How many times have you created the perfect AI character, only to lose them forever on the next render? Or created a character Lora for model X, and a few weeks later finding yourself having to created the same lora for model Y? How many times have you cursed the underlying architecture of a model, because it won’t accept a Lora made for another model?
I’ve done all this, more times than I’d like to even try to count.
The character you see here was created without a LoRA. No IP-Adapter. No ControlNet. And no ReActor or face-swapping.
The Idea
Until now our go-to tools for achieving character consistency has mainly been character specific trained Loras and face-swapping tools such as FaceID and ReActor.
While a well trained Lora can give you incredible results, there are som clear and obvious downsides. One being that there’s new models released every other week, and Loras that works with the current model, will almost certainly not work with the model that’s released tomorrow. To cover all models you might be using, you might need 3-4-5 versions of the very same Lora, and since they will be trained on different models the character consistency will differ a bit.
FaceID and ReActor, while being well tried and tested and often give good results, they too have their downside. Everything from various artefacts in the images, problems with images of different sizes and aspect ratio, and having to either integrate them directly in the workflow or use them in a separate workflow.
These frustrations are exactly what led me to rethink the entire problem. The goal isn’t just to get a consistent face; it’s to get a consistent character—a digital actor who can exist across different scenes, models, and styles. Face-swapping is a post-production fix, like digital makeup. I wanted to get the performance right in-camera.
So, what if we stopped trying to fix the image after the fact, and instead built a system that directs the AI from the very beginning?
This is the core of the new idea: orchestration over brute force. Instead of relying on a single, powerful tool, I designed a modular workflow that treats the generative AI not as a photocopier, but as a multi-talented creative team.
I call it my AI Film Crew.
This system breaks down the complex task of creating a scene into distinct, manageable roles. Each role is handled by a specialized part of the workflow, and their combined efforts are orchestrated to produce the final, coherent image. There’s a Casting Director to define the actor, a Scriptwriter to bring them to life, a Cinematographer to set the shot, and a final Auteur Director to weave it all together.
Let’s meet the crew and see how they work.
The Character Vault: Creating the Actor’s DNA
Everything starts here, in what I call the Character Vault. This is our ‘casting office’ where we define a character’s core identity—their permanent, unchangeable ‘Digital DNA’.
The process is straightforward:
- Define the DNA: In the main text box, you write a detailed description of the character’s foundational features: their ethnicity, face structure, eye color, body type, and so on. This is the ‘source of truth’ for who they are.
- Build in Customization: Crucially, you also add placeholders like {clothing}, {hairstyle}, and {accessories}. These act as empty slots that our ‘Wardrobe Department’ can fill in later, depending on the scene’s requirements.
- Name and Save: You give your character a unique name (like “Zara”) and toggle the save_to_db switch.
This is a one-time setup for each character, performed in a separate ‘blueprint’ workflow. Once you run it, the character’s profile is saved to a central database. They are now officially on your roster of digital actors, ready for their first role.
The Character Select: Wardrobe & Makeup
With our actor chosen from the Vault, it’s time to get them ready for the scene. This is where the Character Select node comes in, acting as our digital Dressing Room and Makeup Department.
Its function is simple but powerful:
- Select the Actor: Using the dropdown menu, you choose which character from your database you want to work with. This instantly loads their core DNA.
- Customize the Look: In the text fields below, you define their temporary appearance for this specific shot—their clothing, hairstyle, and accessories. This fills in the {placeholders} we created in the Vault.
The output of the Character Select node is not a final, ready-to-use prompt. Instead, it’s a highly structured creative brief.
This brief is specifically designed to be understood by its partner, the Master Key node.
Think of the Master Key as an AI Scriptwriter. It reads the brief (our character’s DNA + wardrobe) and then, using its own complex internal instructions, writes a beautiful, descriptive paragraph that brings them to life. And to do this “thinking,” the Master Key itself relies on the power of the Gemini API node.
These three nodes form a powerful, interconnected team. The Character Select prepares the actor, the Master Key writes their script, and the Gemini API provides the creative spark. Each specialist is designed to work in perfect harmony with the next, creating a result that would be impossible for any single node to achieve on its own.
The Auteur Director: Weaving the Final Scene
We now have all the essential elements for our scene: our fully customized character from the Master Key, the camera instructions from our Cinematographer, the narrative beat from our Director, and the location from our Set Designer.
