Preserve Facial Identity Across an AI Campaign: A Reference-Based Workflow

TL;DR: Facial identity holds across an AI campaign when the face is a saved reference that every shot reuses, not a description you retype. Pick clean source photos, save them once as a character, connect that same character to every scene, change one variable at a time, and review each batch against a fixed checklist before anything ships.
The short answer
To preserve facial identity across an AI campaign, use a reference-based workflow: store approved photos of the face as one reusable character, connect that character to every generation alongside the product, style and environment, and keep everything except the variable you are testing fixed. Text prompts alone drift because "30-year-old woman with short dark hair" describes thousands of people. A saved face reference describes one.
This guide walks through that workflow step by step, with the checks that catch drift before a client does. It is written for art directors, brand teams and freelancers running multi-shot campaigns, and it uses Rainfrog's node-based studio as the worked example, because there the face, the product and the look are separate building blocks you connect on a canvas.
Why faces drift in the first place
Image models do not remember the last image they made. Every generation starts again from the inputs it is given. If the only input is a prompt, the model fills in a plausible face each time, and plausible is not the same as the same.
Drift usually comes from three places:
- The face is described, not referenced. Words like "sharp jawline" or "freckles" leave too much room. Two generations can both match the words and still look like different people.
- Too many things change at once. A new pose, new lighting, new outfit and new camera angle in one step give the model room to reinvent the face along with everything else.
- The reference itself is weak. One low-resolution photo, a heavy filter or sunglasses gives the model little to hold on to.
A reference-based workflow fixes the first problem directly and makes the other two easy to control.
Step 1: Choose source photos you can stand behind
Everything downstream inherits the quality of the reference, so spend time here.
- Use photos of a real person only with their documented consent for AI use in this campaign, or use a face that you have the rights to (for example, a character your team created).
- Prefer sharp, evenly lit images where the whole face is visible: no sunglasses, hair across the eyes or heavy retouching.
- Include more than one angle if you have them: straight on and a three-quarter view cover most campaign framing.
- Keep the look neutral. A dramatic makeup look or colour grade in the reference tends to follow the face into every shot.
If the source photos disagree with each other (different hair, years apart, very different weight), the output will wander between them. Pick one consistent set.
Step 2: Save the face once as a reusable character
The single most useful habit is to stop re-uploading the face for each shot. Save it once and reuse it.
In Rainfrog this is the Character Library: you upload the face or model once and it becomes a Character you can drop into any project. On the canvas the Character node carries that face reference, so every image you connect it to draws on the same source rather than a fresh description. Products, styles and environments live in their own libraries the same way, which is what makes the rest of this workflow repeatable.

Whatever tool you use, the test is the same: can you point ten different shots at one stored identity, or are you rebuilding the face each time?
Step 3: Build one campaign workflow and keep it fixed
Treat the campaign like a set with a fixed cast. Build the workflow once, then reuse it for every shot.
- Connect the Character for the face.
- Connect the Product being shown, so the item stays accurate too.
- Connect a Style for the visual language and an Environment for the setting.
- Add Camera and Text inputs only where the brief needs them, and add an Image reference if a specific pose or composition matters.
- Send everything into the Image Generator, which in Rainfrog combines character, products, style, environment and references into one campaign image.
Because the inputs are separate nodes, you can see exactly what each image was made from. When something drifts, you check the inputs instead of guessing which part of a long prompt caused it. If you would rather start from a preset, the workflows page and the in-studio templates such as Fashion Shooting and UGC Content already wire a character, product and style together.

