How to Create Consistent AI Characters (2026 Workflow)

Anchor one image, describe it the same way every time, feed it back as a reference, and fix instead of re-rolling. Here's the full loop, and the advice to stop following.

The nose moved. Again.

You've got your character. Same prompt, same words, same careful line about the scar over the left brow, and the model hands you a cousin. Close enough to notice, wrong enough to bin. Do that eight times for a comic page or a spokesperson ad and you haven't made a series, you've made a family reunion.

Here's our take: consistent AI characters stopped being a training problem in 2026 and became a reference-and-edit problem. Nobody in the ICP needs to fine-tune a model anymore. You need four moves, in order:

1. Make one anchor image you'd actually sign off on.
2. Write a fixed identity spec and paste it in unchanged, every time.
3. Lead with references, not adjectives.
4. QA each output against the anchor before you generate the next one.

Everything else in this article is either how to run those four properly, where they break, or which popular tip to stop repeating. There's one of those we'll retire in about 900 words, and it's probably in your workflow right now.

What consistent AI characters actually means

Two different things drift, and confusing them wastes days.

Identity is the person: facial geometry, build, skin tone, the signature jacket, the earring that keeps vanishing. Style is the look: watercolour, noir, die-cut sticker, that one plasticky 3D thing everyone did in 2024. They move independently. You can hold the same hero through a watercolour panel and a noir panel, or hold one flat-illustration style across a cast of twelve. Consistent AI characters means locking the first axis while the second, plus pose, lighting and scene, is free to change.

The research world splits it further. Identity preservation means real human faces, scored with face-embedding similarity. Subject-driven generation is the broader framing: condition a model on a few reference images, then render that subject somewhere new. Nataniel Ruiz and colleagues at Google set the pattern with DreamBooth in 2022, fine-tuning on three to five images of one subject.

Get the axis right before you start fixing anything. If the face is fine and the medium keeps sliding, you don't have a character problem, you have a style problem.

Why characters drift, in thirty seconds

Every image is generated as an independent event. The model starts from noise or tokens with no memory of the picture it made for you ninety seconds ago, so nothing carries over unless you carry it.

Then there's the prompt itself. "Green eyes, dark hair, mid-thirties" describes a few million people. The model fills every gap you left, and it fills them differently each run, because that's the job. Under-specification plus zero memory equals a new person each time.

Which tells you what the fix has to be. Not better adjectives. External conditioning: pixels the model can look at, or an edit operation that starts from pixels you already approved.

Three approaches, and only one is the default

Every up-to-date vendor doc now leads with reference images. Google, OpenAI, Midjourney, Runway, Ideogram and Black Forest Labs all put reference conditioning first and treat fine-tuning as the specialist tier. That's a real reversal from 2023, when the answer to "how do I keep my character" was "train a LoRA".

The training-free middle ground is worth knowing about even if you never touch it. InstantID, from Qixun Wang and co-authors in January 2024, gets a usable likeness from a single face photo using an ID embedding plus a facial-landmark ControlNet, no training at all. ByteDance's PuLID does something similar with contrastive alignment, specifically to stop the identity injection from bleeding into the background and lighting, which is the known failure of the whole encoder family.

One clarification that saves arguments: ControlNet is not an identity method. It constrains structure, meaning pose, edges and depth. Pair it with an identity method or it just gives you a stranger in the right posture.

For JammyJar's readers the call is easy. Reference plus edit gets you to good in an afternoon. LoRA is for the people who need sprite sheet number four hundred.

The workflow: anchor, spec, reference, QA

Step 1: Build an anchor image you'd sign off on

Everything downstream gets checked against this one file, so treat it like a passport photo: neutral, well lit, unambiguous. Front-facing or slight three-quarter, clean background, whole face visible, no dramatic shadow eating half the jaw.

Give the character one feature a machine can hold onto. An asymmetric scar, a specific jacket cut, one blue streak in the fringe. Vague pretty faces regress to other vague pretty faces.

Generate ten. Keep one. This is the cheapest quality decision in the entire process.

Step 2: Write a fixed identity spec, then stop writing it

A short block of immutable text: face, hair, build, signature wardrobe, defining accessory. Paste it verbatim into every prompt, forever.

The reason it works is dull and important. Re-describing your character in fresh words each time hands the model a fresh interpretation each time. "Mid-thirties woman, dark bob" on Monday and "dark-haired woman in her thirties" on Tuesday are two different briefs.

Keep it tight and concrete. "Emerald eyes, asymmetric scar over the left brow, olive canvas jacket" beats a paragraph of mood.

Step 3: Lead with references, not adjectives

This is where the current model limits matter, so here are the numbers as of August 2026.

