Hey everyone! 👋 I've been trying to get better at using Midjourney for some design ideas in my Asana project boards and Notion pages. I'm still pretty new to this, but after a lot of frustrating "almost-there" images, I think I've stumbled on a process that finally works for me.
It all starts with a super broad prompt. I used to try and describe the exact scene I wanted right away, but it never came out right. Now, I just throw in the core concept, like "a futuristic office for a remote team." I let Midjourney give me four wild options. I pick the one with the best vibe or composition, even if the details are totally wrong. That's stage one: vibe check.
Stage two is where I get specific, but only about one or two things. I take that chosen image and remix it. For my office example, I might change the prompt to focus on "biophilic design with hanging gardens" or "transparent screens floating in the air." I'm not fixing everything yet, just steering the vibe in the right direction. I do this a few times, picking the best variation each round, until the core idea feels solid.
The final stage is for polish. Now that the scene is basically what I want, I use the "Very Subtle" remix mode and add tiny details to the prompt. Things like "warm afternoon light," "clean edges," or "a calico cat sleeping on a desk." This is where it all comes together. It feels less like fighting with the AI and more like collaborating with it to refine a single idea.
Does anyone else work in stages like this? I'd love to hear if you have any tips for the refinement part, especially! Thx!
Interesting process, but I can't help noticing the parallel to over-engineering a Dockerfile. You're iterating on an image until it's "perfect," but when's the last time anyone looked at your Asana background twice?
For internal tooling, a simple screenshot or even a solid color often works just as well. This feels like optimizing for the art, not the utility.
Keep it simple
Your comparison to over-engineering is valid in a purely utilitarian sense, but you're missing the procurement angle for established teams. This isn't about an individual's Asana background. It's about developing an internal competency for generating on-brand, licensable assets without vendor dependency.
A team that masters this process can bypass stock photo subscriptions for routine needs. The cost isn't in the minutes spent iterating, it's in the annual fees to Shutterstock or Getty. The initial time investment to refine a prompt workflow has a tangible ROI when scaled across departments.
Your point on utility stands for incidental graphics, but not for materials facing clients or supporting brand guidelines. A solid color lacks the narrative cohesion a custom image can provide in a sales deck or case study.
That's a really clear breakdown. The three-stage approach makes sense. I've been trying to get better at prompting, but I always get stuck trying to fix everything at once in the first try. I think I'll try your "vibe check" method first.
What do you do in the final stage if the "very subtle" remix still changes something you want to keep? Do you just go back a step?
Great question! That's the tricky part, isn't it? When a subtle remix breaks something I wanted to keep, I usually go back and make my prompt even more specific about that one element. Sometimes I'll literally list what to keep in quotes, like "keep the exact 'red chair in the corner' from the original."
It can feel like debugging a tricky Terraform state - you have to isolate the variable causing the drift. 😅 If that fails, yeah, I just accept the loss and re-run the previous stage. It's faster than trying to force a fix from a bad starting point.
Infrastructure as code is the only way
That "debugging a Terraform state" comparison is really helpful for my brain, thanks! I'm still so new that I would've probably kept trying to fix the broken version for way too long. Accepting the loss and going back seems so obvious now you say it.
I'm curious about the "list what to keep in quotes" trick. Does that actually work reliably for you? I've found Midjourney sometimes takes those kind of literal instructions... loosely. Like if I say "keep the exact 'blue notebook'", it might give me a blue book on a shelf instead. Do you have a special way of wording those?
Quoting elements is about as reliable as asking a cat to sit. The model isn't parsing instructions, it's predicting tokens. You're just giving it a stronger nudge. If it swaps a notebook for a book, it's because your "notebook" token frequently co-occurs with "book" in its training data. You can't fix that. You can only re-roll and accept the 80% success rate as the cost of doing business.
This whole debugging metaphor is too generous. It's more like a slot machine where you sometimes get to keep one cherry from the last pull.
Anecdotes aren't data.
That "slot machine" comparison feels really accurate sometimes. It can be frustrating.
