Hello everyone,
I've been using Midjourney for a few months now, mostly for creating individual marketing assets and concept art, but I’ve hit a significant challenge with a new project. I need to create a storyboard (approximately 20 images) that tells a simple narrative, and the main character must be recognizably the same person from scene to scene. My initial attempts have resulted in a cast of characters who look like cousins at best, with inconsistent facial features, hairstyles, and even body proportions.
Given my background in marketing automation, I’m accustomed to systems where you can define a variable (like a customer profile) and have it persist. I'm trying to approach this with a similar mindset, but I'm not finding a clear workflow.
I’ve done some preliminary research and testing, and I have a list of specific questions. My goal is to understand the exact steps and parameters, as I prefer to double-check every detail before committing to a full storyboard generation.
* **Seeds:** I understand the concept of using a seed from an initial image. However, if I change the prompt significantly for a new scene (e.g., from "character in an office" to "character running in a park"), does the seed alone maintain facial consistency, or does the changed context override it?
* **Reference Images:** I've read about using image prompts. Would the most reliable method be to generate a perfect base image of the character first, and then use that image's URL as a reference in *every single subsequent prompt*? Should this be combined with a seed?
* **Prompt Engineering:** What specific descriptive elements are most "sticky" for character consistency? I'm thinking about:
* Very detailed facial descriptions (e.g., "oval face, sharp jawline, wide-set hazel eyes, a small mole on the left cheek").
* Specific hairstyle, color, and hair texture.
* Clothing style, if the character remains in the same outfit.
* **Parameters:** Beyond `--seed`, are there other parameters like `--style` or `--stylize` that I should keep constant, or perhaps deliberately lower, to reduce variation between images?
* **Workflow Tools:** Is there a common external tool or method, like creating a character sheet with multiple angles first, that the community has found indispensable for this task?
I am less concerned with artistic style consistency at this point and entirely focused on character consistency. A 10-15% variation might be acceptable, but my current results are more like 80% variation.
Any detailed guidance, step-by-step workflows, or warnings about pitfalls you've encountered would be immensely helpful. I want to map out the entire process before I begin.
~Heidi
You've correctly identified the core tension: the need for a stable seed versus the need for varied scene descriptions. The seed locks in the initial noise pattern, but significant prompt changes will override it, as the model interprets the new textual instructions as a higher-priority directive.
One workaround is to construct a very detailed, immutable character "spec" string that you append to every prompt verbatim. Treat it like a database row you're joining to every query. For example:
```
--character_spec: a woman with a square jaw, a narrow nose with a slight bump on the bridge, monolid eyes, and black hair in a precise chin-length bob with blunt bangs --
```
Then your prompts become: `[scene description], --character_spec--`. It's tedious, but it forces a persistent anchor. You'll still get drift over 20 images, but it will be markedly reduced. Have you considered breaking the task into batches and using image prompting with varying weights for the scene versus the character reference?
--perf