I keep seeing posts about Midjourney's 'consistent style' being a feature, but the flip side is a real problem. When I reroll a prompt, especially for technical or architectural concepts, the variations are minimal. It feels like I'm just getting a slightly different crop of the same base generation, not a new interpretation.
My use-case is generating reference images for cloud infrastructure diagrams and data center layouts. A true reroll should give me a different angle, component arrangement, or lighting scheme to choose from. Instead, I get:
* The same central server rack shape, maybe mirrored.
* Identical cable trunking patterns.
* Virtually the same color palette shift.
This isn't variance. It's lazy iteration. I've tried:
* `--chaos` values from 50 to 100.
* Adding `--no` for dominant elements.
* Appending "different perspective, unique layout, alternative design" to the prompt.
The results are marginal. The seed seems to lock in a composition too rigidly. For a tool at this price point, the reroll function should deliver more distinct options, not just tweaks.
Anyone running into this with technical subjects, or is it just me? Screenshots of your original and three rerolls would be helpful to compare.
show me the bill
show me the bill
You've isolated the core problem: seed-locked composition. It's particularly noticeable with technical subjects because the model's training data likely contains fewer variations of server racks than, say, portraits. The seed doesn't just set a style, it often defines a spatial template.
You mentioned using `--no`. Try applying it with surgical precision to the dominant structural elements. For a prompt like "modular data center layout, blue lighting", you'd need `--no central_server_rack uniform_cabling monotone_palette`. Simply `--no` without specific targets is ignored.
Also, chaos values above 70 can paradoxically reduce compositional variance for structured subjects, pushing the model into abstraction rather than novel arrangements. I've found better results with `--chaos 35` combined with explicit negative weights in the prompt, like `:: central server rack::-0.3`.
Yeah, definitely ran into this. The seed locking feels especially strong for prompts with clear physical structures. It's like the model picks a template from its training set and sticks to it.
A weird trick that sometimes works for me: run the original prompt once, then feed that *output image* into a new generation using `/describe`. Let it create a new prompt from the visual, then modify that new text prompt heavily before generating again. It can break the seed's compositional hold.
Have you tried using `--style raw` for technical subjects? It sometimes reduces the default aesthetic bias and allows for more variation on rerolls. Not a silver bullet, but worth a shot.