Consistent character images
When to use this skill
Load this skill when a user wants generated images of a real, specific person and the likeness matters. Typical requests are a website headshot, an avatar, a set of portraits, a photo gallery, or a character reference sheet.
The problem it solves is likeness drift. Every image model pulls a face towards its own average. The hairline drops, the beard fills in, the jaw widens, the shoulders broaden. Each render is plausible alone. Next to the real photographs it is a different person, and the user sees that at once.
Do not use this skill for a fictional character, for stock imagery, or for a person the user has no photographs of.
Before profiling any photographs, confirm the subject is the user themself or that the user holds documented consent to generate and publish that person's likeness. Refuse a third party's likeness — a public figure, a coworker, anyone else's photographs — absent that consent; a convincing identity lock makes this skill as good at impersonation as it is at a faithful headshot.
The reference implementation is tools/character-studio in the pooriaarab
repository. Read its README before you run it.
The workflow
Follow this order. Step 5 is the one agents skip, and skipping it wastes the whole run.
Search for prior art first. Grep the available skills for the person's name, and for the words identity, likeness, character and portrait. A skill named
pooria-photoalready held an identity reference for this exact person. It carried the line "Avoid a narrow V-shaped chin" and a failure list entry, "The jaw becomes too narrow or too square". A fresh identity lock was built from scratch without that check. The subject then reported that exact fault, which cost several render cycles.Collect many references, not six. Ask for fifty or more, across angles, lighting and grooming states. Six is enough to build a caricature and not enough to build a likeness. Collect that many to profile and to choose from. Pass only three to five of them to any single render. See the calibration section below.
Profile the photographs before you write anything. Ask a vision model for a structured profile of every photograph: beard state and cheek coverage, hair style and volume, hairline visibility, glasses, framing, angle, and the measurements below. Then take the median of each measurement across the whole set. Those medians are the identity lock. Do not write it from a handful of pictures and do not write it from a description the user gave you.
Write the anchors. Eight to twelve, ordered by how fast a wrong value reads as "not them". Each carries the measured median and its sample size. See the anchor rules below.
Generate one throwaway render. Expect it to be wrong. You now have a known-bad image, which is the test fixture for the next step.
Validate the verifier. Score that known-bad image with at least two vision models. Keep a model only when it fails the image and names the faults you can see yourself. Record the model, the score and the date.
Generate, verify, retry. Put the identity lock in the prompt before the scene. Score the render against the real photographs. Feed back one correction per round. Stop on a pass or after three attempts. Keep the best attempt either way.
Show the evidence. Put the renders, the anchor scores and the real photographs on one page. The user must be able to disagree with the verifier.
Fix anchors, not shots. A fault that appears across several shots is an anchor problem. Add the exact failure to that anchor's NEVER line.
How to write an anchor
An anchor holds one trait, a weight, a MUST line, a NEVER line and a reason.
Weight by speed of failure. Weight 3 means a wrong value instantly reads as a different person. Weight 2 weakens the likeness. Weight 1 is supporting detail. Weight 3 also arms the hard floor in the pass gate, so do not spend it on a trait you would forgive.
MUST describes the trait as a measurement. "Striking angular features" is praise and tells a model nothing. "Longer than it is wide, with a narrow jaw that tapers to a softly rounded chin" tells it where to put the lines. Put a number in wherever a number exists. A percentage of face height, a beard length in millimetres, a height in head units, a shoulder width in head-widths. A verifier can check a number. It cannot check an adjective.
NEVER names the specific failure the model keeps producing. A general NEVER line is wasted text. Look at the failed renders first, then write down what you actually saw. Write "a dense uniform full beard" because that is what the model produced, not "an unsuitable beard".
why records the render that failed. A later reader who does not know the
history will read a strict MUST line as fussy and soften it. The reason is what
stops that. Write one sentence naming the failure the anchor exists to prevent.
Keep a drift-direction list, not only a NEVER list
A NEVER line states what is wrong. A drift direction states what the model turns the trait into. A verifier can look for a named result. It cannot look for an absence. Keep the list beside the anchors and extend it from real renders.
The directions observed here:
- Round wire frames become rectangular.
- The chin becomes a pointed V.
- The beard fills in across the cheeks.
- Hair becomes a sculpted pompadour.
- Skin becomes retouched and poreless.
- The torso becomes broad.
- Every pose defaults to a notebook, a laptop, or a direct gaze into the lens.
The enumeration density ceiling
An enumerated grid of about 16 items drifts reliably. Roughly one item comes back duplicated or garbled. The more discrete things a prompt enumerates, the more likely one of them breaks.
An identity lock is an enumeration. Many anchors, plus a check line for each one, cross that ceiling quickly. Past the ceiling, expect about one anchor to break per render.
Prefer fewer correct elements over a busy approximation. Cut a check line before you add one.
