You asked for a symmetric pose — arms outstretched evenly, a frontal yoga stance, hands clasped in prayer — and the model returned a lopsided mess. One shoulder sits higher, one eye drifted off-center, the hands land at different vertical heights. Diffusion models have no architectural symmetry prior; symmetry is something they have to approximate from training data, and the variance is high. The reliable fix is not “ask for symmetry harder.” It is to constrain the pose with a ControlNet OpenPose skeleton, inpaint the broken half, or — for must-be-symmetric work — composite a mirrored copy.
Fastest fix (90 seconds): Load an OpenPose ControlNet with a symmetric skeleton (T-pose, prayer hands, arms-out) and regenerate. If you only have text-to-image (e.g. Midjourney), drop directional-lighting and side-specific words first, then re-roll. For a hard guarantee, mirror-composite the good half in any image editor.
Which bucket are you in?
| Symptom | Most likely cause | Go to |
|---|---|---|
| Pure text prompt, no pose input | No structural constraint | Step 1 |
| Prompt has “side-light” / directional light | Lighting bias bends the pose | Step 2 |
| Prompt names a side (right hand, left foot) | Lateralized words force asymmetry | Step 3 |
| Pose is mostly right, one half off | Local error | Step 4 (inpaint) |
| Product/packaging, needs pixel-perfect | Generation will never be exact | Step 5 (mirror) |
| Two-subject mirror pose, never clean | Multi-subject alignment is hard | Step 1 + Step 5 |
Common causes
Ordered by how often each is the actual root cause.
1. No structural pose constraint
Asking for a symmetric pose with words alone gives the model only a soft hint. Without a ControlNet pose or reference image, the model samples from its distribution of “T-pose-like” outputs, and most samples are not perfectly symmetric.
How to spot it: Are you using only text prompts, no pose conditioning? Yes = this is the root cause.
2. Asymmetric lighting or background pulled the model
The model treats the whole image jointly. If you prompted “morning side-light from camera left,” it may bend the pose toward the light — lowering the lit shoulder or making the lit hand more visible. Lighting bias quietly overrides pose intent.
3. Camera angle slightly off-axis
A 5-10 degree camera roll or yaw is enough to break perceived symmetry. The model may be drawing the pose correctly, but the projection looks lopsided.
4. Body parts colliding with hair or clothing
A symmetric pose with long hair, a flowing dress, or an asymmetric item (a single shoulder bag) gives the model permission to break symmetry so the secondary asset stays plausible.
5. Multiple subjects
Two-character symmetric poses (mirrored dancers, partner yoga) almost never come out clean. The model struggles to keep each subject aligned to the other.
6. Prompt mentions one side explicitly
“Right hand raised” or “left foot forward” force asymmetry. If your prompt has any lateralized words, that is your problem.
Before you start
- Decide whether the symmetry needs to be pixel-perfect. For editorial work, near-symmetric is usually fine. For product packaging or anatomical reference, you need exact symmetry — go straight to mirror-compositing.
- Save the seed, prompt, model, and tier of the broken image.
- Generate 4 seeds. If 4 of 4 break the same way, the prompt or workflow is structural, not bad luck.
- Scan the prompt for any word that implies a side (right, left, leading, trailing).
Information to collect
- Full prompt, model and version, seed, sampler, steps, aspect ratio.
- A measurement of the asymmetry — pixel offset of eyes, shoulders, hands on the Y axis.
- Whether a ControlNet or reference image was used.
- Whether lighting is symmetric or directional.
- Intended use case — symmetric logos and anatomical diagrams need far more accuracy than concept art.
Step-by-step fix
Ordered by ROI. ControlNet OpenPose is the single biggest move.
Step 1: Constrain with a ControlNet pose or reference
Replace soft prompting with a hard pose constraint. This is the highest-leverage step.
- SDXL / A1111 / ComfyUI: Load OpenPose ControlNet with a symmetric pose skeleton. Hand-author the skeleton in the editor or extract it from a reference photo. Set ControlNet strength around
0.8-1.0for strict adherence. - Flux (Flux.1 / Flux 2 in ComfyUI): Use the InstantX ControlNet Union Pro model in OpenPose mode — this is the current standard Flux pose ControlNet as of June 2026. Practical tip: keep
start_percentat 0 and setend_percentlow (around0.5-0.7) so the skeleton guides the early steps but does not leave OpenPose color remnants in the final image. Balance ControlNet strength against your prompt influence; Union Pro is heavier on VRAM, so keep the canvas modest while tuning. - Midjourney (V7): The old
--crefcharacter-reference flag is deprecated. As of 2026, V7 uses Omni Reference: append--oref <image_url>with--owfor weight (range1-1000, default100). For strict pose adherence point--orefat an image that already has the symmetric pose and raise--owtoward300-400; keep--owunder 400 unless you are running a very high--stylize, or results get unpredictable. Note:--orefis V7-only and is not supported on V8.
