Flower photography has an unusual fidelity problem: natural variation is expected, but the product still has a recipe. A plausible bouquet can change the species, cultivar, stem count, color balance, bloom stage, foliage, size, mechanics, wrap, or substitutions that define the offer.
This test sent one fictional blush hand-tied bouquet brief through Seedream 4.5, Nano Banana 2, GPT Image 2, and FLUX.2 Pro. The four original outputs below show useful differences in visible petals, droplet cues, foliage, wrap, crop, and composition. No real bouquet, recipe, stem count, variety list, scale, color target, substitution policy, delivery state, packaging, or product data was supplied.
Evidence boundary: this is one text prompt and one displayed output per model. The observations describe visible fictional bouquets—not botanical identification, recipe accuracy, freshness, physical translucency, scent, vase life, model reliability, price/value, studio replacement, or preservation of a real florist SKU.
Quick answer
- Most pronounced displayed translucent-edge and droplet cues: Seedream 4.5 in this run.
- Fullest displayed upright wrapped composition: Nano Banana 2 in this run.
- Brightest displayed orderly arrangement: GPT Image 2 in this run.
- Softest displayed laid-down composition: FLUX.2 Pro in this run.
- What all four establish: each route made one plausible fictional blush-bouquet concept.
- What none establish: the correct recipe, species/cultivars, counts, scale, color, bloom stage, foliage, mechanics, wrap, substitutions, freshness, delivery, or repeated reliability.
One output per route does not establish a universal winner. It also cannot establish that flowers are an “easy” or reliable category: the test omitted the recipe and fulfillment facts needed to detect the most important failures.
First define what “the same bouquet” means
Florists sell at least two different kinds of visual promise:
| Offer type | Fidelity target | Required evidence |
|---|---|---|
| Exact or one-off arrangement | The photographed bouquet itself, subject only to disclosed delivery change. | Approved multi-angle photography, individual stem inventory, dimensions, mechanics, wrap, and delivery state. |
| Recipe-based bouquet | A repeatable recipe with disclosed natural variation and substitutions. | Required varieties/categories, minimum counts, size, focal/support/foliage ratios, palette, bloom stage, mechanics, wrap, packaging, and substitution policy. |
An AI image cannot choose that promise for the business. Define it first, then turn it into acceptance criteria.
The controlled concept brief
The original article did not publish the exact prompt. A bounded version matching the displayed subject is:
Fictional hand-tied blush-and-white bouquet with garden-rose-like focal blooms, ranunculus-like layered blooms, round eucalyptus-like foliage, natural kraft-paper wrap and jute twine, soft camera-left morning light, square florist product photograph. Keep individual blooms distinct and the wrap structurally plausible. No brand, card, words, vase, hand, person, price, stem-count claim, freshness claim, delivery claim, watermark, or extra bouquet.
This prompt can test composition and visible floral cues. It cannot preserve a SKU because the model invents every bloom, leaf, stem, count, variety, scale, opening stage, placement, mechanic, cut length, wrap, and hidden side.
Four first-hand outputs
Seedream 4.5 produced a close, dewy-looking bouquet with bright petal edges, layered blush-and-white blooms, foliage, kraft wrap, and twine. The image supports an art-direction observation: its displayed petal-edge and droplet treatment is the most pronounced of the four. It does not establish real dew, freshness, species, cultivar, count, scale, bloom stage, or recipe fidelity.
Nano Banana 2 made a full upright florist-style concept with multiple blush focal blooms, pale layered flowers, round foliage, kraft wrap, and twine. The arrangement reads clearly at article scale. No recipe or scale source was supplied, so its apparent completeness cannot make it the bouquet a customer ordered.
GPT Image 2 produced a bright, orderly arrangement with distinct pale blooms, foliage, and a clean studio treatment. It is easy to inspect, but the visible forms were not compared with a variety list, stem count, recipe diagram, color target, or approved bouquet.
FLUX.2 Pro placed the wrapped bouquet on its side in a wider, softer composition. That can be useful editorial art direction, while making scale and construction harder to judge. Full-resolution review must still check every required flower, count, proportion, leaf, stem, mechanic, edge, and wrap element.
What this comparison supports
| Question | What the four images show | What remains untested |
|---|---|---|
| Can the routes make a bouquet concept? | Yes, once each for this fictional brief. | Reliability across seeds, recipes, flower types, scales, crops, references, prompts, and current routes. |
| Do visible petal and composition treatments differ? | Yes; edge light, droplets, density, order, crop, foliage, and wrap vary. | Physical translucency, moisture, freshness, scent, vase life, species/cultivar, and real-light behavior. |
| Are the displayed flower forms plausible? | They read as familiar floral forms at article scale. | Botanical identity, variety, petal/leaf/stem structure, growth logic, defects, disease, and full-resolution coherence. |
| Is a real florist SKU preserved? | No evidence; no bouquet source or recipe was supplied. | Counts, ratios, size, palette, bloom stage, foliage, mechanics, wrap, card, packaging, substitutions, and delivery state. |
| Are fulfillment claims established? | No. | Availability, seasonality, freshness, origin, scent, vase life, toxicity, care, sustainability, delivery condition, and every marketed claim. |
The old article treated the absence of an obvious waxy or fused-bloom failure as proof that the category was reliable. The defensible conclusion is narrower: four routes made attractive fictional bouquets once, and their visible petal and composition treatments differ.