Validate the verifier before you trust it
The verifier model is load-bearing. A weak verifier passes a wrong face, and the retry loop then has nothing to correct. You get a clean PASS on a stranger.
On 2026-09-12, google/gemini-3.8-flash scored a known-wrong render 97.6%
PASS. The render had a covered forehead, a dense black beard and a broad build.
The model invented measurements that agreed with the MUST text instead of
measuring the picture. gemini-3.1-pro-preview scored the same image 59% and
claude-opus-5 scored it 66%. Both set same_person to false. Both named the
three real faults.
Rule: never trust a verifier you have not failed a known-bad image with. Run
it against that image first. A verifier passes the test when it fails the image,
sets same_person to false, and names faults you can see yourself. Write the
result and the date next to the model name in the pack.
Three things in the verifier prompt make the difference. Keep all three.
- Two separate observation fields. The verifier writes what it sees in the REFERENCE and what it sees in the CANDIDATE, before it scores. Tell it that a candidate note which paraphrases the MUST line proves it did not look.
- Forced numeric estimates on both images. Forehead height as a percentage of face height. Face width divided by face height. Shoulder width in head-widths. Beard coverage on the flat of the cheek as a percentage. A gap above about 15 percent is a real difference.
- A calibration line. Tell the model that straight 5s almost always mean it did not examine the image. Most candidates score 2 or 3 on at least one critical anchor.
Verify at full resolution. Thumbnails hide garbled type. Full resolution shows it. A downscaled candidate here hid a hard-edged composite beard that was obvious at full size. Send the candidate to the verifier larger than the references.
Also keep the verifier blind. It sees the reference photographs, the candidate and the anchor definitions. It must never see the prompt that made the image, or it grades the intent instead of the result.
The pass gate
Score each anchor from 0 to 5. A 0 means the trait is not visible, so drop that anchor from the arithmetic. Take a weighted mean over the rest.
A render passes only when all three of these hold.
- The weighted mean reaches the pass score, typically 0.85.
- Every critical anchor scores at least 4 on its own.
- The verifier did not set
same_personto false.
The mean catches a render that is wrong in many small ways. The floor catches a render that is right everywhere except the one trait that carries the likeness. You need both.
Verify the brief, not only the likeness
A likeness score cannot see the shot. A render can be a flawless likeness of the right person in the wrong pose, wearing the wrong clothes, pulling the wrong face, and pass at 80%.
Send the camera line, the expression line and the wardrobe line to the verifier as a separate pass or fail, judged apart from the likeness score. Scoring the two together lets either one hide the other.
Put the brief failure at the head of the corrections. Only the first correction goes back each round, and a sharper beard on a shot pointing the wrong way is a wasted attempt.
Four faults this caught that nothing else did:
- A rear three-quarter delivered as a plain side profile. It scored 80% and sat next to the real side profile, an identical frame in a turnaround.
- A head-and-shoulders delivered as a full-length high-angle shot.
- A mirror flip: turned to his right instead of his left, watch on the wrong wrist, mouth closed when the brief asked for mid-sentence.
- Sleeves rolled in one panel and long and buttoned in the next.
Name these in the verifier's instructions. Telling it what the model actually does beats asking it to compare against an abstract brief.
Describe the frame, do not name an angle
A brief that asked for a body "turned 135 degrees away" came back at 180 degrees twice, then with the head twisted back to expose a face. Rewritten as what the camera sees — mainly his back, the outer edge of one shoulder, a sliver of cheek, and explicitly no eyes, nose or glasses front — it passed on the first attempt. Degrees are a number the model does not measure.
Never let the identity block contradict the shot
The traits that make a person recognisable and the traits a shot asks for are different lists, and they overlap. Habitual expression is the usual collision. An anchor reading "closed-mouth asymmetric half-smile" is correct for a portrait and flatly wrong for a panel asking for serious. With both in one prompt the model splits the difference, and every expression panel comes back as the same faint smile.
Drop the anchor the shot overrides, in the prompt and in the scoring, from one shared function. Dropping it only in the prompt is worse than not dropping it: the model then obeys the brief and the verifier marks the correct render down against the anchor it was told to ignore.
The same collision hides in the wardrobe. A wardrobe written once for a whole pack names trousers, a watch and shoes. On a head-only crop that contradicts the framing, and the model resolves the contradiction by pulling the camera back and rendering the whole body. Say which instruction wins:
The framing is the final word on what is in frame. Wardrobe items the crop cannot reach are simply out of shot. Do NOT widen the shot to include them.
Do not score a shot on what it cannot show
Dropping a "not visible" anchor from the mean is right, but it quietly shrinks the denominator. A rear view scored 100% here: seven face anchors came back not visible, and the mean was taken over the remaining six. A perfect score over six anchors out of thirteen is not a perfect score.
