/quickdesign

When to use

• Avatar holding / wearing a product where you want pixel-faithful product detail (multi-ref: avatar + product side + product top)
• Angle change, state change (mouth empty when source had mouth full, glasses off, etc.)
• First-frame composition before feeding into Seedance R2V — banana ~12cr, Seedance ~250-500cr; a wrong first-frame chained into R2V wastes 50× the cost
• Product on white background, lifestyle composite, brand-kit-styled image

Skip in favor of seedream-v4.5 if the user is on a tight budget AND the edit is simple identity-pin (no fine-print / brand-text fidelity needed). Banana handles brand text better but costs slightly more.

Hard facts (live)

quickdesign cost nano-banana-2 -r 2K --num 1
quickdesign image models | jq '.data[] | select(.slug=="nano-banana-2")'

• Resolutions: 0.5K (1.5×), 1K (2×), 2K (3× — default), 4K (4×). Cost = 2 (base) × multiplier × num.
• Default to 2K for talking-head / selfie / lifestyle edits.
• Use 4K only when the video integrates a product with fine detail (engraving, label text, stitching) where pixel fidelity matters in the downstream Seedance render.
• Aspect ratios: 1:1, 9:16, 16:9, 4:5. Match the downstream video aspect.

Multi-reference grammar

Up to ~5 references is well-supported. Pass them as repeatable --reference-image in the order you want labeled:

quickdesign image generate \
  --model nano-banana-2 \
  --reference-image avatar.jpg \         # @Image1 — identity
  --reference-image product-side.jpg \   # @Image2 — product hero
  --reference-image product-top.jpg \    # @Image3 — product detail
  --aspect-ratio 9:16 --resolution 2K \
  -p "@Image1 holds a product matching @Image2 exactly. Replicate the stitching / lacing / sole color visible in @Image3. ..." \
  -o ./edit.png --wait

See ../references/multi-reference-pattern.md for the full pattern, common mistakes, and pre-generation checklist.

Compact prompt skeleton

Compose a single photo-realistic <aspect-ratio> frame using <N> reference images.
@Image1 is the subject — keep <face / hair / outfit identifying details> identical to the reference.
@Image2 (and @Image3 if supplied) is the product — render the held / worn item to match
@Image2 EXACTLY: <key textures / colors / stitching / brand text>. Do not invent details.
Scene: <setting>. Pose: <one-sentence action>. Lighting: <natural / studio / etc.>.
ANATOMY GUARD: exactly five fingers per hand, no extra digits, two normal feet,
hands and wrists clearly connected to forearms.

Gotchas / failure modes

0. Compose-mode regenerates the avatar instead of editing it. When the prompt opens with "Compose a vertical 9:16 UGC selfie frame using two reference images..." or similar, banana renders a fresh AI-look image inspired by @Image1 rather than modifying the source's pixels. The avatar's identity is approximated but the source's lighting, grain, and lo-fi authenticity are gone — output feels synthetic. For UGC, default to edit-style prompts (Edit @Image1: add ..., Take @Image1 as-is and only change ...). Don't re-list scene tokens already shown in the reference, and strip quality-upgrade phrases ("photo-realistic", "studio quality", "8K") — they trigger regen. See ../references/avatar-edit-not-regenerate.md for the verb library + setting-lock principle.

1. Hand / wrist hallucinations when subject holds a partially-occluded prop. Failure modes: severed wrists, six-finger grips, floating hands, props phasing through fingers. Mitigation:

   • Switch to a TWO-HAND pose when single-hand grip keeps failing — banana handles two-hand grips far more reliably
   • Add explicit contact verbs ("palm wrapped around the heel, fingers gripping the toe")
   • Add an anatomy guard clause to the prompt: "exactly five fingers per hand, no extra digits, anatomically correct hands and wrists"
   • Silent regen ~12cr is cheap; the user's attention is not. Use 1-2 silent regens before paging the user. See ../references/confirmation-rules.md.

2. Fine print / engraving / label text drops or smears. When the product has visible text (engraved "925" on jewelry, brand wordmarks, care labels, model numbers, billboard text in scene), banana routinely:

   • Drops the engraving entirely
   • Smears printed labels into illegible squiggles ("garbled text")
   • Replaces logos with similar-but-wrong glyphs

   Mitigation: name the specific text in the prompt ("preserve the engraved 'OTTA 925' wordmark on the silver clasp exactly as in @Image2"). Don't hand-wave with "preserve branding". The validator service checks for this; if the candidate drops fine print → retry with a sharper prompt.

3. Identity drift on full-body shots. Banana is strongest at tight selfie crops; full-body wide shots can drift on face details. Prefer half-body or selfie compositions when identity fidelity is critical.

4. Verbose verbatim re-description fights the reference. Same principle as Seedance: name the action / state / setting CHANGE, not the unchanged details. "Same person, same outfit, same setting — now without the crystal in mouth" works better than re-painting everything.

5. @Image1 syntax is shared with Seedance R2V — but banana also accepts plain "the first image / second image" wording. Stick with @Image1 for consistency across the pipeline.

Cross-references

• Edit-style vs compose-style verbs (avatar authenticity) → ../references/avatar-edit-not-regenerate.md
• Multi-product reference pattern → ../references/multi-reference-pattern.md
• Anatomy self-check before paging user → ../references/confirmation-rules.md
• Resolution decision (2K vs 4K) → ../pipelines/ugc-video.md (resolution rule)
• Pre-stage for Seedance R2V → ./seedance-2.0-r2v.md, ../pipelines/ugc-video.md