The biggest trap in AI image generation isn’t a model that’s too weak — it’s prompts that are too scattered. Only after I took every prompt apart and rewrote it along the six SCULPT elements did my hit rate go from 30% to over 85%. That methodology is the entire content of this handbook.
1. Why Abandon “Free-Form” Prompts
Early on, everyone was the “inspiration type”: picture something in your head, write a description by feel, throw it at the AI and leave the rest to fate. The result — run the same prompt three times and you get three completely different images.
After going through 329 prompts from experienced hands, the pattern was unmistakable: every controllable prompt is a fill-in-the-blank, not an essay. They all start from the six slots of a fixed framework and fill in concrete content.
The framework is called SCULPT — Subject / Composition / Universe / Light / Print / Texture. Each letter maps to one dimension you have to think through clearly before generating.
2. The Six SCULPT Elements: What Goes in Each Letter
S — Subject
What are we drawing? A person, an object, a scene. The more specific the better — down to appearance, posture, and clothing.
✅ “A swordsman in white robes, cold brow, a shallow scar on his jaw, robe lifted by the wind”
❌ “A swordsman”
C — Composition
How is it arranged? Viewing angle, proportions, placement, negative space, leading lines.
✅ “Close-up of the face filling 70% of the frame, head turned 15 degrees away from the camera, blurred background”
❌ “Put it in the middle”
U — Universe
What’s the tone? Period setting, art style, overall atmosphere.
✅ “Eastern cinematic aesthetics, wuxia world, old-dream atmosphere, cool color grade”
❌ “Something pretty”
L — Light
Light direction, color temperature, contrast, depth of field, atmosphere.
✅ “Side backlight catching the hair, cool white color temperature, shallow depth of field f/2.8, cinematic lighting”
❌ “A bit brighter”
P — Print
What text goes on the image? Wrap it in quotes and specify position, color, and size. Portrait close-ups usually have no text; only posters do.
✅ “Large title centered at the top, ‘Snow Fills the Rivers and Lakes’, white regular script, dark ink-wash background”
❌ “Add a title”
T — Texture
Surface texture, render parameters, post-processing style.
✅ “Kodak Portra 400 film grain, keep skin pore detail, 85mm f/2.8 shallow depth of field”
❌ “High definition”
3. Portrait Specialization: Composition Is the Dividing Line
Ninety percent of the “plastic look” comes from two things: flat frontal lighting + a centered full face. Counter them by filling the C and L slots correctly.
| Point | What to fill in | Why |
|---|---|---|
| Face coverage | 60–75%, with just a sliver of shoulder | A close-up carries more atmosphere than a full-body shot and goes wrong far less often |
| Lens | 85mm prime + f/2.8 (f/1.4 for close-ups) | 85mm compression gives the features dimension; shallow depth of field separates the background |
| Lighting | Side backlight + hair highlights | Creates rim light and avoids the “plastic look” of flat frontal lighting |
| Background | Light gray / cream / cool white, or slightly blurred | Keeps attention locked on the subject; don’t let the setting steal the scene |
| Skin | Keep pore detail, no skin smoothing | This one line is the whole difference between “looks like film photography” and “looks like AI” |
Full-body shots are where AI fails hardest (fingers, weapons, proportions). Beginners should start with close-ups and work up to full body.
4. Side by Side: One Brief, Two Ways to Write It
Brief: an Eastern wuxia portrait close-up.
❌ Beginner version (30% hit rate):
“Draw me a handsome ancient-style wuxia guy, make him handsome, with atmosphere”
✅ SCULPT version (85% hit rate):
S: A facial close-up of a swordsman in white robes, cold brow, gaze directed off-camera, a shallow scar on his jaw
C: Face fills 70% of the frame, head at 15 degrees, blurred snowy night street behind
U: Eastern cinematic aesthetics, wuxia world, old dreams, cool color grade
L: Side backlight on the hair, cool white, shallow depth of field f/2.8
P: No text
T: Shot on Sony A7M4, 85mm f/2.8, Kodak Portra 400 film grain, skin detail preserved
Fill the six holes, get one predictable image. That is the whole point of SCULPT — turning “rolling the dice” into “issuing instructions”.
5. High / Low Modes: Don’t Pick Wrong
GPT Image 2 has three generation modes:
- High (high precision) ⭐ use this by default — follows the prompt strictly, renders text accurately, keeps composition under control. Good for portraits, products, and posters.
- Auto — not recommended — the model decides for itself; the least stable option, and it will quietly switch to a style you didn’t want.
- Low (creative) — free rein; full of surprises but uncontrollable. Use it only for illustration, concept art, and exploring ideas.
Use High for 90% of cases.
6. Five Advanced Tricks for More Stable Output
- Edit mode: separate variables from constants. In multi-round editing, state explicitly which parts stay unchanged and which part this round modifies, then end with “only change XXX, don’t alter any other element”.
- Style consistency: pin a template block. Across images in the same series, fix the style paragraph as the unchanging tail of the prompt and change only the subject.
- Text rendering: three keys. Wrap all text in quotes; specify position, color, and size; tell the model the core text in quotes and add decorative text in post.
- Change only one variable at a time. When the result isn’t right, change a single variable and retry. Change three at once and you’ll never know which one fixed it.
- Prompt length: 100–200 characters is optimal. Too short and there isn’t enough information; too long and the model loses focus.
7. Eastern Wuxia Style Dictionary
Style is not SCULPT’s job — SCULPT governs how to fill the prompt, the style dictionary governs what tone to fill in. The two are orthogonal: one is scaffolding, the other is an interior design plan.
wuxia cinematic / 东方电影美学 / 武侠江湖 / 旧梦 / 风雪 / 长夜 /
孤灯 / 荒原 / 剑光 / 刀客 / 山鬼 / 冷色调 / 电影级光影
8. SCULPT Cheat Sheet
S — Subject Get specific about appearance, posture, and clothing.
C — Composition Angle + proportions + subject’s percentage of the frame + placement.
U — Universe Realistic / illustration / 3D / cyberpunk / Eastern aesthetics / film.
L — Light Light direction + color temperature + depth of field + mood words.
P — Print Quotes + position + color + size.
T — Texture Render parameters + depth of field + post-processing style (film grain, etc.).
9. Action Checklist
- Pick a wuxia character and write one complete prompt using the six SCULPT elements.
- Focus on filling C (composition) and L (light) with concrete numbers (coverage, aperture, light direction).
- Generate 4 images in High mode and pick the one in the middle of the pack.
- Look at the hair, skin, and lighting, find the weakest slot, change only that one, and retry.
- Save the prompts that work into your own asset library, with the style paragraph pinned as the unchanging tail.
Final Notes
SCULPT isn’t a fancier prompting trick; at its core it is turning “what I feel” into “what I specify”. Every slot you fill moves the AI from “guessing what you want” to “doing what you said”.
Fill C (composition) and L (light) with concrete numbers first and you’ll see the difference immediately — the gap between “plastic” and “cinematic” is exactly those two slots.
