Every face comes back too symmetrical
Models average toward a symmetrical face that reads as generated. Asymmetry cues to prompt, plus the guidance and checkpoint choices that stop flattening.
The generated face that looks "off" at a glance is often not a bad likeness. It is a face whose two sides agree too well. Same brow height, same lid crease, same nasolabial fold, ears at the same pitch, a hairline that mirrors. Real faces do not do that. Training data, especially beauty stills and illustration, is full of faces that have been pushed toward the midline. The model learns the average, and the average is symmetrical.
You can see it without zooming. Cover one half of the face, then the other. If both halves could be the same person in a police sketch, the sampler flattened the identity. The fix is to prompt a small number of named asymmetries, to stop CFG dragging the face back to the mode of the distribution, and to pick a photographic checkpoint rather than an illustration one.
This is a different job from directing gaze and micro-expression, which is about performance over time. Here the face already reproduces. It reproduces as a mannequin.
Why the average face is a symmetric face
A text-to-image model does not retrieve a person. It denoises toward the most probable face given the prompt. The most probable face, across a training set of retouched portraits and constructed drawings, is frontal, even, and both sides matching. High guidance makes this worse: CFG exaggerates the conditioned prediction, and the conditioned prediction for "a portrait of a woman" is the centre of that cluster. The centre is symmetric.
Two other flatteners sit on top.
Frontality. A dead-on camera is the pose in which left-right agreement is most visible and most rewarded. A slight turn (three-quarter, even ten degrees) hides some of the effect and is closer to how people are actually photographed. "Portrait, looking at camera, symmetric lighting" is a symmetry machine.
Illustration and beauty priors. Drawn faces are constructed on a midline. Beauty retouching pushes brows, eyes, and mouth toward it. If the checkpoint's prior is "attractive face" in that sense, your prompt is arguing with the weights every step. A photographic fine-tune argues less.
Adjectives do not break this. "Slightly asymmetrical face, unique features, photorealistic" is a category, not a geometry. The model has no privileged point inside "slightly asymmetrical", so it ignores the phrase and returns the average again. You have to name the disagreement between the two sides.
Asymmetry is a list, not a vibe
Write one or two physical mismatches. Not ten. A collage of defects reads as damage, not as a person. The list below is a menu, not a checklist.
| Cue | Prompt it as | Why it works |
|---|---|---|
| Brows | left brow a millimetre higher; a small gap in the right brow | brows are the first place a viewer checks for twins |
| Eyes | a slightly lazier right lid; catchlight stronger in the left eye | paired organs that never quite match in a photograph |
| Mouth | a faint smirk on the left; a deeper nasolabial fold on the right | breaks the stamped-on smile |
| Skin marks | a mole above the left lip; a small scar through the right eyebrow | a landmark the sampler has to place on one side |
| Hair | a part on her left; a cowlick at the crown | hair is allowed to be uneven even in beauty work |
| Ears / jaw | the left ear sits a little higher; jawline softer on her right | structure, not makeup |
| Light | window light from camera left, shadow heavier on the far cheek | even lighting is what makes symmetry obvious |
A working clause:
Three-quarter view, ten degrees off camera. Left brow a millimetre higher than the right, a mole above the left corner of the mouth, hair parted on her left. Window light from camera left, the far cheek in light shadow. Natural skin texture, visible pores.
What to avoid:
- "perfectly symmetrical, beauty portrait, studio beauty dish from both sides"
- "slightly unique, interesting face" with no named landmark
- stacking five moles, a scar, a lazy eye, and uneven ears on one pass (that is a character sheet of injuries)
If the person has to recur, freeze the landmarks. Character consistency is already fragile across generations; a mole that wanders from cheek to cheek is worse than no mole. Put the same two marks in the same words every time, the way you would freeze a character block, or better: generate one canonical still and derive later shots from a reference image so the geometry is pixels rather than adjectives.
Guidance and the checkpoint flatten independently
You can write a perfect asymmetry clause and still get a stamped face if the sampler is crushing toward the mode.
Lower CFG on stacks that expose it. The same 3–7 portrait band that keeps skin from going plastic also keeps features from being averaged into each other. High guidance is obedient in the way a beauty filter is obedient. It delivers "a face" very clearly. Distilled and turbo variants need even lower values; the parent's favourite 9 will stamp them.
On Flux-family endpoints the slider often is not there. Flux 2 Max and similar hosted models run without a CFG channel you can usefully turn. The work moves onto the prompt (named landmarks, three-quarter view, directional light) and onto model choice. An illustration model will keep returning a constructed face. A photoreal model such as Flux 2 Max, run from the image generator, has a prior that already includes lopsided real people.
Light from one side is an easy asymmetry. A beauty dish on axis is a symmetry light. Window light from camera left forces the two halves to render differently even if the geometry is still too even. Combined with a slight head turn, it does more than a paragraph of "imperfect" adjectives.
Do not expect the prompt to hold a specific person. More adjectives narrow the range; they do not pin an identity. If the brief is "this face, every time", stop describing and supply a reference. If the brief is "a face that does not look generated", named landmarks plus directional light plus a photographic model are enough.
Repair one feature; do not reroll the skull
Once the bone structure is usable, stop regenerating the whole head. Each reroll resamples identity. That is how a "more asymmetrical" pass quietly becomes a different person.
A masked inpaint on one landmark is the right tool: a mole, a brow, a lazier lid, a deeper fold. Keep the mask tight enough that the skull stays put, loose enough (a feathered edge, a little margin) that the new mark can blend. Denoise in the correction range, not a full regenerate. The same photo-editor pass you would use on a bad hand works on a too-even brow.
A sequence that does not wander:
- Generate at three-quarter view, directional light, CFG in the portrait band if you have it, on a photoreal model.
- Pick the take whose structure you would cast, ignoring that the two sides match.
- Add at most two landmarks to the prompt and run once more with the seed locked, or skip to step 4 if the structure is already right.
- Inpaint one landmark. Stop.
- If the face has to come back in later shots, promote this file to a reference. Do not go back to adjectives.
If you are iterating in the photo editor, treat symmetry like anatomy: local repair, not a new cast. When you still need a text block, freeze the landmark sentences inside it verbatim.
FAQ
Will listing flaws make the face look injured or ugly?
It will if you list too many, or if you pick damage (burns, missing teeth, a collapsed eye) when you meant living variation. One mole, a brow a millimetre off, a hair part, and directional light are enough. The goal is a photograph of a person, not a catalog of defects.
Does a reference image fix the symmetry on its own?
If the reference is a real face, yes, because you are no longer sampling the average. If the reference is itself a generated, symmetric face, you have locked the problem in. Check the reference the same way you check a take: cover each half. For a recurring character, one slightly lopsided canonical still is more valuable than a paragraph.
Why do illustration and anime models do this more?
Drawn faces are built on a midline. The prior is construction, not photography. You can still add a mole, but the bones will want to agree. For a face that has to pass as photographed, use a photoreal model and photographic light. For a face that has to pass as drawn, symmetry is often the style, and fighting it is the wrong brief.
Can I inpaint only one side of the face?
Yes, and that is often better than touching both. A lazier lid or a deeper fold on one side is a one-mask job. Inpaint both sides at once and you risk a new, also-symmetric, different person. Keep the denoise low enough that the skull reads through the patch.