The image is in focus, the subject is right, but the lighting looks like a snapshot taken with on-camera flash: flat, harsh, no depth, no mood. Or worse, it looks like a webinar headshot under fluorescent overheads.
Fastest fix: add a complete light spec to your prompt instead of a lighting adjective. Replace studio lighting with single softbox at 45° camera right, low-angle, 3200K warm, deep shadow on the far side. That one change fixes the majority of flat-lighting prompts on the first regenerate.
This is almost always a prompt problem, not a model problem. Every current model can produce dramatic, dimensional lighting, but only if you tell it exactly what light setup you want. studio lighting averages across every studio setup in the training data and lands on something flat; a named setup with direction, quality, source, and temperature gives you cinematic, repeatable results.
This applies the same way across GPT Image 2 (ChatGPT’s image model since April 21, 2026), Midjourney v8 (2K native, aesthetics leader), Nano Banana Pro (Google’s Gemini-3-based model, photorealism leader), and FLUX 2 Pro. The vocabulary below is model-agnostic.
Diagnose first: which bucket are you in?
Read your prompt and match the lighting words against this table. Most flat images are bucket 1, 2, or 3.
| Symptom in the output | What’s in your prompt | Fix |
|---|---|---|
| Even, shadowless, mid-day overhead look | No lighting word at all | Add direction + quality + temperature (Step 1) |
| Generic, slightly soft, no character | Only studio lighting / professional lighting / beautiful lighting | Name a specific setup (Step 1) |
| Flat top-light, no modeling on the face | Has a lighting word but no direction | Add from camera left/right, 45°, backlit (Step 2) |
| Soft everywhere, zero contrast | Every light is soft / diffused | Make the key hard or directional (Step 3) |
| Wrong color cast (too blue, too orange) | No Kelvin value, or conflicting time-of-day words | Pin one Kelvin value (Step 4) |
| Mid, muddy “movie” look | cinematic with no reference | Name a DP or film, or drop it (Step 5) |
| Preview looks flatter than the file | Platform preview softening | Download the raw file and compare |
Common causes
Ordered by hit rate, highest first.
1. No lighting cue at all
If your prompt has zero lighting words, the model picks a generic safe default, usually broad, flat, mid-day overhead.
How to spot it: read the prompt. No light, lit, shadow, softbox, golden hour, time-of-day word, or temperature spec means you got the model default.
2. “Studio lighting” alone
studio lighting is so vague it averages across all studio setups in the training data: beauty dish, ring light, three-point, softbox, hard light. Result: generic flat output.
How to spot it: your only lighting word is studio lighting, professional lighting, beautiful lighting, or lit well. Be specific instead.
3. Direction not specified
soft lighting without direction is also flat. The model places light from above (most common in training data) and you get top-flat lighting with no modeling on the face.
How to spot it: a lighting word exists but no from left, from right, from above, from behind, front-lit.
4. Multiple lights all written as “soft”
"soft key light, soft fill light, soft rim light, soft ambient"Soft + soft + soft = flat. You need contrast: a hard key against a soft fill, or a directional rim against ambient.
How to spot it: all your light terms are soft or diffused. Make one harder or more directional.
5. “Cinematic” without specifying the genre
cinematic is broad: film noir is dark and hard, teen rom-com is bright and soft, Wes Anderson is symmetric and even. The model averages and gives you mid.
Note (verified against OpenAI’s GPT Image 2 prompting guide, June 2026): when you want authentic photographic realism, words like cinematic grading and dramatic color grading can actually push the model toward a fake, over-graded look. For real-photo results, prefer natural lighting, believable detail, no color grading. Save cinematic for when you genuinely want a stylized, graded frame.
How to spot it: you wrote cinematic but didn’t name a film, era, or DP style.
6. Tool-side denoising / preview washes it out
Some platform previews soften shadows for “friendly” thumbnails, hiding the contrast. The raw image may be better.
How to spot it: download the raw PNG/JPG and compare it to the preview thumbnail.