Florist product fidelity checklist
Build the rejection sheet from approved bouquet photography, the recipe, scale references, mechanics, packaging, substitution policy, and current product data:
- Offer definition: exact arrangement or recipe-based SKU; disclosed natural variation, acceptable substitutions, seasonal rules, and customer expectation.
- Recipe: required species/cultivars or approved categories, minimum and target stem counts, focal/support/filler/foliage ratios, color proportions, and substitution hierarchy.
- Flower identity: bloom and bud form, petal count and shape, center, sepal, stem, thorn, leaf, branching, scale, and distinguishing variety cues.
- Condition and stage: bud/open ratios, opening stage, hydration cues, bruising, browning, wilting, disease, damage, guard petals, intentionally dried elements, and delivery state.
- Color: approved palette and tolerance under the intended illuminant, gradients, centers, foliage tones, wrap, ribbon, card, and color substitutions.
- Scale and silhouette: finished height, width, depth, circumference, face shape, profile, negative space, stem length, grip, and relationship to hands, vase, box, or ruler sources.
- Construction: hand-tied spiral or other mechanic, stem crossings, binding point, tape, wire, foam, cage, pins, picks, support, hydration, cut ends, and intentionally visible hardware.
- Wrap and finishing: paper count, type, color, fold, overlap, cuff, cellophane, tissue, ribbon/twine, knot, card, label, sticker, sleeve, water pack, and presentation orientation.
- Packaging and delivery: vase, box, carrier, insert, care card, flower food, enclosure card, add-ons, protective materials, transport state, and sold configuration.
- Product facts: availability, seasonality, origin, freshness, scent, vase life, toxicity, allergens, care, sustainability, delivery timing/condition, size range, and substitution promises remain in approved data.
If the photo looks fresh but the required recipe, counts, proportions, size, palette, mechanics, wrap, substitutions, packaging, or fulfillment promise differs, it is the wrong florist asset.
A recipe-first flower workflow
1. Capture authoritative sources
Photograph the approved bouquet from the front, back, both sides, top, three-quarter, binding point, cut ends, wrap, label/card, packaging, and scale. Add macros of required flowers and foliage, the recipe and counts, finished dimensions, palette, bloom-stage target, substitution policy, delivery state, and current product data.
2. Build the scene separately
Generate or photograph an empty shop, table, doorstep, event, seasonal, or editorial plate at the final crop. Composite approved bouquet photography, then build contact, shadow, depth, rim light, and atmospheric elements deliberately. This keeps the recipe deterministic.
3. Test a constrained reference edit
masonry image "Change only the environment to a simple warm florist table with soft camera-left morning light. Keep the supplied bouquet unchanged: finished size and silhouette, every required flower and foliage type, stem counts and proportions, bloom stages, color balance, placement, binding mechanics, stem length, kraft-paper folds, twine, card, label, and visible packaging. Add one natural contact shadow. Add no vase, hand, person, text, claim, replacement flower, or extra bouquet." \ --model gemini-3.1-flash-image-preview \ --ref ./approved-bouquet-front-three-quarter.png \ --aspect 1:1 \ --output bouquet-scene-candidate.png
The instruction is not a recipe lock. Compare the candidate against the approved multi-angle bouquet, counts, dimensions, palette, wrap, packaging, and substitution policy; composite the approved bouquet when exactness matters.
For a fictional concept:
masonry image "Fictional hand-tied blush-and-white bouquet with garden-rose-like focal blooms, layered pale blooms, round eucalyptus-like foliage, kraft-paper wrap, jute twine, soft morning light, no brand, card, words, vase, hand, person, price, freshness claim, or extra bouquet" \ --model seedream-4-5 \ --aspect 1:1 \ --output bouquet-concept.png
Bouquet acceptance sheet
| Area | Pass condition |
|---|---|
| Source match | Candidate is compared with approved front, back, sides, top, three-quarter, binding, stems, wrap, label/card, packaging, scale, recipe, counts, dimensions, palette, substitution policy, and product data. |
| Recipe and identity | Required species/cultivars or categories, counts, focal/support/filler/foliage ratios, forms, scale, colors, substitutions, and disallowed replacements match. |
| Condition and construction | Bloom stage, hydration-facing cues, defects, stem length, silhouette, dimensions, binding, mechanics, hardware, cut ends, wrap, ribbon/twine, card, and orientation match. |
| Packaging and delivery | Vase/box, carrier, insert, care card, flower food, enclosure, add-ons, protective materials, sold configuration, crop, dimensions, responsive variants, color, and format are approved. |
| Claims and rights | No freshness, scent, vase life, availability, seasonality, origin, toxicity, care, sustainability, delivery, affiliation, or design-clearance claim is inferred from appearance. |
Bottom line
These four outputs are useful as a petal-and-composition art-direction comparison. They are not evidence that flowers are an easy category, the species and recipe are accurate, a fictional bouquet is ready to list, or one route is always the best value.
Define the product promise first. Use fictional concepts to choose the mood; use the approved recipe, counts, size, palette, mechanics, wrap, substitutions, packaging, and product data to decide whether an asset can ship. Compare the broader AI product photography model review, build a controlled scene in Masonry’s product photography tool, or automate candidates with the Masonry CLI.
