The Complete Prompt
See the full picture first, then break it down:
Transform this image into a refined, high-end low-poly mosaic style.
Strictly preserve the original structure and recognizable facial
details, especially the eyes, mouth, and characteristic hair
contours. Use extremely small, high-density polygons to maintain
photographic clarity and identity while creating a crystalline,
faceted, and multi-dimensional look. Retain the original color
palette for a harmonious, organic, and natural aesthetic. Ensure
clean sharp edges between every polygon. No blurring or smooth
transitions. High-resolution digital art aesthetic,
crystalline masterpiece.
This prompt contains 7 functional word groups, each handling a different job. Remove any one of them and the output quality drops noticeably. Below is a function-by-function breakdown of what each group actually does.
Word-by-Word Breakdown — Why Each Term Is Here
Group 1: Task Frame — "Transform this image into a refined, high-end low-poly mosaic style"
Function: Defines the overall task nature and sets AI's "working context."
refined and high-end are not decorative adjectives — they redirect which training data AI pulls from. Without these two words, low-poly defaults to crude early digital game art (large triangles, flat-fill colors, no lighting gradients). With them, AI shifts attention to fine digital art and contemporary design reference material.
mosaic further specifies the expected density — the word "mosaic" in AI semantics implies dense small-unit composition, distinguishing it from "crude low-poly."
| Version | Phrase | Expected Result Difference |
|---|---|---|
| Basic | low-poly style |
Coarse, large polygons, more filter-like |
| Improved | refined low-poly style |
More detail, smoother transitions |
| Full version | refined, high-end low-poly mosaic style |
Highest density, most refined texture quality |
Substitution experiment: Replacing high-end with experimental produces avant-garde art aesthetics; replacing it with commercial produces cleaner versions suitable for brand presentation.
Group 2: Structure Protection — "Strictly preserve the original structure and recognizable facial details, especially the eyes, mouth, and characteristic hair contours"
Function: The most critical group for portrait transformation — preventing AI from "over-creating" during geometricization.
Strictly upgrades "preserve" from a suggestion to a constraint. Without strictly, AI will "try" to preserve structure; with it, structure preservation becomes the primary execution goal.
especially the eyes, mouth, and characteristic hair contours specifies priority protection zones. Eyes, lips, and hairline are the key recognition anchor points for human faces. Listing these three separately causes AI to use denser polygons in these zones — even if overall density is moderate — ensuring minimum recognition features are never lost.
Substitutions for other subjects:
- Animal portraits: Replace
eyes, mouth, and characteristic hair contourswitheyes, snout, and distinctive coat pattern - Architectural subjects: Replace entire group with
Strictly preserve the building's structural lines, window arrangements, and architectural details - Landscapes: Replace with
Strictly preserve the horizon line, major landmass boundaries, and primary color zones
Group 3: Density Control — "Use extremely small, high-density polygons to maintain photographic clarity and identity"
Function: Directly controls polygon "granularity" — the core parameter determining whether the output looks like "art" or "a filter."
extremely small, high-density is a double reinforcement — small controls individual polygon area, high-density controls how many polygons per unit area. Both terms together produce the "fine crystal" texture; using just one, AI doesn't execute with enough thoroughness.
maintain photographic clarity and identity sets a boundary for the density — AI isn't just supposed to be "dense," it has to be "dense enough to recognize who the subject is." This phrase prevents AI from moving toward "geometric for geometry's sake" extreme abstraction, ensuring the transformed result still has commercial usability.
Density range:
| Density Keywords | Effect | Best For |
|---|---|---|
small, medium-density |
Balanced — geometric feel with recognizability | Landscapes, architecture |
small, high-density |
High density, rich detail, high recognizability | Portraits, complex subjects |
extremely small, high-density |
Extreme density, closest to photographic quality | Close-up portrait shots |
large, low-density |
Low density, strong geometric feel, abstract | Decorative posters, no recognizability needed |
Group 4: Style Anchor — "crystalline, faceted, and multi-dimensional look"
Function: After the task type is set, style anchor words precisely direct "which aesthetic direction this low-poly should take."
crystalline triggers AI to generate subtle light gradients inside each polygon — simulating light refraction through gemstone facets — rather than flat-fill colors. This single word is the most important difference between refined and crude low-poly results.
faceted emphasizes that clear ridgelines between polygons must be visible, preventing AI from softening edges.
multi-dimensional tells AI to maintain 3D depth inside flat geometric shapes — expressing volume through per-polygon luminance differences rather than treating all polygons as equal-brightness flat tiles.