Shortest path to fix
Step 1: Pick a lighting setup, name it explicitly
Use a complete spec: direction + quality + source + temperature. Templates that work in any current model:
# Portrait, moody (Rembrandt)
"Rembrandt lighting, single hard key light from camera left at 45° and
slightly above eye level, no fill, 3200K tungsten warm, small light
triangle on the shadow-side cheek, deep shadow on the far side"
# Portrait, beauty
"Beauty dish lighting straight on from above, soft white reflector below,
even fill, 5500K daylight neutral, soft shadows"
# Outdoor, golden hour
"Backlit by low-angle 5pm sun from camera right behind subject,
warm rim light on edges, soft fill from sky on shadow side, 3000K warm"
# Indoor, window light
"Single large north-facing window light from camera left, soft diffused,
no fill, natural shadow falloff, 5600K cool daylight, low contrast"
# Cinematic / dramatic
"Single hard light from camera left, motivated by off-screen lamp,
deep shadows, no fill, 3200K, high contrast, film noir style"Step 2: Always specify direction
The single biggest lever. Pick one:
from camera left/from camera rightfrom above/top-downfrom below/up-litfrom behind/backlit45° and slightly above eye level(classic key-light position)front-on(least flattering, avoid for portraits)
Step 3: Specify quality (hard vs soft)
hard light -> small source, sharp shadow edges, dramatic
soft light -> large diffused source, gradient shadows, gentle
# Recipes
hard: "direct sun, no diffusion" / "single bare bulb"
soft: "large softbox 4ft wide" / "overcast sky" / "window light through curtain"Step 4: Specify temperature in Kelvin
This is what photographers do, and it works in prompts too:
2200K -> candle, fire warm
3200K -> tungsten warm interior
4000K -> mixed warm-neutral
5000K -> daylight neutral
5600K -> daylight cool
7000K -> overcast cool
10000K -> blue hour, twilightPick exactly one Kelvin value. Two conflicting time-of-day words (for example golden hour and blue hour in the same prompt) make the model split the difference and produce a muddy cast.
Step 5: For “cinematic” — name the reference
Instead of cinematic, write:
"cinematography in the style of Roger Deakins" (warm, motivated)
"cinematography in the style of Bradford Young" (low-key, deep shadows)
"shot like Wong Kar-wai" (neon, smoke, color blocks)
"shot like Wes Anderson" (symmetric, even, pastel)
"shot like David Fincher" (cool, hard light, dark)If you actually want a real-photo look rather than a movie still, drop cinematic entirely and use natural lighting, no color grading (see cause 5).
Step 6: Build a lighting prompt library
Save 8-10 lighting “blocks” you can paste into any prompt. Examples to keep:
- “Beauty headshot” — single beauty dish from above, reflector below
- “Rembrandt portrait” — hard key from left at 45°, deep shadow, triangle on shadow cheek
- “Window light” — north window from left, soft fill from wall
- “Golden hour backlit” — 5pm sun behind, warm rim
- “Studio softbox” — single 4ft softbox at 45° camera right
- “Film noir” — single hard light from below at 30°, deep shadow
How to confirm it’s fixed
Regenerate with the new spec and check three things:
- Modeling on the face. There should be a visible light side and shadow side, not even brightness across the whole face. For Rembrandt, look for the small light triangle on the shadow-side cheek.
- Direction reads. You should be able to point at where the light is coming from. If you can’t, your direction term didn’t land — make it more explicit (
hard key light from camera left). - Color cast matches the Kelvin you set. Warm prompt -> warm image, neutral prompt -> neutral skin tones. A blue cast on a
3200Kprompt means the model ignored it; move the Kelvin value earlier in the prompt and remove competing time-of-day words.
If the file looks dimensional but the preview thumbnail still looks flat, the platform is softening the preview, not the image. Trust the downloaded file.
Prevention
- Never use a “lighting adjective” without backing it with direction + quality + temperature.
- Maintain a personal cheat sheet of 8-10 named lighting setups (paste-and-go).
- For portraits, default to Rembrandt or beauty-dish; rarely go front-lit.
- Pick a Kelvin temperature every time, even an approximate one. It beats
warm/cool. - For real-photo realism, skip
cinematic/dramatic gradingand ask fornatural lighting, no color grading.
FAQ
Why does studio lighting make every image look flat?
Because it isn’t one setup. The training data tags beauty dishes, ring lights, three-point rigs, and softboxes all as “studio lighting,” so the model averages them into a soft, characterless mid-point. Name the exact rig instead (single 4ft softbox at 45° camera right).
Does this work the same in Midjourney, GPT Image 2, and Nano Banana? Yes. As of June 2026 the photography vocabulary (direction, hard/soft, Kelvin, named setups) is model-agnostic. Midjourney v8 tends to render lighting with the most dramatic falloff, Nano Banana Pro and FLUX 2 Pro lean most photorealistic, and GPT Image 2 follows complex lighting prompts most literally, but the same words steer all of them.
My image came out way too orange/blue. How do I fix the color?
You almost certainly have either no Kelvin value or two conflicting ones. Set one Kelvin value near the front of the prompt (3200K warm or 5600K neutral) and delete competing time-of-day cues like having both golden hour and overcast in the same prompt.
Should I use cinematic for a realistic photo?
No. For authentic photo realism, cinematic and dramatic color grading often produce an over-graded, fake look (OpenAI’s own GPT Image 2 guide warns about this). Use natural lighting, believable detail, no color grading. Save cinematic plus a named director for when you want a stylized movie still.
What’s the single most important lighting word?
Direction. from camera left or backlit does more to add depth than any other single term, because it forces a light side and a shadow side and gives the face modeling.