For perfect symmetry, draw the OpenPose skeleton in a vector tool and mirror it across the centerline before importing — that way the constraint itself is exactly symmetric.
Step 2: Neutralize lighting in the prompt
Switch from directional to symmetric lighting:
- Bad:
side-light from camera left - Good:
frontal soft light, even illumination from both sides
This removes one of the most common asymmetry-causing pressures.
Step 3: Remove lateralized words
Scrub the prompt for any word implying a side. Replace with bilateral language:
right hand raised->both hands raisedleading foot forward->feet together, parallelhead turned slightly->head facing camera directly
Step 4: Mask and inpaint the broken half
If the composition is locked but one half is off:
- Identify which half is correct (often the side the model rendered more confidently, or the one opposite the broken half).
- Mask the broken half and run an inpaint pass. In A1111/ComfyUI inpaint, set denoise around
0.5-0.7so it reshapes the limb without redrawing the whole region. - Prompt the inpaint with the same content as the correct half, and keep the OpenPose ControlNet active so the redrawn half lands on the symmetric skeleton.
Step 5: Mirror-composite the correct half
The reliable option for must-be-symmetric work:
- Open the image in Photoshop, Affinity Photo, or GIMP.
- Cut the correct half along the centerline.
- Duplicate it, flip horizontally, and place it on the other side.
- Blend the seam with Generative Fill (Photoshop) or a feathered layer mask.
- Touch up any artifacts at the centerline.
This guarantees pixel-perfect symmetry. Reserve it for product, packaging, and reference work where the symmetric promise is the deliverable.
Step 6: Switch to a stronger model
If you are stuck on an old model, pose adherence may be the bottleneck. As of June 2026, the strongest options for precise pose control are Flux Pro 1.1 Ultra / Flux 2 with ControlNet, Google Imagen 4, and the latest Midjourney (V8.1, with Omni Reference still on V7). SD 1.5 and pre-V6 Midjourney are noticeably weaker for symmetric work — among current models, Flux with ControlNet gives the tightest spatial control because Midjourney “interprets” the prompt more than it follows it.
How to confirm it’s fixed
- Overlay a vertical centerline guide on the image. Both eyes should be equidistant from it; both shoulders should sit at the same vertical height.
- Measure hand positions in pixels — both should land at the same Y coordinate within tolerance.
- For yoga and dance poses, check joint angles — each elbow should bend by the same amount.
- View at the final delivery size. Small symmetry errors that vanish on a thumbnail jump out at full resolution.
Long-term prevention
- For any symmetric-pose work, default to ControlNet OpenPose or a reference image. Do not rely on prompt words alone.
- Build a reusable library of symmetric OpenPose skeletons (T-pose, prayer hands, arms-out, frontal stance).
- Use symmetric lighting language by default for symmetric subjects.
- Audit prompts for side-bias before generating; strip lateralized words.
- For products and packaging, mirror-composite as a final pass — do not trust a single generation.
Common pitfalls
- Trusting that “symmetric” or “mirrored” in the prompt is enough. It rarely is.
- Forgetting that hair, clothing, and props can break symmetry even when the underlying pose is correct.
- Mirror-compositing without blending the seam — the centerline gives it away.
- Re-rolling 20 seeds chasing symmetry. After 4 bad rolls, switch to ControlNet.
- Leaving
end_percentat 1.0 on a Flux OpenPose ControlNet and getting skeleton-colored ghosting in the output.
FAQ
Q: Why won’t the model just draw symmetric outputs when I ask for symmetry? A: Diffusion models have no architectural symmetry prior. They sample from a distribution that includes many almost-symmetric outputs but rarely a perfect one. ControlNet adds a hard structural constraint that text prompts cannot.
Q: Can I get perfect symmetry from prompt alone? A: Rarely. Even with the strongest June 2026 models, perfect symmetry from text alone is roughly a 1-in-10 roll. ControlNet OpenPose or a mirror-composite is the dependable path.
Q: Why does --cref no longer work in Midjourney for posing?
A: --cref (character reference) is deprecated and is not supported in Midjourney V7. V7 replaced it with Omni Reference: --oref <image_url> plus --ow for weight (1-1000, default 100). Point --oref at an image that already holds the symmetric pose. Note --oref is V7-only.
Q: Will mirror-compositing always look natural? A: For mostly-static poses with frontal lighting, yes. With complex hair, fabric, or directional lighting, the seam may need touch-up work.
Q: What about features that are naturally asymmetric, like a side part in the hair? A: Mirror-compositing inherits one side’s asymmetries. A side part becomes a part on both sides or neither. Choose the half to mirror based on which side has fewer asymmetric details.