Google's Gemini image models take up to 14 reference images in one request, with high-fidelity sub-limits of six object references, five human images for character work and three style references. Google's own docs add the caveat every vendor blog leaves out: character consistency "is not always perfect between input images and generated output images". OpenAI's GPT Image 2 handles multiple references by index, so you can say image one is the character and image two is the location, and OpenAI's cookbook demonstrates holding a face and pose fixed while the scene changes around it.

Midjourney replaced the V6-era --cref with Omni Reference in V7, and it's a different beast: one reference image only, a weight (--ow) from 1 to 1,000 with a default of 100, an official warning to stay under 400 or results turn unpredictable, and double the GPU time of a standard image. Ideogram went the other way, launching Ideogram Character on 29 July 2025 as the first character consistency model built to work from a single reference photo. FLUX.2 accepts up to ten references with a sweet spot of four to six. Runway Gen-4 References takes three, labelled in order.

The rule underneath all those numbers: fewer, cleaner references beat more. Conflicting references measurably increase drift, so if your character is wobbling with eight references loaded, the fix is usually to delete five of them.

Step 4: QA before you generate the next one

Drift compounds quietly. Check every output against the anchor, in this order, because these are the things that go first: earrings, scars, piercings and tattoos (accessories erode before faces do), mirrored features and pose flips, hair length, wardrobe colour, then age and skin tone. Last, the sneaky one: a face sliding toward generic and attractive.

Reject and re-anchor. Never build the next scene on an output you're already squinting at.

That loop, anchor and spec and reference and check, is the whole method. It's also what JammyJar's reference-image and Theme workflow is built around: the character rides on the references, the look rides on a saved Theme, and neither has to be retyped.

Two characters in one frame is where it breaks

Put a second person in the scene and the models start mixing them. The peer-reviewed name for it is identity blending, with attribute leakage as the sibling problem: features cross over because every token competes in one shared attention space. Jang and co-authors documented it in 2024 with MuDI, showing that current methods often fail on multiple subjects at once and returning mixed identities, then proposing segmentation-based decoupling as the fix.

A 2026 stress-test benchmark put a number on the cliff. Models look fine at two to four subjects and hit what the authors call catastrophic identity collapse at six to ten, or under heavy occlusion. Their Subject Collapse Rate metric is grounded in DINOv2 for a pointed reason: CLIP-based scores were rewarding identity-collapsed images full of generic clones.

Think of two paint colours in one glass of water. There is no prompt that un-swirls them.

So don't ask for the crowd. Generate each character separately against the same background brief and composite them, or build the scene sequentially: make character A, expand the frame, insert character B by editing. Slower, and it works, which is more than the one-shot group shot manages.

When it drifts anyway, edit instead of re-rolling

Re-rolling asks the model to reinvent your character. Editing asks it to move the character you already have. For non-technical users that difference is the single biggest fidelity gain available, because identity carries structurally through the pixels rather than being re-derived from text.

Black Forest Labs built FLUX.1 Kontext around exactly this, reporting stronger character consistency across iterations and naming character drift during multi-turn edits as the problem to beat. Their KontextBench is unusually concrete for a vendor benchmark: 1,026 image-prompt pairs over 108 base images, 193 of them character references. Google positions Gemini 3 Pro Image, the Nano Banana Pro line, as the premium editor for targeted changes, with Gemini 3.1 Flash Image as the multi-reference workhorse.

The recovery playbook, in order of what to try first:

• Re-anchor. Go back to the original anchor image rather than the last output. Drift is cumulative.
• Cut references. Drop to the three cleanest. Conflict reads as licence to improvise.
• Inpaint the specific failure. Mask the earring, the scar, the wrong sleeve. Regenerate that region only.
• Watch for edit rot. Long edit chains accumulate artifacts and slow quality loss, so re-anchor every few turns even when nothing looks broken yet.

Edit, don't re-roll. Tape it to the monitor.

How consistent can it really get?

Not 99%. Nobody who tells you 99% has published a method.

Real-world reporting across current tools triangulates to roughly seven or eight generations in ten reading as recognisably the same person, with two or three showing drift you'd notice: eye spacing off, a nose reshaped, the scar wandering to the other brow. Good enough for a comic page or a spokesperson series with light curation. Not good enough for fifty game sprites unsupervised. That figure comes with its weakness attached: it's triangulated from vendor field reports, not a controlled study, because no independent apples-to-apples 2026 benchmark across the flagship models exists publicly.

