Monday I tested text rendering. Tuesday I tested genre intelligence. Both used photorealistic or semi-realistic styles, and the reference image held. Today I pushed into flat graphics, tattoo art, chibi, and streetwear. And on the very first variation, the reference image lost.

I tried to put my AI mascot on a t-shirt and lost him completely.

My first variation was a bold graphic t-shirt design. White background, geometric style, flat illustration, "PRISM" in block letters. The graphic quality was great. Clean edges, strong contrast, print-ready.

But the character was gone. Four images, zero recognizable as Prism. NB2 saw "flat graphic illustration style" and "geometric" and decided to build a new character from scratch rather than translate the existing one. The reference image was completely overridden by the style instructions.

Yesterday I noted that genre conventions can occasionally overpower the reference image. Today that "occasionally" became "always" when the style gap is wide enough.

I added "faithful to reference character design" to every subsequent prompt. Nothing else changed.

The results were immediate. The next variation, vintage tattoo illustration on black, maintained every identifying feature: curved horns, pointed ears, iridescent rainbow-shifting skin, mischievous grin. The style was completely different. The character was completely the same.

This phrase works because it gives NB2 a clear priority hierarchy: maintain the reference character first, then apply the requested style to it. Without it, style and reference compete as equals. And style tends to win.

With the anchoring language working, I ran Prism through five product styles. Every one maintained character identity while producing genuinely usable merch.

Vintage Tattoo Illustration (8.94 avg). Ornate decorative borders with roses, crystals, and filigree. Scroll banner text. Rich colors on solid black. NB2 even translated the "Adobe Firefly Ambassador" shirt text into the tattoo art style rather than just preserving it. The iridescent skin became faceted like stained glass. Four out of four print-ready on dark garments.

Streetwear Graffiti (8.83 avg). "PRISM" in massive dripping paint letters with the character emerging from behind. NB2 didn't just render the graphic on a dark background. It rendered it on actual hoodies with visible fabric texture, stitching, and in one image, a fake brand neck tag reading "streetwear tags." The halftone dot textures give authentic screen-print feel. One image limited itself to exactly three colors as prompted. Production-ready for cheap screen printing.

Chibi Stickers (8.84 avg). The biggest style leap. A mischievous horned creature becomes an adorable kawaii character with sparkle eyes, pink blush cheeks, and a gentle smile. Yet the horns, ears, iridescent skin, black tee, and red pants are all present. NB2 even generated multi-sticker sheets: a two-pack and a three-piece set with a separate banner sticker. Die-cut white outlines appeared automatically. The shirt text "Adobe Firefly Ambassador" is legible at chibi scale.

Collectible Figurine (9.56 avg, WEEK HIGH). This is the revelation. Every image scored 9.40 or above. NB2 rendered Prism as a premium collectible with painted metallic finishes, proper product photography lighting, and engraved "PRISM" pedestals. One image has a custom hexagonal base with iridescent rainbow accents matching the character's skin. These look like Hot Toys product photos. If you told me these were renders from a Kickstarter campaign, I'd believe you.

Repeat Pattern (8.35 avg). Prism's face and horn motifs tessellated across cream backgrounds with scattered "PRISM" text. One variation included profile views alongside front faces with crystal and bee filler motifs. The patterns aren't seamlessly tileable without post-processing, but the concept execution is strong enough to start from.

This deserves its own section because it contradicts something I've documented in my previous Firefly research.

Adobe Firefly Image 5 has a known limitation I call the "figurine problem." When you miniaturize living creatures, they become toy-like rather than photographic. The quality drops. They look like cheap plastic.

NB2 doesn't have this problem. At all.

The figurine variation scored 9.56 average across four images. The lowest individual score was 9.40. Higher than any other variation's best score except Tuesday's fantasy book cover. The iridescent skin has a physical, painted-surface quality. The bases are designed with material finishes. The lighting is professional product photography.

If you're a creator wanting to visualize merch concepts, figurines, vinyl toys, collectible statues, NB2 through Firefly is the tool. This isn't "close enough to get the idea." This is "put it on the Kickstarter page."

Something I didn't expect: NB2 renders products, not just graphics.

When I prompted "hoodie print," it didn't just put a graphic on a dark background. It rendered an actual hoodie. Fabric texture, neckline, stitching, kangaroo pocket. One image had a fake brand tag.

When I prompted "sticker sheet ready," it generated multi-sticker layouts with variant poses. Not a single sticker, but a product offering.

When I prompted "collectible figurine," it created pedestals with engraved text and custom material finishes.

Yesterday's self-help book covers did the same thing. NB2 sometimes produced physical hardcover mockups with spine shadows instead of flat art. The pattern is clear: NB2 understands product context. It knows that a "hoodie print" exists on a hoodie, a "figurine" stands on a base, and a "book cover" wraps around a book. This product awareness keeps escalating. Wait until you see what it does with ad creatives tomorrow.

After testing six styles (plus Monday's thumbnails and Tuesday's book covers), a clear hierarchy emerges for how well NB2 maintains character identity:

Strongest preservation. Photorealistic (figurine, thumbnails, book covers). The closer to the reference image's original rendering style, the more faithful the output.

Strong preservation. Illustration styles (tattoo, hand-drawn). Enough detail to maintain features, just in different rendering.

Moderate preservation. Stylized graphic (streetwear, sticker chibi). Features survive but are adapted to the style's conventions.

Weakest preservation. Abstract graphic (bold geometric). Without anchoring language, identity is lost entirely.

The rule: the more abstract the target style, the more critical the anchoring phrase becomes.

The anchoring phrase (add to any merch prompt):

For chibi and simplified styles:

Figurine (highest quality):

Tattoo illustration (best for dark merch):

Streetwear (text-forward):

Being honest about what's not production-ready out of the box:

The figurine images need nothing. Use them as-is.

Tattoo illustrations on black backgrounds are print-ready for dark garments. No editing needed.

Streetwear designs may need minor text-character overlap cleanup in one out of four images.

Chibi stickers need the standard die-cut file preparation (converting to vector paths for cutting machine).

Patterns are NOT tileable without edge alignment work. They're concepts, not production files.

The pupil issue from Monday persists across all styles. Fixable in Boards.

Tomorrow: social media ad creatives. Can Prism sell a product, promote an event, and close a sale?

Testing methodology: Nano Banana 2 (@NanoBanana), a partner model inside Adobe Firefly (@AdobeFirefly). All images scored using a weighted 5-dimension rubric. Minimum 4 generations per variation before drawing conclusions.