Before Consistent Characters

This article is observational and reflective in nature. Statements about generative model behavior are based on operational practice rather than controlled laboratory measurement.

The Early Days

In the early days of generative image AI, creators often talked about generating “the same character.”

On forums like Reddit, people experimented with many techniques:

Reusing long, highly detailed prompts
Giving characters specific names
Keeping generation within the same chat session
Sharing seed values

But even with the same prompt, the result was rarely the same person.

The face changed. The proportions changed. Sometimes the entire character changed.

Creators gradually discovered something frustrating:

Prompts describe appearance — they do not preserve identity.

Three Early Community Themes

Looking back, three recurring themes defined the early community experience with generative character creation.

The Miracle Shot Every creator eventually produced one perfect image — the right face, the right presence. The challenge was that it could not be reliably reproduced. The community called this the “miracle shot.”

Character Consistency The deeper goal was never a single image. Creators wanted the same character to appear across multiple poses, scenes, and contexts — like an actor performing different roles. This became the central unsolved problem of the era.

Prompt Fixation As creators tried to reproduce successful results, prompts grew longer and more detailed. This often made things worse. The model would lock onto certain visual patterns or regress toward defaults, a behavior now understood as fixation drift and attribute drift.

The “Miracle Shot”

Most creators eventually experienced the same moment.

You generate an image.

Everything is perfect — the face, the proportions, the presence.

You save the image.

The next day, you try the same prompt again.

The character is gone.

The community often described this as “losing the character.” In practice, it was the loss of identity continuity.

What Artists Actually Wanted

The real goal was never just to generate a good image.

Artists wanted something else.

They wanted a character who could act.

The same character appearing in different poses. Different gestures. Different emotions.

Like an actor performing across multiple scenes.

But generative models do not preserve identity by default. Each generation is a new reconstruction.

So the moment a pose changed, the character often changed as well.

Example 1 Example 2

Images from the author’s own library. Character originally created in ChatGPT and migrated to Gemini. Without identity stability, the same character may appear differently across generations.

An Unusual Workflow

My own workflow at the time was somewhat unusual and probably not representative of typical community practice.

While many creators tried to stabilize characters through increasingly detailed prompts, I moved in the opposite direction.

Once a generation produced the “right” character, I stopped trying to recreate it with text.

Instead, I treated that image itself as the character.

From that point on, the goal was no longer to regenerate the character, but to let that same character appear in different poses, gestures, and emotions.

Looking back, this was essentially the first form of what would later become the anchor concept.

Pose 1 Pose 2 Pose 3 Pose 4

Example: identity continuity across scenes. Images from the author’s own library, provided as a clear illustrative example of the concept.

Early users often aimed to reproduce a successful “miracle shot”. In this workflow, that reproduction becomes the starting point for generating further scenes with the same character.

Community Solutions

Over time, the community explored several technical solutions to the same problem.

Techniques such as LoRA training, DreamBooth, seed control, and reference-image features (like Midjourney’s --cref) all attempted to stabilize identity.

These approaches generally worked by either:

modifying the model itself, or
injecting additional signals into the generation process

While effective in many cases, they often required model-level intervention or platform-specific features.

These solutions improved identity stability, but they were often tied to specific tools, training pipelines, or platform features.

However, these methods addressed identity stability through model control or feature-specific mechanisms, rather than through a general operational workflow.

From “Same Character” to Identity Governance

A different possibility gradually became clear.

Instead of trying to recreate the character through prompts, what if the correct image itself became the reference?

The moment the “right character” appears, that image becomes the anchor.

Not a prompt. Not a description.

An image.

This idea eventually led to the Character Identity Protocol (CIP).

CIP does not attempt to control the internal workings of generative models.

Instead, it introduces an operational layer that governs identity continuity during generation.

This layer relies on:

Anchor images
Minimal prompts
Validation gates
Hard abort conditions

Generative models are probabilistic systems. Each generation is a reconstruction, not a retrieval.

Without explicit control mechanisms, identity does not persist automatically. Even small changes in pose, lighting, or composition can cause identity drift.

CIP therefore treats identity not as a stylistic detail, but as a governance constraint applied to the generation process.

The Core Principle

In probabilistic generative systems, identity does not persist by accident.

It must be governed.