AI Influencer Master Class - Face Creator: How To Create A Unique Face For Your Character LoRA (Patreon)

🎥 Watch this workflow in action: timestamped to where I use Face Creator in the full video. If you want a visual walkthrough alongside this breakdown, that link jumps straight to it.When you're training a character LoRA, the goal is a model that can generate the same identifiable person across dozens of different poses, outfits, and angles. That means every image in your training dataset needs to trace back to one consistent, locked-down identity. Face Creator is the workflow I run first, before I generate a single dataset image, because it produces that one anchor photo everything else in the pipeline gets built from.

Here's the full sequence for context: Face Creator locks the identity, Dataset Generator uses that locked identity to generate a full pose-varied dataset, and Checkpoint Tester compares the resulting trained checkpoints to find the best one. This post covers step one.

What the workflow does

You feed it one reference photo of the face you want to train on. From there, an image analysis model called Gemma4 (running as gemma4_e4b_it_fp8_scaled.safetensors) looks at that photo and writes out a forensic-level text description of the face: ethnicity, skin tone, face shape, eye color and shape, eyebrows, nose, lips, jawline, cheekbones, hair color and texture, any distinguishing marks. All stated as explicit fact.

That description becomes the generation prompt. The system prompt driving Gemma4 is specific about the goal: "You are a forensic portrait photography prompt writer. Your #1 priority is exact identity fidelity to the reference photo, never soften." That instruction exists because image models have a natural pull toward beautifying or averaging a face into something more generic looking. Forcing an explicit, detailed description first is what keeps the model locked to your actual reference instead of drifting toward "generic attractive person."

Why the output is always a straight-on portrait

Regardless of what angle your original reference photo was shot at, side angle, tilted, looking away, whatever, Face Creator always outputs a straight-on, front-facing portrait. This is deliberate, and it's enforced two ways. First, the Gemma4 system prompt itself states that the reference photo should be used "only as a source of facial identity, not composition," and that the output pose "must still always be straight-on frontal regardless of the reference photo's angle."

Second, there's a dedicated Negative Prompt node running alongside the generation that explicitly excludes side profile, profile view, three-quarter angle, turned head, tilted head, angled shoulders, looking away, and looking off-camera.

You want this flat, neutral, forward-facing portrait because it becomes the single reference image every pose in your future fifty-image dataset gets generated from. Any tilt or angle baked into that base photo just becomes noise the rest of the pipeline has to fight through later.

How the image actually gets generated

The KSampler in this workflow runs at denoise 0.45. That number matters because it tells you what kind of generation this actually is. At 0.45 denoise, this isn't a fresh image being generated from nothing using only the text prompt, it's your actual uploaded photo (resized and padded to 1280x1280 first) being partially transformed using the Gemma4-written identity prompt as guidance.

Structure and general likeness from your original photo carry through the pass, while the guided prompt pushes the result toward that clean, straight-on, unsoftened output. That's the mechanism behind why the result still looks recognizably like your input photo rather than a new face that just sort of resembles it.

Why this step matters for the LoRA you end up training

A character LoRA doesn't learn "a face," it learns one specific face across every pose, angle, and lighting setup you feed it. If the base identity photo you start from has any ambiguity in it, that ambiguity doesn't average out during training, it gets amplified. You end up with a checkpoint that looks right in some generations and subtly wrong in others, and the actual cause traces back to before you ever touched the dataset generator.

For example, my Alyssa character's entire fifty-image dataset was built off a single Face Creator output, straight-on, locked, unsoftened. Every pose in that dataset referenced that one photo for identity. That's a direct part of why the trained checkpoint holds her identity consistently across completely different framing and lighting in the final generations.

If you're building a character LoRA and thinking about skipping straight to dataset generation with whatever photo you happen to have, don't. You can't fix identity drift later with better captions or a higher training rank. You fix it by locking the face down first, before a single dataset image gets generated.