For batches of ad images, our 12-brief comparison favored GPT Image 2 for following instructions and keeping characters consistent. Midjourney v7 produced the most cinematic individual stills, while Flux 2 offered a lower-cost option for producing many images and selecting the strongest.
Those are different production jobs. One campaign may need a single finished image; another may need the same person across many scenes. The model you choose should fit that requirement.
The short version: GPT Image 2 and Flux 2 tied on how realistic the average image looked in our test. GPT followed instructions more closely and kept batches more consistent. The comparison below includes usable-image costs after rejected outputs. Results reflect these briefs and model versions; use the same scoring method when testing newer tools.
What does “best for performance creative” mean?
We use the same five-question rubric every time we test a new model. We call it the Five-Lens Image Test™. Most public benchmarks only look through the first lens.
- Photorealism. Does the face look real on a 6.7-inch phone screen at full brightness?
- Character consistency. Can you make the same person show up in 12 different scenes and still look like the same person?
- Prompt adherence. When you ask for “Korean woman, 28, in a kitchen at golden hour, holding a phone,” do you get exactly that, or do you get whatever the model felt like that day?
- Batch control. Can you script the prompt? Version-control it? Re-roll one variant without re-running the whole set?
- Cost per usable variant. Not cost per generation. Cost per image you ship.
Twitter benchmarks usually only score #1, and even structured leaderboards like the Artificial Analysis image-model arena lean heavily on best-of-set photorealism. That is a hero-image question. Performance creative is a workflow question, and the workflow is what we score.
The 12-brief test
We picked 12 briefs across the kinds of work we ship for clients: train commuter, bedroom confessional, golden-hour window selfie, group chat reaction, desk workspace, two-creator split, kitchen scene, and a few more. We ran each brief through all three models, tweaking the prompt format to fit each one. Then we scored on the five questions above using a blind 1-to-5 rubric.
Here is the summary, with detailed notes per model below.
| What we tested | GPT Image 2 | Midjourney v7 | Flux 2 |
|---|---|---|---|
| Plans the image before drawing it (reasoning) | Yes | No | No |
| Photorealism (average frame across 12 briefs) | 4.5/5 | 3.5/5 | 4.5/5 |
| Character consistency across 12 variants | 4.5/5 | 2.5/5 | 4/5 |
| Prompt adherence | 5/5 | 3/5 | 4/5 |
| Batch control / scriptability | 5/5 | 1/5 | 4/5 |
| Cost per usable variant | $0.20 to $0.50 | $0.40 to $1.20 | $0.10 to $0.30 |
| Best fit for performance creative | Default | Hero stills only | Cost-sensitive backup |
Pay particular attention to “cost per usable variant.” Multiply the number of generations needed for one image you can ship by the cost of each generation. That includes rejected outputs, which a price-per-generation comparison misses.
GPT Image 2
GPT Image 2 is OpenAI’s newest image model, released on April 21, 2026. We use it through the API and the Codex CLI, and it is our default.
Here is the part most reviews skip, and the reason we switched. GPT Image 2 reasons about the picture before it draws it. Midjourney and Flux are diffusion models. You hand them a prompt, they turn it into a vector, and they translate that vector to pixels in one shot. There is no planning step or final check.
According to OpenAI’s prompting guide, GPT Image 2 runs a reasoning pass before generation. It interprets the brief and plans what the scene needs and where each element belongs. After generating the image, it checks the output and re-renders when needed.
That sounds abstract until you see it in practice. Ask Midjourney for “two people in a kitchen, the woman holding a coffee mug.” Half the time the man ends up with the mug, because the diffusion process never asked itself who has it. Ask GPT Image 2 the same thing and the answer is right almost every time, because the planning pass settled the question before the pixels existed.
That reasoning is what powers everything else. Prompt adherence stops being a coin flip, and character consistency holds across 12 scenes as long as you describe the character once and reuse the description. The stuff diffusion models used to garble, like multi-element layouts, on-image text, hand positions, and eye-line direction, GPT Image 2 mostly gets right on the first try.
What GPT Image 2 loses on: pure cinematic feel. Midjourney still produces more “wow” stills if you are picking one image for a billboard. For performance creative, that does not matter. The algorithm and the audience reward consistency and specificity far more than magazine-cover polish.
