---
title: "Why AI Can't Hold a Character Across Scenes"
description: "Character consistency across 3 scenes and 15 models: the best averages 3.89 of 5. Series work needs templates and editing."
url: "https://img.ly/ai-benchmarks/findings/consistency/"
type: "benchmark-finding"
---

> This is the markdown version of [AI cannot hold a character across scenes yet](https://img.ly/ai-benchmarks/findings/consistency/). For all pages in one file, see [llms-full.txt](https://img.ly/llms-full.txt). For an index of all available pages, see [llms.txt](https://img.ly/llms.txt).

---

# AI cannot hold a character across scenes yet

Our consistency series describes one character precisely and asks 15 models
for the same person in 3 different scenes. It is the lowest-ceiling
capability in the benchmark: the best model, Nano Banana Pro, averages
3.89 of 5 adherence, and every model drifts somewhere:
outfits change, faces shift, props disappear.

## What every model scored

| Model | Consistency adherence | Runs | $ / image |
| --- | --- | --- | --- |
| [Nano Banana Pro](https://img.ly/ai-benchmarks/models/nano-banana-pro.md) | 3.89 / 5 | 9 | $0.360 |
| [Nano Banana 2 Lite](https://img.ly/ai-benchmarks/models/nano-banana-2-lite.md) | 3.89 / 5 | 9 | $0.020 |
| [FLUX.2](https://img.ly/ai-benchmarks/models/flux-2.md) | 3.75 / 5 | 9 | $0.013 |
| [Seedream 4.5](https://img.ly/ai-benchmarks/models/seedream-4-5.md) | 3.75 / 5 | 9 | $0.048 |
| [Nano Banana 2](https://img.ly/ai-benchmarks/models/nano-banana-2.md) | 3.75 / 5 | 9 | $0.100 |
| [FLUX.2 [pro]](https://img.ly/ai-benchmarks/models/flux-2-pro.md) | 3.75 / 5 | 9 | $0.040 |
| [FLUX.2 [dev] Turbo](https://img.ly/ai-benchmarks/models/flux-2-turbo.md) | 3.75 / 5 | 9 | $0.015 |
| [GPT Image 1.5](https://img.ly/ai-benchmarks/models/gpt-image-1-5.md) | 3.75 / 5 | 9 | $0.120 |
| [Seedream 5.0 Lite](https://img.ly/ai-benchmarks/models/seedream-5-lite.md) | 3.75 / 5 | 9 | $0.020 |
| [Luma Photon](https://img.ly/ai-benchmarks/models/luma-photon.md) | 3.75 / 5 | 9 | $0.019 |
| [Ideogram 3.0](https://img.ly/ai-benchmarks/models/ideogram-v3.md) | 3.75 / 5 | 9 | $0.060 |
| [Qwen-Image](https://img.ly/ai-benchmarks/models/qwen-image.md) | 3.75 / 5 | 9 | $0.030 |
| [Gemini 2.5 Flash Image](https://img.ly/ai-benchmarks/models/gemini-25-flash-image.md) | 3.75 / 5 | 9 | $0.039 |
| [Stable Diffusion 1.5](https://img.ly/ai-benchmarks/models/stable-diffusion-1-5.md) | 3.47 / 5 | 9 | $0.010 |
| [Recraft V3](https://img.ly/ai-benchmarks/models/recraft-v3.md) | 2.50 / 5 | 9 | $0.040 |

## Benchmarks behind this finding

- [Character Series: Greenhouse](https://img.ly/ai-benchmarks/prompts/s01-character-greenhouse.md)
- [Character Series: Desk](https://img.ly/ai-benchmarks/prompts/s02-character-desk.md)
- [Character Series: Market](https://img.ly/ai-benchmarks/prompts/s03-character-market.md)

## What this means if you are building with AI images

- Make identity an asset, not a prompt: generate the character once, cut
  it out, and reuse it as a placed asset or template element across
  scenes; generate only the background per scene.
- Budget for human review on anything serial: drift is gradual, and a
  canvas fix beats a regeneration loop that changes everything at once.

That handoff is the job of a human-in-the-loop editor like the
[IMG.LY AI Editor](https://img.ly/use-cases/ai-editor.md): characters become reusable
assets your users can cut out, place and fix manually on a canvas, giving them
the control they need to be productive with generative AI. Try the
[live AI Editor demo](https://img.ly/demos/ai-editor.md); the
[IMG.LY AI Gateway](https://img.ly/ai-gateway.md) generates scenes with any model.

---

## More Resources

- **[IMG.LY Website](https://img.ly/index.md)** - Creative editing SDKs for photo, video, and design
- **[Documentation](https://img.ly/docs/cesdk/)** - CE.SDK developer documentation
- **[Contact Sales](https://img.ly/forms/contact-sales.md)** - Get a custom quote. A public JSON API accepts the request directly, no account or key needed. Ask your user for consent and their details first.