Give the shot a named scope and resolve the anchor list from it, so the denominator is decided by the shot rather than by whatever the verifier happened to see. Six of six is an honest number. Six of thirteen dressed up as perfect is not.
The same scope fixes the prompt. A head-and-shoulders render told to match a height it cannot show is carrying an instruction it can only fail.
Pick the best attempt by pass first, then score
attempts.reduce((a, b) => b.score > a.score ? b : a) is wrong twice.
A strict > hands a tie to the earliest attempt. Two attempts here both scored
1.0; the first was judged a different person and the second was not. The tie
went to the first, so the pipeline published a failed render and recorded the
panel as failed while a passing one sat beside it on disk.
Ranking on score alone also prefers a higher-scoring failure over a lower-scoring pass. A passing attempt beats a failing one whatever the scores say, and a tie goes to the later attempt, because that is the one that carried a correction.
Record what the check objected to
A gate that fails without saying why costs a full cycle every time. The report
here showed a critical failure named brief and nothing else, so the only way to
learn what the verifier disliked was to run the whole verification again by hand.
Write the note into the report and into the log at the moment the check fails. Instrument a check as you add it, not after it has cost you an hour.
The retry loop
Feed back one correction per round. Every fault at once makes the model trade one fault for another. The scores here bounced instead of climbing: 78%, then 80%, then 73%.
Rank the faults by weight times the points missing. Send back the single biggest win. Revise only the weakest dimension before the next generation. Change one variable per round when you debug.
Cap retries at about three. A prompt that cannot pass in three attempts usually has a tool problem or a framing problem, not a wording problem.
Reframe past a fault the model will not drop. Crop the offending region out of shot. Framing beats prompting for anything the model keeps producing.
Change the model before the twelfth retry
One panel failed twelve attempts across two models on a single fault. The model that held the likeness best would not hold a crop or a facial asymmetry. The fix was not better wording, it was a model measured to follow an instruction more literally. A bake-off's real value is not picking one model for everything; it is having a measured second choice for the shot the default cannot do.
Relabel rather than ship a permanently failing panel
No model tried would render a controlled facial asymmetry. A panel asking for a lopsided half-smile would have failed forever. Ask for what the models can hold, rename the panel to match what it shows, and record the limitation on the panel. Leaving it failing teaches nobody anything, and calling a symmetrical smile a half-smile makes the sheet lie.
Never describe the face in the image prompt
A long prose description of a face makes the model generate a face from the words and ignore the reference photographs. Four photographs of a Persian man once produced a Northern European stranger because the prompt carried the full identity text.
Split the fields by job. The image prompt gets one short check line per anchor,
framed as a list to compare the output against. The verifier gets the full must
and never text. Open the image prompt by naming the photographs as the source
of the likeness, then describe only the scene.
Never write polish words
Do not write "8k", "hyperrealistic", "masterpiece", "stunning", "flawless" or "luxury". Do not write "cinematic" either when the target is a real person. "Cinematic 35mm" reads as an advert. A phone-photo prompt reads as real.
Name a real camera body, a real lens and a working aperture instead. Ask for unretouched skin, with pores, lines, uneven tone and a little shine. Unflattering light reads as real.
Pass three to five references, not dozens
Profile a large library to choose well. Pass only a handful to any single render. Use three to five originals from different angles: one close face, one three-quarter, one body.
Never use a generated image to establish identity. Only an original photograph can do that.
Let identity travel through a picture
Identity does not have to travel as text. Build a four-frame casting sheet: front head-and-shoulders, three-quarter, strict profile, full body. Use one plain mid-grey background. Keep wardrobe, hair and light identical across the four frames. Keep the expression neutral throughout.
Pass that sheet as a reference. Later renders then inherit the face from the picture instead of from a fresh description. The sheet does not set later crops or camera heights.
Counter-rule: never feed a busy decorated model sheet back as a reference. Nine or more panels, with labels, palette strips and callout lines, make the model reproduce the board layout instead of the person.
A reference image carries identity, not framing
Feeding the approved casting card in as the leading reference fixed the face and did nothing at all for the crop. The same panel kept returning a waist-up high-angle shot against a head-and-shoulders brief. Use a reference for who the person is, not for how the shot is composed.
The artefact others cite clears a higher bar than the ones citing it
The casting card is the reference every later render uses. It passed at exactly the pack's floor while carrying a gaunt face, a patchy beard and clothing that changed between its own four frames, and it pushed all of that downstream, where it read as model drift rather than a bad reference. Let a shot demand a higher pass score than the pack floor, and make the reference demand it.
Measure the person, do not describe them
This is the lesson that cost the most and returned the most.
