# Prompting Giant-Scale Photos That Keep a Real Likeness

> Published 2026-09-15 · https://www.promptzone.com/samir_arellano/prompting-giant-scale-photos-that-keep-a-real-likeness-1d90

The giant-person-in-a-city shot is one of the few compositions that separates [image models](/muhsin/midjourney-and-flux-the-new-kids-on-the-ai-image-block-okc) that understand scale from ones that only understand objects. Done right you get a person the size of a building, sitting in traffic, with a face that still belongs to the reference photo you supplied. Done wrong you get a normal-sized person badly composited over a model village. This is the template, the scale cues that make it read, and the reasons it usually fails.

## Scale is built from cues, not from adjectives

Writing `giant` in a [prompt](/tara_suzuki/chatgpt-prompt-engineering-2026-30-production-tested-patterns-master-guide-1pmc) does very little. A model has no absolute sense of size; it infers size from the relationships between things in the frame. If nothing in the image has a known real-world size, there is nothing to compare the subject against and the result just looks like a person photographed close to the camera.

So the prompt has to supply reference objects whose size everyone already knows, and it has to put them in the same frame as the subject. Cars, buses, bicycles, pedestrians, street lamps, traffic lights, ground-floor shopfronts. The more of these appear at a plausibly tiny scale next to the subject, the more convincing the illusion. Buildings alone are weaker cues than vehicles, because building size varies enormously and the model knows it.

Camera position does the other half of the work. A drone or aerial viewpoint is what makes the scale legible: from street level a giant subject fills the frame and you lose the environment that proves the scale. From above you see the subject and the miniature city in the same shot.

![Overhead aerial view of a city street with cars and pedestrians](https://images.rawpixel.com/editor_1024/cHJpdmF0ZS9sci9pbWFnZXMvd2Vic2l0ZS8yMDIzLTEyL2xyL21vbnoxMDcwMzk1LWltYWdlLmpwZw.jpg)

## The template

Slots in square brackets, so the same prompt can be run across a series without rewriting the structure:

```plaintext
A realistic drone photo of a giant man (face and body identical to the reference image)
sitting casually in a street in [city], one knee raised, hand resting on his leg.
He is wearing [clothing]. Tiny cars, buses, bicycles and pedestrians move around him,
while the typical buildings of [city] and the streets look minuscule next to his size.
The soft overcast light of [city] highlights the wet pavement.
```

Filled in, that becomes something like:

```plaintext
A realistic drone photo of a giant man (face and body identical to the reference image)
sitting casually in a street in Lisbon, one knee raised, hand resting on his leg.
He is wearing a grey hoodie and dark jeans. Tiny cars, buses, bicycles and pedestrians
move around him, while the tiled facades and narrow streets of Lisbon look minuscule
next to his size. The soft overcast light highlights the wet cobblestones.
```

It suits any model that accepts a reference image alongside a text instruction. OpenAI's GPT Image 1.5, announced in December 2025, handles this family of prompt; so do Google's Nano Banana Pro and ByteDance's Seedream line. On open weights, [FLUX Kontext](https://huggingface.co/black-forest-labs) and [Qwen-Image-Edit](https://huggingface.co/Qwen) accept the same image-plus-instruction shape, though identity retention across a big scale change is the hardest part for any of them.

## What goes in each slot

| Slot | What to put | What breaks it |
| --- | --- | --- |
| City | A named city with recognizable architecture | Generic `a city street` gives you generic anywhere |
| Clothing | Simple, describable garments | Logos and complex patterns smear at that size |
| Pose | Weight-bearing, casual, grounded | Standing upright wastes the frame and hides the environment |
| Light | The city's typical weather | Hard sun with no shadow direction named |
| Ground | A surface with texture and reflection | An unspecified ground plane reads as flat grey |

The overcast-light choice in the original is not decorative. Diffuse light is far easier for a model to make consistent across a subject that spans several city blocks, and wet ground gives reflections that tie the subject to the street instead of leaving it floating on top.

## Holding the likeness

Identity is the part that degrades first. The face occupies a small fraction of the frame in an aerial shot, and small faces are where models take liberties.

1. Supply a reference photo where the face is well lit, roughly frontal, and large in frame. A crop from a group photo will not survive.
2. Say explicitly in the prompt that the face and body must match the reference. Models weight the instruction as well as the image.
3. Do not stack heavy stylization on top. Every step away from photographic makes the face drift further.
4. If the result is right except for the face, run a second editing pass on the output with the same reference rather than rerolling the whole image.
5. For a series, keep one accepted result as the canonical version and reference it alongside the original photo on later runs.

![A person sitting on a curb with one knee raised](https://cdn.stocksnap.io/img-thumbs/960w/KTKWKG6OR9.jpg)

## Failure modes

- **It looks like a miniature diorama, not a giant person.** Usually caused by tilt-shift blur. Avoid the words `tilt-shift`, `miniature effect` and `macro`, and ask for deep focus across the frame.
- **Nothing communicates scale.** The environment came out empty. Name the small objects explicitly and put them in contact with the subject: traffic passing at his feet, pedestrians on the pavement beside his hand.
- **The subject is pasted on.** No shared lighting or contact shadow. Name the light source direction and ask for a shadow cast across the street.
- **The face is a stranger.** Reference image too small, too angled, or too far from the requested pose. Fix the input before touching the prompt.
- **The city is unrecognizable.** Models render generic European or generic American streets unless given specific architectural features. Name them: tiled facades, fire escapes, red brick terraces.

## Variations worth running

The sitting pose is the safest because contact points with the ground sell the weight, but the same template supports others: leaning an elbow on a rooftop, lying along a boulevard with feet at one intersection and head at another, peering over a row of buildings with only the upper face visible. Each one changes which scale cues matter. The peering-over variant, for instance, drops the vehicles entirely and depends on roofline detail instead, so the prompt has to describe the roofs in the same detail the street version gives the traffic.

## Takeaways

- Scale comes from named objects of known size sitting next to the subject, not from the word giant.
- Shoot from above. An aerial viewpoint is what keeps the environment in frame to prove the scale.
- Diffuse light and reflective ground make the composite hold together.
- Identity is the fragile part: strong reference photo, explicit instruction, minimal stylization, and a second pass rather than a reroll.
- Keep the prompt as a slotted template so you can change one variable at a time across a series.


## Related reading

- [Turning Objects Into Characters With AI Image Editors](/aleksandr_herrera/turning-objects-into-characters-with-ai-image-editors-2clk)
- [Turning a Reference Photo into a Trading Card Illustration](/khalid_laurent/turning-a-reference-photo-into-a-trading-card-illustration-52fh)
- [Building Multi-Panel Story Grids from a Reference Photo](/klaus_kamau/building-multi-panel-story-grids-from-a-reference-photo-23bk)