Most weak AI images fail for one reason: the prompt names a subject and leaves everything else to chance. Style, light, and framing decide whether an image looks like stock filler or a deliberate shot. This guide gives you a repeatable prompt formula, a vocabulary for lighting and composition, and model-specific adjustments for Midjourney, GPT Image, FLUX, and Imagen.
Quick Takeaways
- Use a six-part formula: Subject + Style + Lighting + Composition + Color + Quality. This structure works across all major platforms.
- Lighting changes the result more than any other single term. Name the source, direction, and quality of the light.
- Prompting is model-specific now. Midjourney likes compact, stylized phrases. GPT Image and Imagen follow full sentences. FLUX and Stable Diffusion respond to weighted, structured detail.
- Iterate one variable at a time. Change lighting or framing, keep the rest fixed, and compare.
| Element | What it controls | Weak example | Strong example |
|---|---|---|---|
| Subject | Content and action | “a woman” | “a ceramicist in her 60s trimming a bowl” |
| Style | Medium and rendering | “nice” | “35mm film photograph, Kodak Portra 400” |
| Lighting | Mood and depth | “bright” | “soft window light from camera left, long shadows” |
| Composition | Framing and viewpoint | “close” | “low-angle, rule of thirds, shallow depth of field” |
| Color | Palette and contrast | “colorful” | “muted teal and terracotta, low saturation” |
| Quality | Finish and detail | “HD” | “fine skin texture, sharp focus on eyes” |
The Prompt Formula Behind Consistent Results
Every strong image prompt reads like a short creative brief. Several guides converge on the same ordered structure: subject, style, lighting, composition, color, and quality. Others add mood or action as extra slots. The order matters because models weight early tokens more heavily. Lead with what must appear, then refine how it looks.
Template:
[Subject + action], [setting], [style/medium], [lighting], [composition/camera], [color palette], [mood], [quality details]
Worked example:
A ceramicist trimming a stoneware bowl on a wheel, cluttered sunlit studio,
35mm film photograph, soft window light from camera left with gentle rim light,
medium close-up from slightly above, shallow depth of field, rule of thirds,
muted terracotta and sage palette, calm focused mood, visible clay texture
How long should the prompt be?
For modern models, 50 to 150 words is a workable range. Beyond that, instructions start to compete. Leaving unimportant details open-ended is a legitimate strategy: name what matters, and let the model improvise the rest.
Style Prompts: Naming the Look
A missing style directive pushes the model toward a generic average look. Pick one primary style and support it with one or two modifiers. Stacking five clashing styles produces mush.
Style families that behave predictably
| Style family | Example phrases | Best for |
|---|---|---|
| Photographic | “35mm film”, “medium format”, “85mm portrait lens” | Portraits, product, editorial |
| Illustration | “gouache illustration”, “linocut print”, “flat vector” | Book art, branding |
| Cinematic | “anamorphic lens flare”, “film still”, “color graded teal-orange” | Key art, thumbnails |
| 3D and render | “isometric 3D render”, “clay render”, “octane-style lighting” | Icons, product concepts |
| Painterly | “oil on canvas, impasto”, “watercolor wash” | Concept art, posters |
Reference a medium or a film stock rather than a living artist’s name. Medium terms are portable across models. Artist names invite inconsistency and, for commercial work, legal ambiguity.
Lighting Prompts: The Highest-Leverage Term
Lighting is one of the most underused prompt elements, and adding it often makes the biggest visual difference. Describe three things: source, direction, and quality.
Lighting vocabulary that models understand
| Lighting term | Visual effect | Typical use |
|---|---|---|
| Golden hour | Warm, low-angle, long shadows | Portraits, landscapes |
| Blue hour | Cool, even, moody | Cityscapes, atmosphere |
| Rembrandt lighting | Triangle of light on the shadow-side cheek | Dramatic portraits |
| Softbox / diffused | Soft shadows, flattering | Product, beauty |
| Rim / backlight | Glowing edge separation | Silhouettes, hero shots |
| Hard noon sun | Harsh contrast, deep shadows | Documentary, street |
| Neon / practical lights | Colored spill from in-scene sources | Cyberpunk, night scenes |
| Volumetric light | Visible beams through haze | Fantasy, interiors |
Lighting rules that save rerolls
- Tie the light to a source. “Window light from camera left” beats “nice lighting.”