Substitution experiment: Replacing this group with stained glass mosaic look produces a colored glass window effect (thicker edge lines, brighter colors); replacing with ancient mosaic tile produces a Roman tessellation aesthetic (more saturated, with wear texture).
Group 5: Color Strategy — "Retain the original color palette for a harmonious, organic, and natural aesthetic"
Function: Controls whether AI changes the source image's color information.
Without this group, AI sometimes "creatively" alters the color scheme — converting warm-toned originals to cool tones, or making polygon colors more saturated. This freedom can be useful for decorative work but typically destroys skin tone naturalness in portrait transformation.
harmonious, organic, and natural describes the ideal color relationships: balanced (not jarring), organic (not mechanical), natural (not over-saturated). Together they constrain AI to color restraint, avoiding cyberpunk palettes and neon gradients that AI commonly over-applies.
When you want to change colors: Replace the entire group with transform the color palette to [specific color scheme]. Example: transform the color palette to cool midnight blue and silver tones produces a cold-toned crystalline night effect.
Group 6: Edge Constraint — "Ensure clean sharp edges between every polygon. No blurring or smooth transitions"
Function: This group specifically prevents AI's "softening tendency" — AI defaults to light edge softening when rendering geometric shapes (because this typically corresponds to "high quality images" in training data), but in low-poly style, soft edges are exactly what causes quality breakdown.
No blurring or smooth transitions is a negative constraint (telling AI "what not to do"). Negative constraints are highly efficient in AI prompts because they directly eliminate the most common error behaviors, rather than relying on positive descriptions to "guide" correct behavior.
A common misconception: thinking positive and negative constraints are redundant repetition. They actually serve different functions: clean sharp edges (positive) tells AI the target state; No blurring or smooth transitions (negative) eliminates paths toward the wrong state. Both together are far more effective than either alone.
Group 7: Quality Seal — "High-resolution digital art aesthetic, crystalline masterpiece"
Function: Quality terms at the end act as a "quality gate" — they don't change content composition but increase AI's "effort investment" in detail handling.
crystalline masterpiece is more effective than plain masterpiece because it binds the quality expectation with the style keyword (crystalline) — AI optimizes "crystalline quality" as a unified goal rather than two separate requirements.
Substitution experiment: Replacing crystalline masterpiece with award-winning digital illustration shifts AI toward Behance-style refined digital illustration; replacing with museum-quality print shifts toward larger-scale, more detailed processing.
Word Order Experiments — What Happens When You Swap Positions
Low-poly prompts are more sensitive to word order than most styles. Three common word order mistakes:
Mistake 1: Structure protection before task framing
Strictly preserve facial details... Transform into refined low-poly mosaic...
Result: AI treats "preservation" as the highest priority, significantly reducing geometric transformation intensity. Output looks like "a slightly geometricized photo" rather than a "crystallized artwork" — appropriate if you want subtle changes, but wrong if you want strong geometric impact.
Mistake 2: Quality terms in the middle
...high-density polygons. Crystalline masterpiece. Retain color palette...
Result: Quality terms interrupt the logical continuity between parameter groups. AI starts a new interpretation after crystalline masterpiece — subsequent color instructions execute inconsistently, and generation-to-generation variation increases significantly.
Mistake 3: Negative constraints first
No blurring or smooth transitions. Transform this image into...
Result: Negative constraints appearing first put the entire generation process into excessive "avoid-blurring" vigilance, which can produce unnaturally thick edge lines and even affect lighting gradients inside polygons.
Correct order logic: Task type (set frame) → Structure constraint (set limits) → Density control (set granularity) → Style anchor (set aesthetic direction) → Color strategy (set palette) → Edge constraint (set boundary quality) → Quality terms (set overall standard)
3 Substitution Experiments: Switching Subject Types
The same prompt framework with different subject replacements produces completely different results:
Variant A: Architecture Low-Poly Transformation
Replace Group 2 with:
Strictly preserve the building's structural lines, window
arrangements, and facade architectural details, especially
the edge silhouette and vertical alignment
Best for: tech company annual report illustrations, architectural concept presentations, cityscape art posters. Architecture's straight edges "naturally align" with polygon meshes, typically producing the cleanest geometric results.
Variant B: Animal Portrait Low-Poly Transformation
Replace Group 2 with:
Strictly preserve the animal's distinctive features, especially
the eyes, facial expression, and the characteristic texture
pattern of the coat or skin
Best for: brand mascot design, nature photography art, pet memorial pieces. Note: short fur or scaled animals (fish, lizards) work better than long-furred animals (cats, dogs) — flowing hair converts to polygons with a "comb-cut" effect that destroys the original soft impression.
