GPT Image 2.5 + MiniMax H3 Tutorial: Turn a 4-Panel Storyboard into a Coherent 12-Second 2K Short Film

Key Takeaway (TL;DR): The biggest hurdle in AI video generation is character face distortion and chaotic temporal pacing. By pairing [GPT Image 2.5 for spatial storyboard locking + MiniMax H3 for temporal camera direction], we converted a single 4-panel grid image into a coherent 12-second, 4-shot 2K cinematic short film. The narrative follows a clockwork courier approaching a rooftop planter, retrieving a glowing seed, planting it, and watching a luminous flower bloom. The entire pipeline cost 257.99 credits ($1.29 USD), providing an ultra-cost-effective production model for indie filmmakers and creative agencies.

---

1. Final Short Film Screening

Watch the resulting short film rendered by MiniMax H3 (optimized as a silent 720p web preview below):

• 🎬 Download full-resolution 2560 × 1440 (2K) 24 FPS master MP4
• 📥 Download 720p silent web preview MP4

Technical Master Specs: Requested 12s duration, container output 12.25s, 2560 × 1440 resolution at 24 FPS with full camera continuity across all 4 story beats.

---

2. The Core Pipeline: Why Storyboard Conditioning Works

Direct text-to-video models frequently alter costumes and morph character facial structures between cutaways. Our structured pipeline decouples spatial asset design from temporal motion:

  [Step 1: Script Timing]  ──> Divide 12s into four 3-second visual beats
             │
  [Step 2: GPT Image 2.5] ──> Generate 2×2 storyboard grid (locking character, props, lighting)
             │
  [Step 3: MiniMax H3]    ──> Animate storyboard as a 4-shot full-screen sequence in reading order

Because all four panels are generated together in GPT Image 2.5, character costume geometry, color palette, and rooftop environments remain identical. MiniMax H3 then interprets the grid as sequential story beats rather than hallucinating new character designs.

---

3. Step-by-Step Implementation Guide

Step 1: Script a 4-Beat Narrative

Focus the 12-second sequence on one character, one consistent location, and one transforming focal prop:

Timecode	Shot Scale	Visual Action & Beat
0–3s	Wide Establishing	Courier walks across evening rooftop toward empty terracotta pot
3–6s	Medium Close-Up	Courier reaches into satchel and lifts a luminous glowing seed
6–9s	Extreme Close-Up	Hands gently place the glowing seed into dark soil
9–12s	Medium Wide	Golden flower sprouts instantly; courier watches in quiet awe

Step 2: Generate the 4-Panel Grid with Sunburst

Use gpt-image-2.5-sunburst at 2K 16:9 to establish cinematic stop-motion aesthetics:

• 📥 Download original 2048 × 1152 storyboard PNG
• 📋 Inspect original image generation input JSON

Prompt Best Practice: Specify “2 by 2 storyboard grid, thin dividing gutters, sequential reading order” and explicitly forbid panel numbers or comic speech bubbles.

Step 3: Direct MiniMax H3 via Reference Conditioning

On PoYo, MiniMax H3’s API model identifier is hailuo-03.

⚠️ Critical Video Engineering Trap: Do NOT pass the storyboard into start/end frame fields (image_urls). Passing a grid as a start frame forces the model to animate the grid lines! The Correct Technique: Pass the storyboard into reference_image_urls (global reference mode) and explicitly instruct the model: “Image 1 is a 4-panel storyboard reference. Make a 12-second short with four consecutive full-screen shots in reading order. Never show the grid or gutters!”

Core Video Prompt:

Image 1 is a four-panel storyboard reference, NOT a single frame
to animate. Make a 12-second 16:9 silent cinematic short with four
consecutive full-screen shots in reading order. Never show the grid,
gutters or multiple panels.

0-3s: courier walks toward the empty pot.
3-6s: courier reaches into the satchel and lifts one glowing seed.
6-9s: seed is gently planted in the dark soil.
9-12s: a luminous golden flower sprouts; courier watches.

Preserve the character, clothes, pot, rooftop and blue-hour light.
No subtitles, written words, logo, dialogue or standalone outro.
The final frame stays on the courier and flower.

PoYo API Request Payload:

{
  "model": "hailuo-03",
  "input": {
    "prompt": "Your complete sequential-shot instructions",
    "reference_image_urls": ["https://storage.poyo.ai/blog/gpt-image-25-p0-20260910/storyboard-c996b2d10610.png"],
    "duration": 12,
    "resolution": "2K",
    "aspect_ratio": "16:9"
  }
}

• 📥 Download complete video generation JSON request

---

4. Second-by-Second Visual Inspection

We extracted 12 representative frames (sampled at 1 FPS) into an inspection contact sheet:

Read left to right, then top to bottom (Seconds 1 through 12).

Strengths & Current Limitations:

• ✅ Flawless Camera Sequencing: The video broke free from the grid completely, cutting cleanly between 4 full-screen angles matching the storyboard beats.
• ✅ Character Consistency: Navy cap, teal coat, scarf, brass details, and rooftop backdrop stayed consistent without facial collapse.
• ⚠️ Mechanical Realism: The glowing seed remains visibly oversized and finger contact with soil exhibits minor clipping. High-end commercial spots may still benefit from localized VFX cleanup.

---

5. Production Cost and Turnaround Breakdown

Production Asset	Model & Parameters	Charged Credits	Cost in USD ($0.005/credit)	Measured Turnaround
1. 4-Panel 2K Storyboard	GPT Image 2.5 Sunburst (2K medium)	5.99 credits	$0.02995	42.97 s
2. 12-Second 2K Short Film	MiniMax H3 / hailuo-03 (2K 12s)	252 credits	$1.26000	621.46 s (~10 min)
Total Production Cost	Complete 12-Second 2K Short	257.99 credits	$1.28995 (~$1.29)	~11 minutes

---

Frequently Asked Questions (FAQ)

Can GPT Image 2.5 generate video files directly?

No. GPT Image 2.5 is a static image model that excels at spatial character design and multi-panel storyboarding. MiniMax H3 provides the temporal motion and camera dynamics.

Do I need to split the storyboard into 4 separate image uploads?

No. Passing the entire 4-panel grid as a single URL into reference_image_urls works out of the box. Modern video models possess the spatial attention needed to parse multi-panel grids sequentially.

How do I get started?

Design your storyboard concept on GPT Image 2.5, then submit the image URL to MiniMax H3.