Midjourney Tests Thinking Mode: A Rerun (Thinking) Button for Better Text Accuracy

Midjourney has started testing Thinking Mode for image generation on its Alpha site, alpha.midjourney.com. You turn it on with a single "Rerun (Thinking)" button in the bottom-right corner of the lightbox for a finished job. According to the team, internal tests show improvements in three areas: prompt adherence, typography and overall image coherence. This is an invitation to test, not a full release. The announcement includes no figures, benchmarks, generation-time data or pricing.

How it works in the interface

The mechanics are simple. You open an already generated job in the lightbox and, instead of the regular Rerun, click the variant labeled Thinking. The model processes the same request again, but with an extra "thinking" step. Midjourney does not say what happens during that step. It could be composition planning, prompt parsing, self-checking or something else; that is unknown.

Facts from the announcement:

• platform: the Alpha site, alpha.midjourney.com;
• launch: "Rerun (Thinking)" under the job's lightbox;
• claimed improvements: prompt accuracy, typography, coherence;
• plans: possible wide availability as a "thinking" mode or as a way to "add more thinking" after a job;
• feedback: the #ideas-and-features channel, examples welcome.

Why images need "thinking"

The list of weak spots is not random. Typography has long been a sore point for generative models: letters break, words turn into pseudo-text, and long captions fall apart. Prompt accuracy shows up in complex requests, where the model has to hold many objects, their relative positions and their attributes. Coherence keeps a scene from falling apart into details that look fine on their own but contradict each other when put together.

If you give a model extra time to think, as language models already do with reasoning modes, these are the tasks that benefit first. The price is usually known in advance: more compute per result. Midjourney says nothing about latency or GPU time, so don't expect the speed of regular generation. The phrase "as a way to add more thinking after a job" hints at a metered-use model: a fast draft first, then an expensive pass for the works that deserve it.

There is also a strategic layer. Midjourney has long competed with models that understand text instructions better. Betting on a "thinking" pass lets it tighten discipline within its own aesthetic without replacing the entire base model.

How to test it usefully

The team is asking for more than likes: it wants examples of where the mode helps and where it doesn't. To make your report useful, compare in pairs. Take the same prompt and seed, if available, run a regular Rerun and a Rerun (Thinking), and save both versions.

Good candidates for testing:

• posters and covers with a headline of 3–6 words;
• scenes with several objects and a clear layout ("on the left… on the right… in the foreground…");
• prompts with negations and clarifications that models usually ignore;
• stylized works, where the question is whether the mode loses Midjourney's signature character.

Note side effects separately: an overly dry composition, lost atmosphere, extra waiting time. Remarks like these will decide whether Thinking becomes a separate mode or an option after generation.

For now this is an experiment on the Alpha site, and it is too early to build workflows on it. But designers who need text in an image without manual touch-ups should try it now.