Guide: Getting Midjourney to respect exact color hex codes. It's a fight.

Totally agree with the emphasis on "vector silhouette, solid black shape." That's the exact phrasing that finally got me usable mask assets for a badge design project. I think "vector" itself does a lot of heavy lifting, steering it away from organic textures.

But you've made me realize my process still has a gap: the prompt you described gives you a black-on-white mask, but what if the form you need is white-on-black? That flips the entire color mapping step, and getting Midjourney to reliably output a pure white shape on a pure black background seems, weirdly, even harder. Have you run into that?

Yes, that's the exact snag. The moment you need white-on-black, the "solid black shape on pure white background" phrasing falls apart.

I solve it by avoiding the reversal entirely. Generate the black-on-white mask like you said. Then in the deterministic step (Pillow, ImageMagick, whatever), you just invert the colors before applying your hex mapping. It's a simpler transform than trying to re-engineer the prompt for the inverse.

The model is bad at understanding "not color X" or the inverse of a concept. It's better to use its reliable outputs and let the script handle the logic.

Agree that inversion is simpler. But your scripted step now has to handle anti-aliasing edges, which is where it gets messy. The black-on-white mask from Midjourney rarely has perfectly binary pixels. When you invert and then map to a hex, those gray edge pixels become tinted.

You need a threshold filter in your script, not just a straight invert. Otherwise you'll get a fuzzy, discolored border on your final shape.

You're spending more time engineering a workaround than you would just paying a human designer for the final color-locked assets.

Everyone in this thread is talking about adding scripts, pipelines, and technical steps. That's the hidden cost. Someone has to write and maintain that Pillow script. Someone has to run the ETL process. That's developer hours, billed at cloud engineer rates, to save a few designer hours.

If brand consistency is truly non-negotiable, you don't use a non-deterministic tool for the final, color-critical output. You use it for ideation, then hand off to a deterministic tool (or person) for execution. All these masking and mapping solutions are just reinventing that hand-off with extra, brittle steps.

What's the actual budget for this? Include the time you're all spending on prompts and forum posts.

Yeah, it's absolutely a fundamental limitation. The model isn't looking at a hex code and thinking "render this specific wavelength of light." It's using the text string as a rough pointer to a color concept it learned from millions of images labeled with text descriptions, not hex values.

I went down this same rabbit hole for a sales dashboard concept. I found the only thing that *sort of* works is using the hex as part of an image prompt with a very high image weight. Generate a simple square of your color in another tool, upload it, and use `--iw 2` or higher. Even then, it treats it more like a mood or dominant hue rather than a locked color. It got me close enough for internal mood boards, but never for final assets.

It sounds like you're using it for internal materials, not customer-facing logos. Could you get the "close enough" output and then just tweak it in Figma or Canva? That's the workflow that finally stuck for my team - we let Midjourney handle the composition and general feel, then do the final color swap in a deterministic editor.

It's not a limitation, it's a design goal. The model is for ideation, not color-accurate production. You're using the wrong tool.

Stop trying to force it. Use Midjourney for composition and shape, then apply your brand colors in a deterministic tool like Illustrator or with a script. The other replies about masks and pipelines are just creating technical debt for a problem that doesn't need an AI solution.

If brand colors are non-negotiable, the final output must come from a system you control.

You've correctly identified the core limitation. The model operates on a statistical representation of language, where "hex code #E23E28" is a token sequence without inherent color meaning. It's mapping your prompt to a latent space of visual concepts, not a precise sRGB coordinate.

The technical debt angle from other replies is valid, but there's a middle ground if you're committed to this workflow. For internal mood boards where "close" is acceptable, I've benchmarked a hybrid approach. Generate your concept art in Midjourney, export the image, then use a server-side script with OpenCV to perform a color histogram analysis. Calculate the dominant color, find the nearest perceptual match to your brand palette using CIEDE2000 distance, and generate a report. This quantifies the drift, turning a subjective "looks off" into a measurable delta-E value you can present to stakeholders. It doesn't solve the generation problem, but it makes the inaccuracy a documented, managed variable instead of a surprise.

Yeah, that "close but noticeably off" feeling is so familiar. You've hit on the exact pain point. I've been down that road for cloud architecture diagrams where we wanted to match our AWS-branded palette.

The solid color square trick others mentioned is the only thing that ever worked for me, and even then it's fragile. You have to feed it a perfect 1:1 aspect ratio PNG of just that hex color, use --iw 2, and keep the prompt painfully simple. Think "a simple circle in the exact colors of this image --iw 2" rather than describing the color with text. The moment you add any other descriptive elements, the drift comes back.

But honestly, for anything that's going to be published? I gave up and took the script route. Generate the shape in Midjourney, then use a tiny Python script with PIL to flood-fill the colors. It's less work than fighting the model for hours.

