Say you're making a poster. You told the image generation model to put the title at the top, the date at the bottom left, and the person on the right. The result looks cool at a glance, but the date overlaps the person, and the poster text is misspelled.
Redoing it, this time you don't like the person's expression and the lighting. Starting over from scratch feels like a waste, and fixing it by hand feels daunting.
Thinking Mode, which Midjourney released on October 8, is a feature aimed at moments like this. It re-checks an already-made picture against the prompt, finds the requirements it missed, and generates once more.
It's as if they attached a step where the image AI checks its own result instead of delivering a perfect one from the start. Let's see how much it really helps.
Not a New Model, but a Feature That Rechecks the Result
First, we need to separate the dates. The Midjourney V8.2 model was released on July 24, and what was announced on October 8 is the Thinking Mode test. A new V8.2 model didn't come out in October.
✦ How to Use Thinking Mode
On the Midjourney Alpha site
Using it is simple. On the Midjourney Alpha site, open a standard generation or edit job made with V8.2, and press Rerun (Thinking) at the bottom right of the 'enlarged image view' screen.
Then the model looks at the result, finds the parts the prompt missed, and generates again. Wrong objects, misaligned layout, unnatural bodies, and text in the image are the typical things it checks.
✦ What Thinking Mode Checks
Problem types and what to keep an eye on
In other words, after making the picture, the image generation model itself looks over a checklist once more and fixes errors. But it isn't a feature that edits only a specific part of the image.
It generates the result from scratch again, so parts you didn't mean to change, like composition or color, can change.
The Advantages and Limits a Designer Found by Using It
If you only read the feature description, it sounds like everything would get better, but actual results can vary by type of work.
Nakano, who runs the Japanese design company DotProof, published the results of testing Thinking Mode with several prompts on October 12.
✦ Nakano's Test Results
They differed by image type
In a prompt drawing a full-body elf character, Midjourney's regenerated result brought out the armor's details better, and the composition placing the character in the frame came out more balanced.
In a poster image, the text in the first result was too small, and in the picture redone with Thinking it got a bit better. Still, the text wasn't at a perfectly readable level.
Regenerating after thinking one more time doesn't overcome the image generation model's own limits.
In a photorealistic image of apples, bananas, and tangerines in a basket, the result after Thinking changed to a more balanced composition. But if you want a natural, photo-like scene, the first image's composition could have been better.
A picture getting more tidy and getting closer to the photo you wanted aren't the same thing. In another portrait image, the awkward right hand and bow shape got better after regeneration, they recorded.
What's important in this experiment is that a 'more balanced picture' isn't always a 'better result.' Thinking Mode is a tool for re-matching conditions the prompt missed, not an all-purpose button that judges your taste and purpose for you.
In Nakano's test, too, repeated regeneration didn't improve the result every time.
Is Thinking Mode Worth Using or Not?
Midjourney said its own testing found a tendency toward better prompt adherence, text rendering, and scene consistency. But it didn't disclose figures such as by what percent the success rate rose.
The independent usage cases out so far are also experiences comparing specific prompts, not a large-scale test showing how much better every image gets.
If you've done a lot of image generation locally, you've probably run a workflow that inserts an LLM in the middle to review whether the generated image followed the prompt well, and regenerates if it doesn't pass.
But I'd guess the results weren't very satisfying. For starters, local LLMs aren't very good at recognizing and interpreting images. And connecting an expensive high-end API model isn't worth it for a feature that isn't that important.
And the cheap 'dummy friends' like GPT Luna, Claude Haiku, and Gemini Flash-Lite didn't show that big a difference from local LLMs in image-review performance. That's in my experience, anyway.
They'd even reject a properly made image at times, so it was common to needlessly throw away a well-made image.
The significance of Midjourney's Thinking Mode is that it's not a general-purpose LLM but an image-generation-specific LLM with a built-in verification function.
How much Thinking Mode will reduce these problems can only be known once more real usage results pile up, and when you try it yourself, it probably won't be as satisfying as you'd think at first.
But if this feature develops step by step to the point where Midjourney can properly filter out 'images that look good to the user,' image generation work really will become easy.
But will... that day really come? When I use the AIs everyone calls amazing these days, I often feel like we've hit the limit of LLMs. They get more and more 'plausible,' but they're not quite as satisfying as human work.
(And AI's lies and deceptions are also getting more and more cunning.) What do you think?
That's all for this post. See you next time.
- Nano Banana 2.1: 4K Images at $0.113 Each, Worth It?
- FLUX 3 Image Klein vs Qwen Image 2.1: Preserve or Blend?
- MOSS-Image Arrives in ComfyUI, and Commercial Use Is OK
🌐 Korean version: 10월 8일, 미드저니 8.2 추론모드 등장! 이미지 AI가 자기 실수를 스스로 고친다
No comments:
No comments yet. Want to leave the first one?
Post a Comment