• Client: None. Self-initiated spec work.
• Role: Everything: concept, prompt architecture, generation, curation, edit, sound, finishing.
• Dates: August 2026. Two working sessions of 4 to 5 hours across a single weekend.
• Format: Two variants from one pipeline. Cinematic (10s) and Toon (15s), 1920×1080.
• Tools: Weavy.ai (ChatGPT Image, Gemini 3 / Nano Banana, Gemini Omni Flash, Seedance 2.0, Seedance 2.0 Mini, Seedance 2.5, Topaz Video Upscaler), After Effects, Premiere Pro.
• Status: Spec. Not commissioned by, affiliated with, or endorsed by Warner Bros. Games.

On the IP: This is unofficial spec work, made independently to test a generative production pipeline. Game of Thrones and Game of Thrones: Dragonfire are the property of their respective rights holders. Nothing here is an official advertisement.

Weavy File for GenAI Video: https://app.weavy.ai/flow/KxGeKciJXg4iWPLxx1ZlKe

The setup

Mobile UA creative runs on volume. You need a lot of variants, fast, and every one has to look like it came from the same game. That second part is where most GenAI ad work falls apart, because each generation starts over on lighting, palette, camera height, and world.

I gave myself a weekend and one constraint: build a prompt structure that holds a whole spot together, then re-skin it into a different art direction without rebuilding anything.

Two finished spots came out of it. Same beats, same copy, same end card, one photoreal and one illustrated.

The system

1. Start with the storyboard

The first prompt asks for the entire ad as a single multi-panel image: isometric camera, stone keep on green terrain, an enemy army advancing, escalating attacks, a dragon, and a final pull-back onto the next wave.

That one generation sets the style. Every panel comes out of the same image, so they already share lighting, palette, camera height, and world. The pipeline starts consistent, which is far cheaper than correcting drift downstream.

2. One prompt, two image models

The same storyboard prompt runs into two models, three images each:

• ChatGPT Image: dark, photoreal, cinematic
• Gemini 3 (Nano Banana): bright, illustrated, stylized

That gives me a second art direction for one node change. Everything downstream inherits its branch, which is why both spots hit identical beats in completely different visual languages.

3. Hold the shot rules constant

A second prompt sits between the storyboards and the video models, carrying the rules that do not change: follow the storyboard, sound effects only, no voiceover, no music, no on-screen text, no logos.

Blocking music and voiceover at the model matters most. Clean SFX-only audio lets me move, trim, and reorder scenes in the edit without a baked-in music bed fighting me. Text works the same way. Copy and end cards belong in After Effects, where I can edit, resize, and localize them. Bake a hook into the render and every copy test costs another generation.

4. How the models take direction

Every video node is reference-to-video off a storyboard panel.

That ruled out first-frame and last-frame conditioning, which is the wrong control surface here. I did not want to specify where a shot starts and ends. I wanted the model to read a panel as style and staging reference, then interpret the motion itself. Seedance takes reference images that way, so Seedance carried most of the work: 2.0 for hero shots, 2.0 Mini for cheap coverage while exploring, 2.5 and Gemini Omni Flash for comparison passes.

Choosing a model for its conditioning method is most of what made the weekend work.

5. Why Weavy

Weavy exposes the generation graph as a shareable link. Every node, prompt, and discarded branch stays visible to anyone I send the URL to.

That was a deliberate trade. Other platforms aggregate the same models behind a cleaner interface, but they show finished outputs without the reasoning underneath. Here the pipeline is the deliverable, so I picked the environment that could show its own work.

None of these methods depend on Weavy. The storyboard style lock, the two-model fork, the reference-conditioning rule, and the shared overlay layer all rebuild in any multi-model environment. Weavy is the version you can audit.

6. Standardize the finish

Selected clips run through Topaz Video Upscaler under one naming convention, straight into Premiere. Uniform resolution and consistent file names keep the edit from stalling on file management.

7. The reusable layer

Copy overlays and the end card are After Effects comps, built once and used in both variants. “Defend the Keep,” “Tap to Upgrade,” “Swipe to Strike,” and the Eyrie/Caraxes end card are identical across both cuts. A new variant reuses the whole overlay layer and changes only the footage branch.

The edit

Each spot is cut from one selected generation, so the editing carries real weight.

Cinematic (10s) takes a single generated take and cuts it into roughly nine beats, averaging just over a second each. Photoreal generative footage holds up best in short bursts, before temporal artifacts have time to surface. The cutting rhythm is doing quality control.

Toon (15s) leaves its take almost intact as one continuous run. Illustrated styles survive a long shot because there is less for the model to break. Same beats, opposite approach.

The copy hooks are mechanic-first on purpose. “Defend the Keep,” “Tap to Upgrade,” and “Swipe to Strike” each name an action the player takes. In a real test, those are the first variables I would rotate, and they sit in their own comps so rotating them costs nothing.

Audio is three stems: an SFX bed generated by Seedance, a music track, and an end-card sting. The SFX are a time trade. On a two-day build, model-generated sound got me to a finished cut fast. With a real schedule, I would replace them from my own library and design the hits properly.

Results

This is spec work, so there is no campaign data, and I am not going to invent any. What it shows is throughput and curation.

• Two finished, 1920×1080 spots in two 4 to 5-hour sessions across one weekend, solo, from a blank canvas.
• 24 generations, 2 selects. Six storyboard images and eighteen video generations across four video models, narrowed to one hero clip per variant. Most of the pipeline is discarding.
• A second art direction for one node change. The Toon variant needed no new brief, no new storyboard prompt, no new overlays. It reused the structure and swapped the image model.
• An auditable pipeline. The Weavy tree is open: every prompt, every node, every branch that did not make it.

What I learned

Consistency is structural. Generating the storyboard as one image solved more drift than prompt tuning would have, because it puts consistency upstream of generation.

Storyboard adherence beat my expectations. Seedance costs a lot per generation, but it followed the panels closely and at high enough quality that I stopped budgeting for a re-roll pass. That is the only reason two spots shipped instead of one.

Duration is a cost decision. Holding the spots to 10 and 15 seconds instead of 30 cut credit burn enough to fund the second variant outright.

Cut length doubles as quality control. Matching editorial pace to how long a model’s output actually holds up is what keeps the footage reading as finished.

What I’d test next

Vertical and square resizes (1080×1920, 1080×1080) off the same overlay comps, hook-order rotation across the three mechanics, a designed SFX pass to replace the generated one, and a third art-direction branch to confirm the fork scales past two.