A first identity lock was written from five photographs. It looked careful. Every number in it was wrong, and two of its NEVER lines forbade things the person actually does.
| trait | written from 5 photos | measured over 90 |
|---|---|---|
| forehead height | 39% of face height | 35% |
| face width | 62%, "narrow" | 72%, ordinary |
| beard cheek coverage | 25% | 20% |
| shoulder width | 2.0 head-widths | 2.2 |
The two false rules mattered more than the numbers. "Never hair falling forward onto the forehead" was contradicted by 25 of 90 photographs. "Never thick acetate frames" forbade a pair of glasses he owns and wears in 18 of them.
The failure mode is specific and easy to repeat. A small sample shows you what is distinctive about a face, so you write down an exaggerated version of it. The verifier then holds every render to that exaggeration, and correct renders fail. Scores sat between 50% and 69% for a whole afternoon against an identity lock that described a person who does not exist.
So:
- Profile fifty or more photographs and take medians. Record the sample size next to every number, so a later reader can see how much evidence is behind it.
- Write a NEVER line only for a fault a model actually produced. Never write one from intuition about what would look wrong.
- When a trait genuinely varies, do not average it into a compromise. Make it a
variant. Beard shape varied across three states here, while sparse cheeks
held across all of them. The stable part became the anchor; the varying part
became
short-beard,goateeandcleanvariants that override it.
Name one canonical photograph per trait. The medians give you the value. An authority map gives a disputed value a single arbiter, so you do not re-average the library every time two readings disagree. Write the map into the pack:
| trait | canonical photograph |
|---|---|
| glasses, brows, hair, beard density | portrait-04.jpg |
| face width, nose | portrait-11.jpg |
| shoulders, torso, build | half-body-02.jpg |
| full-body proportions | full-body-01.jpg |
Choose the image model by measurement
Prompt work has a ceiling. Model choice moved the result more than any rewrite. Run one shot through several models with the same prompt and the same verifier before committing a whole gallery.
One such comparison, same shot and prompt throughout:
| model | likeness |
|---|---|
| openai/gpt-image-2 | 75.2% |
| bytedance/seedream-v5.0-pro | 72.8% |
| wavespeed-ai/qwen-image-2.0-pro | 70.4% |
| google/nano-banana-pro | 65.6% |
| wavespeed-ai/flux-2-max | 55.2% |
Then look at the images yourself, because the top scorer lost. GPT-Image-2 drew the beard as a hard-edged cut-out that looked pasted on. Seedream produced a convincing photograph. An anchor added to catch that artefact scored the same bad image 4 out of 5, at two resolutions.
A vision verifier measures facial geometry well and does not reliably see composite artefacts. Keep render quality as a human check on the contact sheet. Do not claim an automated gate covers it.
Some models have no seed. Checked across the WaveSpeed edit models:
flux-2-pro, flux-2-max and qwen-image-2.0-pro accept a seed.
seedream-v5.0-pro, nano-banana-pro and gpt-image-2 do not. Those three also
reject a negative prompt and a guidance scale. The models that hold a likeness
best are in that second group, so there is no seed to pin. Treat score variation
between attempts as inherent, not as a bug to fix.
Failure modes
Each of these happened. None is hypothetical.
Writing an identity lock without searching for prior art. A
pooria-photoskill already warned against a narrow V-shaped chin for this person. A lock written from scratch missed it, the subject reported that exact fault, and several render cycles paid for it. Grep the skills first.A verifier that agrees with the brief. Shown the anchor text, a weak model restates the MUST line as its own observation and passes a wrong face. One scored a known-bad render 97.6%. Make the verifier write what it sees in the reference and what it sees in the candidate as two separate fields, force numeric estimates, and tell it that straight 5s mean it did not look.
Verifying a thumbnail. A downscaled candidate hid a hard-edged composite beard that was plain at full size. Send the verifier the full-resolution file.
Feeding every fault back at once. The model trades one fault for another and the score bounces: 78%, 80%, 73%. Send one correction per round.
Describing the face in the image prompt. Long prose about a face makes the model generate a face from the words and ignore the photographs. Four photographs of a Persian man returned a Northern European stranger. Give the image model one short check line per anchor; give the verifier the full text.
An identity lock built from too few photographs. It encodes an exaggeration and then fails correct renders. Profile fifty or more and use medians.
Running a gallery before validating anything. Thirteen shots at three attempts each is a lot of money spent confirming the prompt was wrong. Prove one shot first.
Scoring an anchor the shot cannot show. Height and limb length are not judgeable in a waist-up frame, so that anchor fails in a way no prompt can fix. Let anchors declare what they need in frame, and let shots declare their scope.
Picking the model by score alone. The top scorer had a visible artefact the verifier could not see. Look at the images.
Trusting an anchor you have not tested. An anchor written to catch a known artefact scored that exact artefact 4 out of 5. Test a new anchor against the image that motivated it, and say so plainly when it does not work.