- Match light to mood. Soft and warm reads safe. Hard and low-key reads tense.
- Avoid contradictions. “Harsh midday sun” plus “soft diffused glow” splits the difference badly.
- Test with a fixed seed or fixed subject. Swap only the lighting line and compare outputs.
Composition Prompts: Controlling the Frame
Composition tells the model how to frame the scene. Specify shot size, angle, and layout rule.
| Dimension | Options | Effect |
|---|---|---|
| Shot size | extreme close-up, close-up, medium, wide, aerial | Intimacy vs. context |
| Angle | eye level, low angle, high angle, top-down, Dutch tilt | Power, vulnerability, tension |
| Lens feel | 24mm wide, 50mm natural, 85mm portrait, macro | Perspective distortion and compression |
| Depth of field | f/1.8 shallow, f/8 deep | Subject isolation vs. full detail |
| Layout rules | rule of thirds, centered symmetry, leading lines, negative space | Visual balance |
Use negative space on purpose when you need room for headlines. “Subject in the lower right third, large empty sky above” gives designers a usable canvas.
Model-Specific Prompting
The same prompt behaves differently across engines, so adjust the writing style. The field has no single default winner now. The best tool depends on the job.
| Model family | Prompt style | Strength | Trade-off |
|---|---|---|---|
| Midjourney (V8.x) | Short, evocative phrases plus parameters | Best artistic quality and creative exploration | Weaker text rendering; subscription from about $10/mo |
| GPT Image (via ChatGPT or API) | Full natural-language sentences | Precise instruction following, accurate text, conversational edits | Less stylized by default |
| FLUX.2 | Detailed descriptive prose | Photoreal products and skin texture | Needs tuning for stylization |
| Imagen 4 | Clear sentences | Legible in-frame text and product photography | Enterprise access via Vertex AI |
| Stable Diffusion 3.5 | Weighted tags or prose, negative prompts | Open weights, local control, LoRA support | More setup and tuning |
Version numbers move quickly. Midjourney made V8.2 its default model on July 24, 2026, and OpenAI replaced Images 2.0 with GPT Image 2.5 on September 8, 2026. Check your tool’s current default before you copy any parameter set.
Midjourney parameters
ceramicist trimming a stoneware bowl, sunlit studio, 35mm film photograph,
soft window light, medium close-up, muted terracotta palette
--ar 3:2 --stylize 250 --chaos 10 --no text, watermark
--ar 3:2sets the aspect ratio.--stylize(0–1000) controls how strongly Midjourney’s house aesthetic overrides your prompt. Use 50–150 for literal results and 250+ for artistic ones.--chaos(0–100) increases variation across the four-image grid.--noexcludes elements.
GPT Image via the API
import base64
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY from the environment
prompt = (
"A ceramicist trimming a stoneware bowl on a wheel in a cluttered, sunlit studio. "
"35mm film photograph with soft window light from camera left. "
"Medium close-up from slightly above, shallow depth of field. "
"Muted terracotta and sage palette. Visible clay texture."
)
result = client.images.generate(
model="gpt-image-2", # swap in the newest GPT Image model ID you have access to
prompt=prompt,
size="1536x1024", # landscape
quality="high",
)
with open("ceramicist.png", "wb") as f:
f.write(base64.b64decode(result.data[0].b64_json))
OpenAI removed DALL-E 3 from the API on May 12, 2026, and new API projects should target the current GPT Image model directly. Annotate your own prompts in version control so you can reproduce a look later.
