Fifty Shots With the Same Fake Bird: A Mascot Consistency Field Test
Every brand mascot has a tell. For our fictional mascot, it was the blink. June Peck is a park-ranger cardinal for Roost & Ramble, a made-up US outdoor gear company with a catalog full of waxed canvas and overpriced trail mix. We wrote a character bible with one weird rule: June does not blink. It made her look like she was staring through your soul in a friendly way. It also gave us an easy way to catch identity drift in generated video.
The Character Bible
We gave June a fixed silhouette: cardinal body, round head, short beak with a tiny notch on the left side, mustard rain jacket, canvas satchel, brass pin shaped like a pinecone. No human hands. No extra accessories. Height relative to a campsite chair. Expression range from mildly concerned to deeply unimpressed. Palette locked to six swatches. The rule was not 'make a cute bird.' The rule was 'make this exact bird in a new situation.'
The Reference Kit
Before generating anything, we made a six-image reference kit: front, three-quarter left, profile, back, expression sheet, and a prop detail shot. Neutral background, flat lighting, no dramatic shadows. We also made one full-body scale reference next to a chair. That kit mattered more than any prompt.
For stills, we used image-to-image and character-reference slots where available. For video, we started from an approved still and used image-to-video. Text-to-video was a loss almost every time. The model would invent a bird that was close, then lose the jacket by second three. Reference images are not a suggestion; they are the character.
The Prompt Pattern That Held
We stopped writing paragraphs and started writing blocks. Character block first, scene second, camera third, negatives last. Something like: 'June Peck, female cardinal mascot, mustard rain jacket, canvas satchel, pinecone pin, round head, short beak with left notch, no human hands, same character as reference. Standing beside a camp stove at dawn, three-quarter view, 35mm lens, soft overcast light. No blinking, no extra limbs, no logo changes, no costume variation.'
The useful part was repetition. Every shot repeated the same identity phrases verbatim. We did not paraphrase 'mustard rain jacket' into 'yellow coat.' We did not let the scene description rewrite the character. For video, the motion prompt stayed short: 'June turns her head toward the stove, jacket shifts slightly, background campers stay still.' The more we described, the more the model had to invent.
What Broke
Across 50 shots in November 2026, 31 passed our first review. Nineteen needed another pass. Seven never became usable. The failures clustered.
| Failure | Count | What it looked like |
|---|---|---|
| Face drift | 9 | beak widened, eyes moved, head shape softened |
| Costume drift | 12 | jacket shifted orange, satchel vanished, pin turned into a leaf |
| Scale drift | 5 | June became pigeon-sized or chair-sized |
| Blink or expression pop | 4 | eyelids appeared in video |
| Wing anatomy | 3 | extra feather joints |
| Background mutation | 6 | props changed between frames |
The blink failure was the most entertaining. Four video clips gave June eyelids. Not a full blink, just a slow, judgmental shutter. In one clip she blinked at the exact moment the logo appeared, which felt like a brand strategy statement. We rejected it.
Costume drift was more expensive. A mustard jacket is a specific yellow-brown. Several models read it as safety orange. The pinecone pin became a leaf, a star, or once a tiny waffle. The satchel disappeared whenever the camera moved behind her. Reference kits help, but they do not fully solve small prop consistency.
The 50-Shot Scorecard
Stills held up better than motion. First-pass approval was 72 percent for stills and 48 percent for video. Close-ups were easier than wides. Wide shots with multiple characters were the hardest because the model had more room to invent. Short clips under four seconds were noticeably more stable than eight-second clips. Camera moves were the real trouble: a slow push-in could turn June's beak into a duck bill by the final frame.
The best results came from a boring loop. Approved still, short motion prompt, first-frame lock, human review. The model did not need to know June's backstory. It needed to know her jacket color and beak shape.
The Workflow We Would Ship
Use one approved hero image per costume. Keep a contact sheet. Lock the character block. Generate stills before motion. Use first-frame locking. Reject drift early. Keep video shots under four seconds. Review at 100 percent, not on a phone. Log seed and reference IDs. Add a costume check before publishing.
Consistency Checklist for Recurring AI Characters
- Write a character bible with fixed proportions, colors, and props.
- Build a six-image reference kit, including back and prop details.
- Repeat the identity block verbatim in every prompt.
- Use image-to-image or image-to-video; avoid pure text-to-video for hero shots.
- Lock first frames for video and keep shots short.
- Add negative prompts for blinking, extra limbs, costume changes, and logo edits.
- Review at full resolution for small props and color.
- Keep a contact sheet of approved shots for the next batch.
- Version the character bible when the design changes.
- Schedule a human consistency pass before any paid media.
June is not real. But the workflow is. The mascot does not blink, and now neither does our review process.
