The character’s arm bends the wrong way at the elbow. The neck turns past a human range of motion. The hand sits on the hip but the wrist is broken. The pose is technically there, but it’s anatomically wrong in a way your brain flags instantly even when you can’t say why.
Fastest fix: stop describing the pose in words and feed the model a skeleton. Drop an OpenPose pose map into ControlNet (SDXL or Flux both support it as of June 2026) at strength 0.8-1.0, and most broken-limb problems vanish in one generation. If you can’t run ControlNet, the second-best fix is a concrete action verb (“standing with arms crossed, weight on her right hip”) plus a wider camera. Everything below is ordered by how often it actually fixes the problem.
Common causes
Ordered by hit rate, highest first.
1. Abstract verbs like standing or sitting
a woman standing doesn’t tell the model where her arms are, where her weight is, or what her hips are doing. It picks a default, which is often anatomically lazy: arms straight down, weight even, head forward.
How to spot it: your pose word is one of standing, sitting, lying down, posing. Fix by adding concrete limb detail.
2. No pose reference
For anything beyond a default stance, text alone is underdetermined. There are infinite ways a body can kneel while reaching forward, so without an OpenPose skeleton or pose reference image the model guesses, and guesses badly on overlapping limbs.
How to spot it: you’re not using ControlNet OpenPose, a pose-reference image, or Midjourney --cref with a posed person.
3. Camera too close, limbs forced into frame
Close-up portrait plus a full-body pose description means the model crops limbs at strange angles to make them fit. Arms end mid-frame, hands disappear behind the head, elbows fold into the torso.
How to spot it: tight framing (close-up / head shot) combined with arm/hand/leg descriptions. Pull the camera back.
4. Multiple conflicting pose words
"hands on hips, arms crossed, holding a coffee cup"That’s three mutually exclusive arm positions. The model averages them and produces broken or extra arms.
How to spot it: count the distinct arm/hand actions in your prompt. More than one and you have a conflict.
5. An action word the model doesn’t know
pickleball serve, yoga warrior 2 pose, kabuki mie pose. Uncommon poses are under-represented in training data, so the model approximates and breaks.
How to spot it: the pose word is sport-specific, art-form-specific, or culturally specific. Swap in a more generic equivalent, or drive it with a pose reference.
6. The model is weak on anatomy (Flux schnell, lighter SD checkpoints)
Some lighter checkpoints prioritize style over anatomy. They produce beautiful but anatomically loose results, especially on hands and overlapping limbs.
How to spot it: run the same prompt and pose on a known anatomy-strong model (Juggernaut XL, LEOSAM HelloWorld XL, RealVisXL). If the anatomy comes back clean, your original model is the source.
Which bucket are you in
| Symptom | Most likely cause | Go to |
|---|---|---|
| Generic stiff pose, arms glued to sides | Abstract verb (#1) | Step 1 |
| Specific pose you described keeps coming out broken | No skeleton reference (#2) | Step 2 |
| Limbs cropped or folded at the frame edge | Camera too close (#3) | Step 3 |
| Two left hands, extra/fused arms | Conflicting pose words (#4) | Step 1 + Step 6 |
| Niche sport/dance/ritual pose mangled | Unknown action word (#5) | Step 2 |
| Anatomy soft everywhere even on simple poses | Weak checkpoint (#6) | Step 4 |
Shortest path to fix
Step 1: Replace the abstract verb with a concrete action
Bad to good:
| Bad | Good |
|---|---|
a woman standing | a woman standing with arms crossed, weight on her right hip, head tilted slightly down |
a man sitting | a man sitting cross-legged on the floor, hands resting on knees, leaning forward slightly |
posing | leaning against a wall with one hand in pocket, other hand holding coffee, looking off camera left |
running | mid-stride running, left foot forward and lifted, right arm forward and bent, looking ahead |
Be explicit about each arm, each leg, weight distribution, head direction, and gaze. One concrete clause per limb beats a single vague verb.
Step 2: Feed an OpenPose skeleton via ControlNet
This is the single highest-leverage fix. As of June 2026 both SDXL and Flux have working OpenPose ControlNet, so this is no longer SDXL-only.
# SDXL (A1111 / Forge / ComfyUI)
1. Find or take a photo with the pose you want
2. Drop it into ControlNet -> OpenPose Editor -> extract the skeleton
3. Preprocessor: openpose (or openpose_full for hands + face)
4. Model: xinsir/controlnet-openpose-sdxl-1.0 (current community pick)
5. Generate with your prompt + the skeleton
# Flux.1 dev (ComfyUI)
1. Same skeleton extraction step
2. Use the InstantX / Shakker-Labs ControlNet Union Pro model
3. Set the SetUnionControlNetType node to the pose mode
4. Strength 0.3-0.8 (developer-recommended range for Union Pro)
# No local setup? Build the skeleton online:
- Posemy.art: drag a 3D mannequin into the pose, export PNG
- OpenPose Editor (online): adjust a 2D skeleton, export PNG
- Either PNG drops straight into the ControlNet image slotControlNet weight 0.8-1.0 locks the pose strictly on SDXL; 0.5-0.7 lets the prompt soften it. Flux Union Pro is more sensitive, so start lower (0.3-0.5) and raise it if the pose drifts. If you use the thick default pose lines and the result is unstable, regenerate the skeleton with the thinner openpose line style.