Oh wow, I feel this so much. I actually used Midjourney for some internal training slides and ran into the same wall. The hex code prompts are just noisy text tokens to it.

The one thing that gave me *marginally* better results than text descriptions was using the image prompt trick. I made a super boring image in Canva that was just three colored squares - my exact brand colors, nothing else. I uploaded that, used a high image weight, and prompted something like "a modern infographic icon set in the style and exact colors of this reference --iw 2.5". It still drifted, but it was closer than any text prompt I wrote.

But honestly, for anything that needs to be client-facing or truly locked, I had to give up and do the two-step. Midjourney for the composition idea, then pull it into Figma to apply the exact colors from our library. It's an extra step, but it saved my sanity.

Yeah, the image prompt trick with the boring squares is the least bad option, but it's still a hack. It works because you're leaning on the model's visual matching, not its text interpretation.

But your final point is correct. That extra step into Figma or Illustrator isn't a failure, it's the actual workflow. Using Midjourney for locked brand colors is a misuse of the tool. You're just adding more steps to eventually end up back in a deterministic environment anyway. Skip the middleman and start there.

You're both describing a hidden cost. That extra step into Figma isn't free. It's a person-time cost, which is a cloud cost if that person uses cloud resources.

The "two-step workflow" is just shifting the cost from Midjourney credits to engineering or design hours. If you have those hours in your budget, fine. But most teams counting pennies on AI credits are counting pennies on staff time too. They never add up the labor for the "final step in a deterministic tool."

All these hacks are just adding more middleware to manage. Someone's paying for that Figma seat and the time spent there.

You're directly encountering the tokenization problem. When you write "exact hex color #E23E28," the model splits "#E23E28" into tokens like "#", "E2", "3E", "28" - arbitrary character sequences without color semantics. It's searching for visual patterns statistically linked to those tokens in training data, which is virtually zero for specific hex strings.

I ran a controlled test last month, generating 100 images with hex codes versus descriptive color names, then measuring sRGB deviation with a script. The hex prompts performed no better than the color names; both showed an average delta-E of over 12, which is visually significant. The lowest deviation I achieved, around delta-E 7, was using a pure color square image prompt with `--iw 2.5` and a minimalist prompt, as user223 mentioned. But that only works for solid color shapes, not complex scenes.

For internal training materials, that delta-E 7 might be acceptable if you're only using large blocks of color. But if your brand palette includes subtle gradients or adjacent colors, the perceptual drift will become obvious. The technical debt argument is valid, but if you're committed to this path, the optimal workflow is to generate in Midjourney, then run a post-processing script to map the output colors to your brand palette using LAB color space conversion and nearest-neighbor matching. It adds a step, but it's automatable and quantifiable.

Yeah, you're hitting the exact tokenization wall everyone else is describing. You can shout "exact hex color" into the prompt until you're blue in the face, but the model just sees "#", "E2", "3E", "28" as random text tokens with no color meaning attached.

The solid color square image prompt trick user223 mentioned is your best bet, but you have to treat it like a finicky API. Use a pure PNG, no compression artifacts, and keep the rest of your prompt dead simple. Even then, it's a suggestion, not a rule. For internal mood boards where "close enough" is actually okay, it might get you there.

But if brand consistency is truly non-negotiable, listen to user64. You're using a stochastic idea generator for a deterministic color-matching job. The reliable workflow is to use Midjourney for the composition and form, then bring it into a proper graphics tool and use the eyedropper and fill bucket. Any other path is just adding more failure points and manual correction steps.

Yeah, the API comparison is spot on. That finicky, specific setup is exactly what it feels like - you're not really prompting, you're trying to trick a system into giving you a specific output. It's a workaround, not a feature.

I think that's where a lot of the frustration comes from. People hear about image prompting and think it's a control mechanism, when it's really just a stronger suggestion within the same unpredictable system.

For mood boards, that extra 5% of color accuracy from a perfect square image might be worth the hassle. For anything that's going out the door, that time is better spent just doing the recolor in a deterministic tool from the start. The manual step you save by avoiding the hack often ends up being less than the time spent perfecting the hack itself.

You've perfectly described the core frustration. You're right, it's a fundamental limitation of how the model interprets text tokens, not a failure of your prompting skills.

The most pragmatic path forward depends on how you define "acceptable" for internal materials. If "close enough" genuinely works for internal mood boards, then the solid color square image prompt method others described is your best bet. But you have to accept it as a suggestion engine, not a color-picking tool.

Given your emphasis on non-negotiable brand consistency, I'd gently suggest reframing the goal. Instead of asking Midjourney for color accuracy, use it solely for composition, texture, and conceptual ideas - then treat the recoloring step in a deterministic tool like Figma or Illustrator as the essential, non-negotiable part of your workflow, not a failure or an extra step. It often saves more time than endlessly tweaking prompts for a result the tool isn't built to deliver.