Variant C: Product/Object Low-Poly Transformation
Replace Group 2 with:
Strictly preserve the product's form factor, brand logo
positioning, and distinctive shape silhouette
Best for: brand art presentations, product concept imagery, creative e-commerce visuals. Regular geometric product forms (bottles, packaging boxes) are highly compatible with polygon meshes; irregular organic forms (food, fabric) are less compatible.
For a deeper technical dive into how density parameters work, the low-poly mosaic technical deep guide covers everything from 6-level density comparisons to boundary protection fine-tuning.
Common Failures and Fixes
Failure 1: Face severely distorted, unrecognizable
Cause: Group 3 density not strong enough, or Group 2 structure protection missing.
Fix: ① Change small, high-density to extremely small, high-density; ② add with absolute fidelity to the primary facial proportions to the end of Group 2; ③ check source lighting — images with flat even lighting (indoor overhead light, overcast diffusion) lack light-dark boundaries, leaving AI without contour "cues" to follow.
Failure 2: Edges have a soft, blurry quality
Cause: Group 6 edge constraint missing, or AI's default edge softening is active.
Fix: Strengthen the negative constraint: Ensure absolutely clean, razor-sharp polygon boundaries with zero anti-aliasing or edge feathering. "Zero anti-aliasing" is a graphics technical term — AI understands it more precisely than "no blurring."
Failure 3: Colors are very different from the original
Cause: Group 5 color strategy too vague, AI applied creative color treatment.
Fix: Replace with more specific language: Preserve the exact color temperature, saturation levels, and tonal values of the original image — no color grading, no filters, no artistic color interpretation. "No artistic color interpretation" directly eliminates the most common type of AI color modification.
Failure 4: Background polygons look more interesting than the subject, stealing focus
Cause: Prompt doesn't specify relative weighting between subject and background, AI processes the entire frame uniformly.
Fix: Add Apply highest polygon density to the subject [SUBJECT] with progressively coarser polygons toward the background edges, emphasizing the subject as the visual focal point. This "density gradient" description creates natural subject emphasis with background naturally receding.
Use the 7-group framework from this article to generate 3 images — one each with a portrait, a building, and an animal — and see which subject type pairs best with this prompt. Start at nano-banana.cn.
FAQ
Why is portrait transformation harder to control than landscape low-poly?
A face's "minimum recognition threshold" is much higher than a landscape's. In a landscape, missing a few polygons rarely affects recognition. But facial eyes, nose, and mouth lose recognizability immediately if polygon density is insufficient. Solution: portrait transformation must include Group 2 structure protection (specifically naming eyes, mouth, and hair contours) and higher density than landscape work (extremely small, high-density rather than medium-density).
Why write "No blurring" separately — isn't "sharp edges" enough?
They serve different functions in AI. Sharp edges (positive constraint) describes the target state; AI knows where to go. No blurring (negative constraint) directly eliminates the error behavior — because even while "executing sharp edges," AI sometimes still applies slight anti-aliasing to boundaries (a default behavior). The negative constraint explicitly turns off that default. Both together are far more effective than either alone.
Can I make the background also low-poly but in a different style from the subject?
Yes. Add to the end of the prompt: Subject uses ultra fine, high-density polygons for maximum detail; background uses larger, more impressionistic polygon blocks creating depth gradient between foreground and background. This foreground/background density contrast is a standard technique in film promotional posters and brand visuals — fine subject, abstract background, naturally clear focal point.
Does this prompt work on group photos (multiple faces)?
Yes, but success rate inversely relates to the number of faces. Single close-up portrait is most stable. Two-person shots require adding with equal structural preservation applied to both subjects. Three or more faces usually can't maintain sufficient density and recognizability for all simultaneously — recommend cropping into individual portraits and processing separately. If you must handle a group shot, add apply identical crystalline treatment uniformly to all subjects within the frame and raise density to extremely small, ultra-high-density polygons.
Every generation looks different — how do I stabilize output?
AI generation has inherent randomness that can't be fully eliminated. Approaches to increase stability: ① add more specific constraint terms (more precise descriptions leave less variation space); ② when a particularly successful result appears, record all the parameters you used and reference them next time; ③ generate 3-5 candidates and select the best — this is more efficient than repeatedly trying to "recreate" a specific successful result through prompt tweaks.