Midjourney v7
Midjourney v7 produces the best individual hero shots in this comparison, full stop. If you are making one image and you want it to look like a film still, Midjourney is still the answer.
For performance creative, it is not. The two structural problems are batch control and prompt adherence.
Batch control is the bigger issue. Midjourney runs through Discord. There is no first-party API. Third-party wrappers exist but are unofficial, rate-limited, and break when Midjourney updates. We tried scripting MJ for a 12-variant sprint and burned a half day on plumbing. Same sprint in GPT Image 2 took 20 minutes.
Prompt adherence is the smaller issue but matters at scale. Midjourney has its own aesthetic point of view, which is great when that aesthetic matches your brand and brutal when it doesn’t. Asking for “subtle, candid, slightly imperfect lighting” returns something cinematic and intentional anyway.
Where Midjourney still wins: hero brand stills, key art, anything that gets used once and seen many times. Different job.
Flux 2
Flux 2 is the open-weights wildcard. It runs on fal.ai, Replicate, or self-hosted, which makes it the cheapest of the three per generation by a wide margin. According to fal.ai’s published pricing, Flux variants run a small fraction of a dollar per image at standard resolution, and on Replicate’s Flux 1.1 Pro listing the per-image cost is similarly low. Both run well below GPT Image 2’s API rate at the same resolution.
Across the 12 briefs, Flux 2 tied GPT Image 2 at 4.5 out of 5 for average-frame photorealism. It followed prompts well, though GPT Image 2 did better and kept characters slightly more consistent. Batch control was a strength, supported by an open, well-documented API.
Where Flux 2 wins: cost-sensitive sprints where you want to generate 100 variants cheaply and curate down to 20. The low cost per image lets you over-generate and pick.
Where Flux 2 loses: face consistency across long sprints, and edge cases in non-Western character types where the model’s training data is thinner.
Our pick and how we use them
For most performance creative work, we reach for GPT Image 2 through Codex. The reasoning pass plus the character consistency is exactly the tradeoff paid social needs.
For high-volume, cost-sensitive sprints, we use Flux 2 through fal.ai. The per-image cost lets us over-generate and pick the best ones.
We do not use Midjourney for performance creative anymore. We still love it for hero brand work and pitch decks, where one beautiful still does the whole job.
If you want the one-glance answer, this is the cheat sheet we share with clients:
| If you are doing this | Use this | Why |
|---|---|---|
| 8+ variants of the same person across different scenes | GPT Image 2 | Reasoning pass keeps the character consistent |
| 100+ variants where unit cost matters most | Flux 2 via fal.ai | Cheapest per usable image at scale |
| One cinematic hero still for a deck or billboard | Midjourney v7 | Best single-image craft |
| Anything with on-image text, signage, or packaging copy | GPT Image 2 | 95%+ text accuracy, no other model is close |
When this comparison will be wrong
These choices have a shelf life. Image model leadership changes every six months, and Midjourney was the obvious default a year ago. A year from now, Flux 3 or an unfamiliar open-weights model could lead. Keep the Five-Lens Image Test™ and be ready to change the tool.
If you are setting up a performance creative pipeline in 2026, GPT Image 2 is the default. If you are reading this in 2027, run the Five-Lens Image Test™ against whatever models are leading then, and pick again. We will too.
Where this fits in the larger workflow
This is the image-generation step of a three-step pipeline. The other two are creative direction (a human creative director paired with a Claude Code agent) and animation (Seedance 2.0 or Kling 3.0, batched through Lovart). We covered the full workflow in our AI performance creative case study, including the campaign that dropped CPA 50% on TikTok.
Up next: The Playbook. This is the model-comparison chapter. The full hub, The AI Performance Creative Playbook, covers the workflow, economics, the rest of the 2026 tool stack, and the first 30 days of standing this up.
The right model is the one that can deliver your brief reliably at the volume you need. Our AI performance creative service connects that tool choice to direction, production, and campaign testing.
A representative brief shows us more than a list of preferred tools. In a free strategy call, we can discuss your formats, production bottleneck, and review requirements together. We’ll help you decide where AI belongs in the process.
Like this? Get the next one.
Short emails. New posts as they ship.
