Breaking Down the Process: AI-Driven Action Sequence

This project started with a simple idea: create a cinematic, grounded action sequence—set in real-world Chicago—featuring a fully AI-generated version of myself, a custom-designed mech suit, and a large-scale creature.

What made it interesting wasn’t just generating shots—it was building a pipeline that allowed for consistency across characters, environments, and motion.

Character Creation

1. Digital Double (Myself)

To create a consistent version of myself across shots, I trained a FLUX LoRa on a curated dataset of my face and body.

Using a custom ComfyUI setup, I generated controlled outputs:

• Full body (multiple angles)

• Head turnarounds

• Expression variations

This gave me a reliable “actor” that could be dropped into any shot while maintaining identity and proportions.

2. Mech Suit + Creature Design

The mech suit and creature were developed in MidJourney through heavy iteration—pushing for:

• Mechanical realism (for the suit)

• Anatomical weight and texture (for the creature)

Once I landed on designs I liked, I ran those images through a custom NanoBanana + Weavy to generate:

• Multiple angles

• Pose variations

• Structural consistency

From there, I assembled everything into 2048x2048 character sheets, which became the backbone of the entire pipeline.

These sheets were included in every Seedance 2.0 prompt to maintain:

• Design consistency

• Proportions

• Material detail

Smaller elements like the watch transformation trigger and wrist-mounted weapon system were treated as their own assets—designed and referenced independently to ensure clarity during close-ups.

Environment Creation (Chicago as a 3D Backlot)

Instead of relying on purely generated environments, I used Google Earth as a base.

The approach:

1. Navigate to specific locations in downtown Chicago

2. Frame exact camera angles for each shot

3. Capture perspective-based images

Then in NanoBanana, I enhanced those images using:

• High-end architectural photography references

• Clean lighting and material refinement

The result was a workflow where I could effectively place a virtual camera anywhere in the city and generate a cinematic-quality plate.

Animation & Shot Generation

Shot generation was handled across a combination of Seedance 2.0, Veo3, and Kling, depending on what each moment required.

• Seedance 2.0 was the primary tool for structured, cinematic shots where I needed tighter control over:

  • Camera movement

  • Character consistency

  • Framing and staging

• Veo3 was mainly used as an iterative tool—when I needed to explore ideas through rapid trial and error. It allowed me to generate many variations quickly, helping dial in:

  • Motion ideas

  • Timing

  • General shot direction before refining elsewhere

• Kling was used in select cases where it handled motion or temporal coherence better for certain types of movement.

Across all tools, I maintained consistency by:

• Feeding in the same character sheets

• Using environment references as anchors

• Structuring prompts more like shot direction than simple descriptions

This process was less about “prompting” and more about direction—starting with a clear vision for each shot, then executing it using a combination of carefully chosen references.

Upscaling & Motion Refinement

Once the sequence was assembled:

Topaz Video AI

• Upscaled footage (720 → 1080)

• Reduced temporal artifacts / “choppiness”

• Improved motion interpolation

Nuke (Final Polish)

Final compositing and polish were handled in Nuke, where I brought everything together and refined the sequence:

• Color correction

• Time remapping for pacing

For the final hit, I needed a clean projectile element. Instead of sourcing stock footage, I generated a custom missile element in Veo on a green screen. This gave me full control over:

• Angle and perspective

• Motion characteristics

• Lighting consistency

I then:

• Keyed the element in Nuke

• Integrated it into the shot of the creature

• Matched lighting, motion blur, and scale

• Blended the impact so it felt physically grounded in the scene

This step helped sell the final moment—tying together the generated character, environment, and action into a cohesive shot.

Final Thoughts

This project wasn’t about a single tool—it was about building a pipeline where each tool handled a specific role:

• FLUX LoRa + ComfyUI → Identity & performance

• MidJourney + NanoBanana → Design & consistency

• Google Earth + NanoBanana → Real-world environments

• Seedance 2.0 → Shot creation

• Topaz + Nuke → Final polish