Step 3: Pull the camera back
If the pose only breaks in a tight crop, widen the frame so the model has room to place limbs:
# Replace
"close-up portrait" -> "half body shot"
"head shot" -> "medium shot, chest up"
# Or describe the framing explicitly
"full body shot, framing head to feet"
"three quarter body, framing head to knees"A full-body pose needs a full-body frame. Mismatched framing is the most common cause of limbs vanishing at the edge.
Step 4: Switch to an anatomy-strong checkpoint
# SDXL, best anatomy as of June 2026
- Juggernaut XL v10 / Ragnarok (most versatile photoreal)
- LEOSAM HelloWorld XL 7.0 (trained against bad hands/limbs; best for poses)
- RealVisXL (strongest on faces and skin)
- Pony Diffusion V6 XL (best on stylized full-body)
# Flux
- Flux.1 dev (clearly better than schnell on anatomy)
- + an anatomy / hands LoRA from Civitai if neededIf hands and overlapping limbs are your recurring failure, HelloWorld XL 7.0 was specifically trained against bad anatomy and is the most reliable single swap.
Step 5: Negative-prompt anatomy errors (SD-family only)
Flux ignores negative prompts by default, so this step is for SD 1.5 / SDXL.
broken arm, broken anatomy, bad anatomy, deformed limbs,
disjointed limbs, extra limbs, missing limbs, twisted body,
unnatural pose, broken neck, contorted, mutilated,
warped arms, malformedStep 6: Use ADetailer or inpaint on the broken zone
If 95% of the pose is fine and only one limb is broken, fix it locally instead of rerolling the whole image:
# A1111 / Forge ADetailer
- Auto-detect hands or body (hand_yolo / person model)
- Inpaint with a prompt describing correct anatomy
- Denoise: 0.4-0.6
# Manual inpaint
- Mask the broken limb only
- Inpaint prompt: "anatomically correct arm, natural elbow bend"
- Denoise: 0.5-0.7Tool-specific notes
- Midjourney: there is no OpenPose. For a specific pose, supply a posed reference with
--cref, then set--cw 0so the model takes only the face from the reference and follows your text prompt for the body and pose. Higher--cwvalues copy the reference pose more literally. - ChatGPT / Gemini image gen: no ControlNet. You’re limited to concrete verbs (Step 1), wider framing (Step 3), and re-rolling. For anything pose-critical, move to SDXL or Flux with ControlNet.
How to confirm it’s fixed
- Zoom to 100% on each joint: elbow, wrist, knee, neck. Each bend should fall within a human range.
- Count limbs and digits in the fixed zone (poses that break often break into extra/fused limbs).
- Generate the same prompt at 4 seeds. A real fix is stable across seeds; if one seed is clean and three are broken, the pose still isn’t locked, so go back to Step 2 and raise ControlNet weight.
FAQ
Does ControlNet work with Flux now, or is it still SDXL only?
Flux has working pose ControlNet as of June 2026 through the InstantX / Shakker-Labs ControlNet Union Pro model in ComfyUI (pose mode, strength roughly 0.3-0.8). It is heavier than the SDXL OpenPose ControlNet, so keep the image size modest while testing.
My pose looks fine until I zoom in on the hands. Is that the same problem?
Partly. Pose ControlNet fixes large-limb placement but not always fine finger anatomy. Use openpose_full (which includes hand keypoints) plus ADetailer hand inpainting, or see AI image extra fingers.
What ControlNet weight should I start with?
On SDXL, start at 1.0 for a strict pose lock and drop toward 0.6 if the figure looks stiff or copied. On Flux Union Pro, start lower at 0.3-0.5 and raise only if the pose drifts from the skeleton.
Why does the exact same pose work on one model but break on another? Lighter checkpoints (Flux schnell, slim SDXL merges) trade anatomy for speed and style. Re-run on Juggernaut XL v10 or LEOSAM HelloWorld XL 7.0; if it comes back clean, your original checkpoint was the cause (cause #6).
I can’t install ComfyUI or A1111. Any way to use OpenPose? Yes. Build the skeleton in a browser at Posemy.art or an online OpenPose editor, export the PNG, and feed it into any hosted SDXL/Flux service that exposes a ControlNet or pose-reference image slot.
Prevention
- Keep a personal pose dictionary: each named pose gets a concrete one-line description.
- For any non-trivial pose, reach for ControlNet OpenPose first instead of trying to describe it.
- Match camera framing to the pose: full pose to full-body shot, head detail to head shot.
- Use anatomy-strong base checkpoints for any pose-critical work (sports, dance, action).