapid pre-visualization and 2K for master delivery—alongside discrete clip durations spanning 4 to 15 seconds. Choosing the right configuration for each storyboard row prevents unnecessary credit depletion while preserving cinematic punch.

Below is the verified operational breakdown for MiniMax H3 configurations on Say Action canvas, comparing visual stability, typical duration envelopes, and recommended staging roles.

Platform credit economics are transparently displayed before any rendering job commences. In Say Action, estimated compute points are calculated based on chosen resolution and duration prior to submission. When an upstream provider explicitly reports a generation failure or terminal timeout, reserved platform credits are automatically returned to your team balance (When the upstream explicitly reports failure, reserved credits are refunded; tasks with unclear network status stay pending review to avoid double settlement.). For creators operating under Bring Your Own Key (BYOK) mode, raw model tokens are billed directly by MiniMax with zero platform markup (BYOK supports a custom API endpoint, model key and headers. For OneAPI, NewAPI or another relay, select the OpenAI, Gemini or Claude protocol that matches its actual API, then check model IDs and image/video endpoint support. Not every relay or private protocol works automatically. BYOK model calls have no platform model markup or platform generation-credit deduction; workspace access and provider usage are still billed separately.), while workspace access remains covered by the standard studio subscription.

A proven production strategy for 60-to-90-second drama episodes is two-pass rendering. During the first pass, generate all 10 to 14 storyboard cuts at 768P with 5-second durations to test scene continuity and dialogue cadence. Once the director approves the edit timeline, selectively re-render only the 3 or 4 pivotal emotional close-ups at 2K. This balanced methodology keeps production costs predictable while ensuring festival-grade visual fidelity where it matters most.

MiniMax official API parameters verified on 2026-10-11; platform credit estimation and deduction are displayed prior to generation, with automatic refunds for confirmed job failures.

Connect your own API channels and third-party relays

In model-channel settings, enter the Base URL and API key, then configure the protocol, model IDs and required headers from your provider's documentation. Check the image and video modes you need before generating. Paths, resolutions and first/last-frame inputs depend on the connected channel. An adaptation request for an unsupported channel is not a compatibility guarantee.

BYOK supports a custom API endpoint, model key and headers. For OneAPI, NewAPI or another relay, select the OpenAI, Gemini or Claude protocol that matches its actual API, then check model IDs and image/video endpoint support. Not every relay or private protocol works automatically. BYOK model calls have no platform model markup or platform generation-credit deduction; workspace access and provider usage are still billed separately.

Connect only trusted providers and use HTTPS in production. Custom requests are forwarded through the authenticated backend; required prompts, reference assets and credentials take part in the request. Never put API keys, passwords or secret-bearing headers in public canvases, screenshots or support requests.