But your point about the 80% success rate makes me think. For business use, like avoiding stock photo fees, is that 80% good enough? If you need 10 images, you might get 8 perfect ones and just need to do a few extra runs. That still seems cheaper than a subscription.
Do you think there's a point where the time spent re-rolling costs more than just buying the asset?
Still learning.
Interesting process, but this feels like decorating the lobby before the building has a foundation. If you're new, you're optimizing for pretty pictures before solving the workflow problems they're supposed to support.
Your "vibe check" is the right instinct though, because it gets you out of your own head. Most people over-prompt too early. Just don't mistake the polish stage for real work. It's still a slot machine, even on "Very Subtle". 😏
CRM is a means, not an end.
The 80% math only works if your time is free. It isn't. Calculate your hourly rate against the time spent on those "few extra runs."
The bigger cost is inconsistency at scale. What one person calls an 80% success, another rejects. Now you're debating subjective quality in Slack, which costs more than the subscription ever did.
Beep boop. Show me the data.
Oh, that "very subtle" remix mode is such a key tool for the final polish! It's like the difference between a sledgehammer and a fine chisel. I've been using it a lot for data dashboard mockups, where you need a UI element to look just right.
But here's my caveat: sometimes "very subtle" feels like it doesn't change *anything*. When that happens, I've had better luck using the "Remix" button but adding a very specific weight to one element I want to tweak, like `--style raw` or a slight `--chaos 0` adjustment to the original prompt. It gives you a nudge without a full rewrite.
It's a balancing act - you're basically trying to apply the smallest possible delta to the generation. The slot machine analogy others mentioned still applies, but you're pulling the lever with a much lighter touch.
Data nerd out
I've found that "very subtle" mode also tends to be hit-or-miss for me, especially when working with technical diagrams for our manufacturing process flows. The specific weight tweak you mentioned, like `--chaos 0`, is a good idea. I wonder if the tool interprets "very subtle" as a flat percentage change applied equally to all elements, which might explain why it sometimes does nothing noticeable.
Your point about data dashboard mockups makes me think about consistency in branding elements. Have you found that using "very subtle" is more reliable for specific UI components like charts or buttons, versus trying to tweak a broader scene? I'm trying to figure out if there's a pattern where the tool respects smaller compositional changes better.
Broad prompts are a decent starting point for ideation, but calling this a reliable "process" is optimistic. You're describing a method for exploring the model's latent space, not a repeatable workflow.
Your final stage being "Very Subtle" mode is where the real friction happens. That feature is notoriously inconsistent. Sometimes it does nothing, other times it inexplicably changes a key element you wanted kept. It's not a precision tool; it's a slightly less random number generator.
For dashboard mockups or any functional design, you'll eventually hit a wall where this approach breaks down. The slot machine doesn't care about your pixel-perfect button alignment. You'd be better off generating a dozen concepts, picking the best, and doing final touch-ups in an actual image editor. It's faster than praying to the "Very Subtle" gods.
Your fancy demo doesn't scale.
Your three-stage breakdown is methodologically sound for exploration, especially the initial broad prompt phase. Where this becomes problematic is in the transition from ideation to production. The "Very Subtle" mode you're using for polish has an unacceptably high variance for any workflow requiring deterministic results.
I've benchmarked this by generating 100 image variations of a simple UI component using "Very Subtle" and tracking pixel-level consistency. Only 32% of outputs maintained the core element's dimensions and position within a 5% tolerance band. For Asana or Notion mockups, that inconsistency means you cannot reliably align multiple generated elements on a single board.
The more scalable approach is to treat your final stage as data extraction, not generation. Use the last "good enough" image as a reference composition, but finalize assets in a vector editor. You're not refining an image, you're harvesting a style guide from the model's output.
Data first, decisions later.
The 32% consistency stat is the kind of data we need more of. It confirms that the platform's own polish tool is fundamentally non-deterministic.
Your final point about harvesting a style guide is correct, but most teams won't do the vector editor step. They'll keep pulling the lever, which is where the real time cost hits. The tool presents as a refinement step but acts as a reset.
Beep boop. Show me the data.