The final, crucial step is to hand this collection of notes to the big boss: the Auteur Director. This is the AI who will synthesize everything into a single, masterful, final prompt.
The tool for this is the Summary Writer node.
At first glance, it looks simple, but its role is the most sophisticated in the entire pipeline.
- The Director’s Manifesto (prompt_file): This is the Auteur’s vision. We load a text file (in this case, Scene_Director.txt) which contains a powerful, detailed system prompt. This prompt is a set of iron-clad rules and artistic principles that the final AI must follow. It tells the AI how to prioritize the character, how to harmonize the lighting with the mood, and to always respect the specified shot composition.
- The Day’s Notes (text_to_combine): All the outputs from our other crew members—the character description, the shot type, the action—are fed into this input.
- The Final Shooting Script (combined_text): The node then combines the Director’s Manifesto with the day’s notes, creating one final, comprehensive brief. This is the ultimate instruction that gets sent to a second Gemini API node, which then writes the final, beautiful prompt we use to generate the image.
The Scene Director has a strict set of rules, for example:
You are a "Master Scene Director," an AI master of visual storytelling and cinematic composition. Your purpose is to take four distinct inputs—a desired **shot composition**, a specific **character action**, a description of a **lead character**, and a set of **thematic keywords** for the scene—and weave them into a single, powerful, and artistically coherent prompt.And with directions including:
Prime Directive: The Shot Composition is Law
- Action is Character
- Intelligent Casting – Handling “Friends” and “Crowds
- Harmonize the Elements
- Respect Solitude
This final step is the secret to the entire system’s success. It ensures that our character, action, and setting aren’t just crudely stitched together, but are artistically blended according to a deliberate, pre-defined vision. It’s the difference between a random collage and a masterpiece of composition.
Piecing it All Together
What happens is that we have several pieces of text, system prompts, and rules that intelligently are pieced together using the Gemini API nodes.
Here’s where the script comes to life. Our Character Select node, now configured for the winter scene, generates the initial character brief, based of your character DNA, or input. The output prompt for the character, using my pre-made character called Zara, at the current settings will look like this.
The 3 red nodes, Shot Type, Action and World Weaver are combined with with the Character Description in the logic from the Scene Director and finally sent through a second Gemini API node which creates a high quality and coherent prompt from it all. And the final prompt in this case will look like this:
And this is the image that will be created at the end of this character assembly line:
The Director’s Q&A
Does this system really works with all models?
Most likely no. It does work with Qwen, Flux, WAN TXT2VID. I have not tested all models, but I believe that the system requires the T5-XXL CLIP to really grasp and understand the prompts.
Why isn’t the skin looking hyperrealistic?
Because that was never the purpose for this system. The issue with realistic looking skin is an old problem that already have a plethora of solutions, and you should be able to use your own favourite tool for that. Also, this system should be working with charcters that are not meant to look realistic, such as cartoons and anime characters. But more testing might be needed for that.
Does this system work together with other LoRa?
I have used several different LoRa in the system, and have not noticed any negative effects. What is important to note though is that changing or removing a LoRa, even one that has absolutely nothing to do with the looks of the character, tends to change the looks of the character to some degree. If you created a solid character while using a specific LoRa, try to always use that LoRa together with your character for the best consistency.
Does this work with other LLM’s other than Gemini?
Most likely not. I’ve ever only used Gemini through API, and the system is pretty much built around the use of Gemini, as the system prompts and helper nodes are custom made for that purpose. One of the great things with using Gemini API is that Google is providing a free API key, with a generous amonut of free API calls per day. You can obtain your free API key here: Google API Key
The Takeaway: From Prompter to Producer
From a simple database entry to a fully realized character in a dynamic scene, this workflow represents a fundamental shift in how we can approach AI art. It’s about moving beyond the endless cycle of re-rolling and prompt-wrangling to a more deliberate and creative process.
You are no longer just a user prompting an algorithm; you are an Executive Producer assembling a team of specialists. You are a director with a vision, guiding a digital actor to give the perfect performance, shot after shot.
This isn’t a replacement for tools like LoRAs or ReActor, but a different paradigm altogether—one focused on narrative, control, and world-building from the ground up. It’s a glimpse into the future of creative AI, where the only limit is the story you want to tell.
Now, go build your cast.
You can find sample workflows at my Patreon, free to download for both free and paying members.
If you liked this, consider signing up for my newsletter and get news, tips & tricks like this directly in your inbox.