Open-weight model settings (Stable Diffusion, FLUX)
| Setting | Typical range | Effect |
|---|---|---|
| Steps | 25–35 | More steps add detail until returns flatten |
| CFG / guidance | ~5–8 for SDXL-class; ~3–4 for FLUX dev-style | Higher means stricter prompt adherence, with a risk of oversaturation |
| Resolution | ~1 megapixel native | Generate native, then upscale |
| Seed | Fixed during A/B tests | Isolates the variable you change |
| Negative prompt | “blurry, extra fingers, watermark” | Suppresses artifacts (SD-family) |
Ranges vary by checkpoint, so treat these as starting points.
Practical Workflows and Real-World Use Cases
1. Product photography template
[Product] on [surface], [style: commercial studio photograph],
softbox key light from 45° left with white fill card on right,
soft gradient shadow, centered composition with generous negative space,
neutral [color] backdrop, crisp focus across the product, 85mm lens feel
Use case: An e-commerce team approves one product shot, then holds the template constant while swapping the product, backdrop color, and language. This is how one approved image becomes dozens of consistent campaign assets.
2. Editorial portrait template
[Person description], [expression], [clothing],
medium-format film portrait, Rembrandt lighting from camera right,
85mm lens, f/2 shallow depth of field, eye-level framing,
muted earth tones, quiet confident mood, fine skin texture
3. Cinematic key art template
[Hero + action] in [location], cinematic film still, anamorphic lens,
low-angle wide shot, backlit with volumetric haze and warm rim light,
teal and orange grade, rule of thirds, dramatic mood
4. Iterative editing loop (conversational models)
Once you have a good base, edit rather than rewrite. Follow-up instructions can target specific attributes such as lighting, color balance, or object placement while preserving the overall structure. Use short, scoped commands:
- “Keep the subject and framing. Change the lighting to backlight at sunset.”
- “Same composition. Shift the palette to cool blues.”
- “Move the subject to the left third and leave empty space on the right.”
5. Reference-driven consistency
When words fall short, supply a reference image and state what must stay fixed and what may change. Tools like Kling IMAGE 3.0 are built around this pairing of reference images with natural-language instructions. Apply the same pattern in any tool that accepts image inputs.
Common Mistakes and Fixes
| Mistake | Symptom | Fix |
|---|---|---|
| Stacking 5+ styles | Muddy, generic output | Pick one primary style and one modifier |
| No light source named | Flat, shadowless image | Add source, direction, and quality |
| Contradictory terms | Inconsistent results | Remove one side of the conflict |
| Keyword spam (“8k, ultra, masterpiece”) | Little gain on modern models | Replace with concrete detail such as “visible fabric weave” |
| Changing many variables per rerun | You can’t tell what worked | Edit one element per iteration |
| Ignoring aspect ratio | Awkward crops | Set the ratio before generating |
A Repeatable Prompting Process
- Write the subject sentence first. Make it concrete and visual.
- Choose one style anchor. Use a medium, film stock, or rendering term.
- Add a lighting line. Give the source, direction, and quality.
- Add a framing line. Give shot size, angle, and layout rule.
- Set the palette and mood. Two or three color names are enough.
- Generate a batch of four. Keep what works.
- Change one element and rerun. Save the winning prompt as a template.
FAQ
What is the best formula for an AI image prompt?
Use Subject + Style + Lighting + Composition + Color + Quality. Put the must-have content first. Then describe how it should look. This structure works across Midjourney, GPT Image, FLUX, and Stable Diffusion.
How long should an AI image prompt be?
Aim for 50 to 150 words on current models. Include every visual factor that matters and leave unimportant details open. Past about 150 words, instructions begin to conflict.
Which lighting terms work best in AI image prompts?
Terms tied to a physical source work best: golden hour, Rembrandt lighting, softbox, rim light, and window light from camera left. State the direction and quality of the light, not just “good lighting.”
Do I need different prompts for Midjourney and ChatGPT?
Yes. Midjourney responds well to compact phrases plus parameters like --ar and --stylize. GPT Image and Imagen follow full sentences and precise edit instructions. Prompting is model-specific now, so tune your phrasing to the engine.
