This started with a Pinterest find. It ended with a production-ready pipeline that can generate a gamified character roster for any sports franchise—consistently, at scale, without manual illustration.
Here is exactly how I did it.
⚠️ Disclaimer: Personal portfolio project and fan art. No official affiliation with USA Basketball, FIBA, the NBA, or any player's personal brand.
Step 1: Finding the Spark
I spotted a chibi yoga character set on Pinterest. The style immediately clicked: clean proportions, expressive faces, strong silhouettes. To see if this aesthetic could hold up on real athlete faces, I stress-tested it against the 12-player USA Basketball roster.
Step 2: Hitting the Generative Wall
I fed the Pinterest reference into Nano Banana 2 with a basic character prompt. The early results were promising, but three critical problems surfaced fast:
- Art Direction Drift: Proportions were too vague. The model wandered, delivering inconsistent head sizes and body ratios.
- Jersey Complexity: The USA Basketball kit has complex textures, gradients, and player numbers. Text prompting led to sloppy approximations.
- Generic Faces: Chibi illustrations naturally flatten features. Getting the model to capture unique identifiers (like Bam’s headband or Durant's narrow frame) required more than a text description.
Step 3: Building the Visual Guardrails
Instead of fighting the model with endless text prompts, I engineered dedicated visual references directly into the pipeline:
- For Proportions: I created a detailed character spec sheet defining exact ratios (head-to-body, eye distance) relative to a base unit $H$. This gave the model a concrete blueprint.
- For the Jersey: I built a template jersey pre-rendered in the target art style. I fed this image to the model, forcing it to replicate the exact patterns and textures.
- For Facial Features: I adjusted the prompt parameters to forcefully weight the player’s actual facial structure, extracting their specific markers instead of drawing a generic placeholder face.
Step 4: Automating and Scaling
With the guardrails locked in, I built a modular Weavy pipeline using routers for our visual assets (Player Reference, Art Direction, and Jersey). To scale across all 12 players, I just needed to swap out the player reference image and jersey number—the underlying system did the rest.
Step 5: Final Model Selection & Upscaling
After benchmarking multiple options, Nano Banana 2 was selected for the generation phase because it offered the highest style stability across different skin tones and body builds. From there, the approved lineup was pushed through Magnific Precision V2 for production-ready upscaling.
One Pipeline. Infinite Touchpoints.
Once you lock down a consistent 12-player asset sheet, the commercial applications are endless:
- Digital Trading Cards: Full-bleed art with stats and holographic overlays.
- Social Engagement: Custom sticker packs for WhatsApp, Telegram, and iMessage.
- AR Lenses: Interactive filters for TikTok, Snapchat, and Instagram.
- Broadcast Graphics: Player intros, lower-thirds, and transition cards for OTT platforms.
- Physical Merch: Proportions are pre-optimized for 3D collectible toy manufacturing (Funko Pop style).
The System is the Deliverable
The characters are the proof; the pipeline is the product. By swapping out the underlying references, this exact workflow can be deployed for any sports roster or entertainment franchise globally.