Step 4: Change one variable at a time
Once the base workflow produces a face you approve, vary the campaign along one axis per batch:
- Scene first: swap the Environment while keeping character, product and style fixed.
- Then product: swap the Product to cover the range, same face, same look.
- Then framing: adjust Camera or add a pose reference for close-ups versus full-body shots.
If the face changes, you know which single input caused it. That is far faster than rerunning a prompt that changed five things at once.
Step 5: Make a character sheet before the full campaign
Before you produce dozens of finals, generate a reference set that shows the face and body from several angles. It becomes the visual bible for the campaign and the thing reviewers compare against.
Rainfrog has two nodes for this: the Character Sheet Generator builds a film-style character sheet from a character plus products, and the Full Body Generator shows the character wearing the products in front, side and back views. Approve the sheet with the client or brand lead first, then generate the campaign against the same Character.
Step 6: Review every batch against a fixed checklist
Do not rely on "it looks about right". Compare each new image to the approved reference and check the same things every time:
- Face shape: jaw, cheekbones and chin match the reference.
- Eyes and brows: spacing, shape and colour are unchanged.
- Nose and mouth: proportions match, especially in profile shots.
- Skin and marks: freckles, moles and skin tone are present and in the same place.
- Hair: hairline, colour and length stay on brief unless the brief changes them.
- Age: the person has not visibly aged or de-aged between shots.
View images side by side at the same size, and look at small thumbnails as well as full size: drift that hides at full size is often obvious in a grid. Reject and regenerate rather than retouching a different face into shape.
How other tools describe face references
Rainfrog is not the only tool built around reusable references. Two examples show how other vendors describe the same idea. We have not run a side-by-side test of these tools for this article: the screenshots show each vendor's own documentation, not our results.
Midjourney documents a feature called Omni Reference for putting a person or object into your images. Its documentation page currently notes that Omni Reference is not supported in V8.2 and points V8.x users to Edit Model instead, which is a good reminder to check which model version a reference feature applies to before you plan a campaign around it.

Leonardo.Ai announced Character Reference on 30 May 2024, describing it as a way to keep a character consistent across different locations, poses and styles.

The difference to look for is not whether a tool accepts a face reference, but whether that reference is stored once and combined with your products, style and setting in a workflow you can rerun. For a broader comparison, see the brand consistency roundup linked below.
Where reference-based workflows still struggle
A stored reference makes consistency much more likely, but it is not a guarantee. Be honest with clients about the edge cases:
- Extreme angles (looking straight down, full profile, back of the head) have less of the reference to work from and drift more often.
- Small faces in wide shots carry fewer details, so identity is harder to hold. Plan close-ups where the face matters.
- Heavy stylisation can overpower the reference. A strong illustrated or painterly style may trade likeness for look.
- Group shots with several characters raise the chance that features blend between people.
- Consent and rights do not come from the software. You still need permission to use a real person's likeness and you still need to follow the platform's terms.
How this fits a brand consistency system
Facial identity is one part of campaign consistency. The same reference-first approach keeps products, colour and setting steady too. For a wider comparison of tools on that front, see 7 AI image generators that actually maintain brand consistency. For the asset-library side of the method, see how to create consistent brand visuals with AI without writing a single prompt.
FAQ
Which AI image generation tools preserve facial identity?
The tools that preserve facial identity best are the ones that let you save a face as a reusable reference and attach it to every generation, rather than describing it in a prompt. When you evaluate a tool, check whether it stores a character you can reuse across projects, whether you can combine that character with product and style references, and whether you can see which inputs made each image. Rainfrog is built around this approach with its Character Library and node-based workflows.
How many reference photos do I need?
Enough to show the face clearly and consistently. A small set of sharp, evenly lit photos from a front and a three-quarter angle is a better starting point than many mismatched ones. Check your tool's guidance for its specific requirements.
Can I keep the same face across both images and video?
Yes, if the same stored character feeds both. In Rainfrog the character is an input to the canvas, and image outputs can be passed on to the video nodes, so start from approved stills of the face rather than a new description.
Is a text prompt ever enough for a consistent face?
For one-off images it can be close enough. For a campaign with many shots, prompts alone tend to produce similar-looking but different people, which is exactly what a reviewer notices.
Do I need permission to use a real person's face?
Yes. Get documented consent from the person for AI-generated use in the campaign, keep it on file and follow the terms of the tool you use.
Start with one face and one workflow
The fastest way to see whether a reference-based workflow holds your face steady is to try it with your own approved photos. Save one character, connect it to one product and one style, and generate a small batch against the checklist above.
Sign up for free to explore the Rainfrog studio and build your first character workflow. When you are ready to run full campaigns, compare plans for credits and storage.
