Okay, so I've been using Midjourney for client projects for about a year now, and I've hit a wall. You know the one. The client says, "This looks great, but... it looks like a Midjourney image." That specific blend of hyper-detailed, slightly surreal, overly cinematic lighting. It’s become a recognizable aesthetic, and for brands wanting something unique, it’s a problem.
I’ve been experimenting with ways to break out of that default style. It's a bit like tweaking CRM workflows—you have to push against the platform's defaults to get a truly custom result. Here's what's worked for me:
* **Heavy, specific stylistic prompting:** Instead of `a product photo of a watch`, you go deep. `A flat lay product photo of a minimalist watch on concrete, shot on a 50mm lens, f/2.8, natural morning light, muted color palette, slight grain, 35mm film scan`. The more photographic and technical you get, the further you move from MJ's fantasy core.
* **Embrace `--stylize low` (or even `--style raw` in v6):** The high stylization is a huge part of "the look." Lowering it gives you more control and a more natural, less "artistically interpreted" result. `--s 50` is my new starting point for client work.
* **Reference artists *outside* the usual digital art sphere:** Everyone uses Greg Rutkowski. Try referencing specific photographers (e.g., `in the style of Shot by Campbell`), illustrators from different eras (mid-century advertising, woodcut prints), or even film directors for mood.
* **The double-remix trick:** Generate something close, then use `Vary (Region)` on a part, but **remix it** with a totally different medium instruction. Like, take a character and remix the face region with `a wood carving of...`. It introduces chaos that breaks the uniformity.
* **Post-processing is non-negotiable:** Just like no CRM works perfectly out-of-the-box. I always pull the image into another tool to adjust colors, add grain or texture overlays, and composite elements. This is where you truly divorce it from the MJ fingerprint.
The real challenge is consistency across a batch of images for a single client campaign. Anyone else cracked that code? Are you using custom preferences or Niji mode for more graphic styles? I'm considering building a library of personal style codes, similar to how I have templates for HubSpot deal stages.
Still looking for the perfect one
You're spot on with lowering the stylize value. I've found that `--style raw` paired with a very low `--stylize` (think `--s 20`) is essential for corporate clients who need predictability. It forces the model to stick closer to your prompt's semantics.
A technique I'd add: use photographic flaws as positive prompts. Things like "chromatic aberration," "lens flare from practical light source," or "subtle motion blur on background elements." Midjourney's default is a sterile, perfect render. Injecting real-world camera and film artifacts immediately breaks that CG feel.
Also, don't forget the power of negative prompting for this. Explicitly banning terms like "cinematic lighting," "hyperdetailed," "octane render," and "unreal engine" can save you dozens of gens.
Mike
The negative prompting suggestion is critical, but you have to treat it like a cache invalidation problem. Banning "cinematic lighting" might still leave related artifacts in the latent space. I've logged generations showing that you often need to ban correlated terms like "dramatic lighting," "volumetric lighting," and "studio lighting" to see a statistically significant drop in that signature look.
Your point about `--style raw` with low stylize is correct for predictability, though I'd add it introduces its own constraint: the output becomes highly sensitive to prompt order and token weight. It's a lower-latency path to a predictable result, but the variance can still spike if your prompt isn't deterministic. Think of it as moving from an eventual consistency model to strong consistency - you get correctness but pay a performance cost in prompt engineering time.
--perf