# Codename Models on Image Arenas and How to Read Rankings

> Published 2026-09-23 · https://www.promptzone.com/thu_choudhury/codename-models-on-image-arenas-and-how-to-read-rankings-2mm4

Every so often an [image model](/muhsin/midjourney-and-flux-the-new-kids-on-the-ai-image-block-okc) nobody has heard of appears near the top of a public arena leaderboard under a name like a codename rather than a product. A few days or weeks later a lab confirms it was theirs. If you want to know how much weight to put on that ranking - and on leaderboard position generally - it helps to understand what the arena is measuring and what the codename is for.

## What a codename on an arena is

Public model arenas serve anonymous outputs: you submit a [prompt](/tara_suzuki/chatgpt-prompt-engineering-2026-30-production-tested-patterns-master-guide-1pmc), get results from two unnamed models, and vote for the one you prefer. Votes across many users feed a pairwise rating system in the Elo family, which turns individual comparisons into a single ordered ranking.

Because outputs are anonymous by design, an arena is a convenient place to field-test a model that has not been announced. It gets listed under a pseudonym, accumulates votes against everything else on the board, and the lab watches the rating settle. Users notice a strong unfamiliar name, speculation starts, and eventually the lab confirms it.

This has happened repeatedly. A model listed as `gpt2-chatbot` drew a lot of attention on text arenas in 2024 before being connected to OpenAI. An image model appearing as `red panda` in late 2024 turned out to be Recraft V3. In April 2025 an image model listed as `Mogao` was identified as ByteDance's Seedream 3.0. The pattern is now routine enough that a strange name near the top of a board is generally assumed to be someone's unreleased release candidate.

![Two framed photographs displayed side by side for comparison](https://images.rawpixel.com/editor_1024/cHJpdmF0ZS9sci9pbWFnZXMvd2Vic2l0ZS8yMDIyLTA2L3Vwd2s2MTg5NTYwNy13aWtpbWVkaWEtaW1hZ2Uta293cDB1YnguanBn.jpg)

## Why labs bother

Four reasons, roughly in order of importance:

- **Clean preference signal.** Brand affects judgement. Voters who know which lab produced an image rate it differently. Anonymity removes that, which is the whole point of blind evaluation.
- **Volume no internal test can match.** Arenas deliver a large, adversarial, self-selected stream of real prompts, including all the weird ones an internal eval set would never contain.
- **Comparison against live competitors.** Head-to-head against whatever is currently deployed, on the same prompts, without running any of it yourself.
- **Deniability.** If the model underperforms, it can be quietly withdrawn. Nothing was announced, so nothing failed.

The cost to the lab is essentially zero and the information is good. Expect the practice to continue.

## What the score actually measures

A blind arena vote is a human preference judgement made in a few seconds, on a screen, usually at moderate resolution. That biases what wins in a specific and predictable direction.

What gets rewarded is immediate visual appeal: strong contrast, saturated colour, pleasing composition, clean faces. What gets under-weighted is anything requiring inspection - whether the image contains all six objects you asked for, whether text is spelled correctly at full resolution, whether hands survive a zoom.

So an arena rating is a good proxy for *aesthetic preference at a glance* and a mediocre proxy for *instruction following*. Models tuned hard for the first can outrank models better at the second. That is not cheating; it is what the metric asks for.

## What a top rank does not tell you

| The arena measures | The arena ignores |
| --- | --- |
| Which image people prefer at a glance | Cost per image |
| Broad aesthetic quality | Generation latency |
| Rough prompt plausibility | Licensing and commercial terms |
| Performance on typical prompts | Whether weights are downloadable |
| | Editing, inpainting and reference support |
| | Character consistency across a series |
| | Maximum native resolution |
| | API availability and rate limits |

The right-hand column is usually what decides whether a model is useful in an actual project. A model that wins on preference but ships behind an expensive API with restrictive commercial terms may be worse for your work than an open-weights model several places below it.

![A crowd of people raising their hands to vote](https://cdn.stocksnap.io/img-thumbs/960w/QBGU3ISCKI.jpg)

## Build a personal eval set instead

The durable habit is to stop treating leaderboards as a verdict and start treating them as a shortlist. Then run your own comparison, which takes less effort than it sounds:

1. Write six to ten prompts covering things you personally need: a portrait, a product shot, a scene with a specific object count, something with legible text, a style you use often, and one deliberately awkward request.
2. Keep them in a text file, verbatim. The value comes entirely from reusing the same prompts across models over time.
3. Run each prompt on the candidate model at its recommended settings, not yours. Fine-tunes and closed models have different defaults for a reason.
4. Score against your own criteria, not general preference. If your work needs accurate text, weight that heavily and ignore everything else.
5. Save the outputs with the model name. In six months this archive tells you more about real progress than any rating curve.

A good eval prompt is short, unambiguous, and has an obviously correct answer you can check by looking. This one is a useful member of that set, because it tests instruction following, symbol interpretation and a specific rendering style all at once:

```plaintext
A 3D Pixar-style mascot combining these emoji: 🌶 😉
```

Swap the emoji pair and re-run. Weak models produce a generic 3D character that ignores one of the two inputs; stronger ones fuse both concepts into a single coherent design. It takes seconds to judge, which is exactly what you want from an eval prompt.

![A brightly coloured cartoon character figurine on a plain background](https://cdn.stocksnap.io/img-thumbs/960w/JBFGUYFDMW.jpg)

## Practical takeaways

A codename on an arena is a lab running a blind test, not a mystery - and the eventual reveal changes nothing about the images you already saw. Arena ratings measure quick human preference, which correlates with aesthetics far better than with instruction following, so read a high rank as evidence that a model makes appealing pictures rather than obedient ones. Everything that usually determines whether you can actually use a model - price, licence, latency, editing support, whether the weights are downloadable - sits outside the metric entirely. Keep a small fixed set of your own prompts and re-run it on each new candidate; that comparison is worth more to you than the ranking that pointed you at it.

## Related reading

- [Anime Generation With SDXL: How Tag Prompting Works](/arif_wu/anime-generation-with-sdxl-how-tag-prompting-works-cee)
- [Consistent Characters in Fooocus Without Training a LoRA](/thandi_fischer/consistent-characters-in-fooocus-without-training-a-lora-2jh9)
- [Fooocus Image Prompt: Steering Output With Reference Images](/thu_choudhury/fooocus-image-prompt-steering-output-with-reference-images-11n7)