The metrics for building one do exist, and they came out of DreamBooth. DINO similarity, on self-supervised vision embeddings, is the preferred subject-fidelity score because it separates individual instances instead of lumping same-class objects together. CLIP-I flatters, because it rewards mere same-class similarity. CLIP-T checks you're still obeying the prompt, which stops you "winning" consistency by ignoring the scene. For photoreal faces, ArcFace cosine similarity is the standard.

They all have blind spots, and researchers say so. The DreamBench++ team, Peng and colleagues from Tsinghua, UIUC and StepFun, argued in 2024 that automatic scores like DINO frequently misalign with human judgement, and proposed GPT-4o as an evaluator with human alignment checks. Report metrics jointly, add human raters, or don't claim a number.

That's what our own test is built on: one character with a testable feature, eight fixed scene prompts from studio portrait to night exterior, ArcFace against the anchor as the primary score, DINO on the full crop for build and wardrobe, CLIP-T for prompt fidelity, plus a one-to-five "same person?" rating from three reviewers. Mean and worst case, per model, prompt set published. We haven't finished running it, so there are no numbers here yet. Be suspicious of anyone who publishes theirs without the prompt set.

The stat you should stop quoting

While we're here: the line about character consistency being the number one barrier to professional adoption, "per a 2025 Adobe survey of 1,200 creative professionals", is fabricated. It traces to a single vendor SEO blog and matches no Adobe study. Adobe's real 2025 research is the Creators' Toolkit Report, run with The Harris Poll and published on 28 October 2025, which surveyed over 16,000 emerging and semi-professional creators across eight countries. It found 86% of them using creative generative AI and 76% saying it grew their business or personal brand. It says nothing about character consistency being anyone's top barrier.

A "92% of audiences" claim doing the rounds in the same niche has the same problem. Check the primary source before you build a slide on it.

Which model to reach for

Checked August 2026, and expect this to move by Christmas.

Pick by job, not by loyalty. Character work usually means two models: one that holds references while you build the set, one that edits cleanly when a single element goes wrong. If you'd rather not pay two subscriptions to find that out, JammyJar puts the Gemini image models and GPT Image 2 behind one prompt box with your references and Themes saved alongside them.

A quick word on using real people

If your character is based on someone who exists, the rules are moving fast. The NO FAKES Act, which would create a federal right against unauthorised AI replicas of a person's voice or visual likeness, cleared the Senate Judiciary Committee by unanimous voice vote on 18 June 2026 and is waiting on a floor vote. It is not law yet. Tennessee's ELVIS Act has been in force since 1 July 2024. The DEFIANCE Act passed the Senate on 13 January 2026 and hasn't been enacted.

Platform policies already restrict more than the statutes do, so the practical advice is short: use your own reference photos, or a clearly fictional character, and read the real-person policy of whichever model you're using before a commercial spokesperson goes out.

FAQ: consistent AI characters

Do seeds keep a character consistent?
No. A seed reproduces a whole composition only when the prompt and parameters stay identical. Change the pose, the scene or the lighting and the identity drifts anyway. Runway's own documentation notes seeds don't transfer reliably between models or versions either. Reuse the seed if you want the same picture again; use a reference image if you want the same person.

How many reference images should I use?
Three to six clean ones, in most tools. Google allows up to five human images for character work inside a 14-image request, and FLUX.2's effective range is four to six despite a ten-image ceiling. Conflicting references increase drift instead of reducing it, so if a character starts wobbling, delete references before you add any.

Can I keep two characters consistent in the same image?
Usually not in one shot. Models blend identities and leak attributes between subjects, and a 2026 benchmark found collapse setting in around six to ten subjects. Generate each character separately and composite, or add them sequentially by editing: make character A, expand the frame, then insert character B.

Is LoRA still worth training in 2026?
Only for volume. If you need hundreds of near-identical assets, game sprites or a full catalogue, trained weights still give the tightest consistency. For a comic page, an ad series or a mascot, reference-and-edit gets you to a usable result in an afternoon without the ten to twenty training images and the GPU time.

Which AI model is the most consistent for characters?
There's no clean single answer, because no independent benchmark across the 2026 flagship models has been published. Vendor claims of 95 to 99% consistency come with no sample size, metric or prompt set. Google's Gemini line handles multiple references well, FLUX.1 Kontext leads on multi-turn editing, and Ideogram Character works from one photo.

The nose will still move occasionally. Seven or eight in ten, on a good day with clean references, and the other two go in the bin. That's the actual state of consistent AI characters in 2026: much better, not solved, and mostly a discipline problem rather than a model problem.

So build the anchor first. Ten generations, keep one, write the identity spec underneath it, and the next scene starts from something you already approved instead of a description you're hoping the model reads the same way twice.