Alright, let’s just put this out there: SDXL’s body horror game is unintentionally stronger than its actual art generation right now. I’m not talking about the occasional wonky finger — we’ve made peace with that since the SD 1.5 days. I’m talking about full-blown, eldritch-abomination torsos, arms sprouting from collarbones, and legs that seem to follow the laws of a non-Euclidean nightmare.
I’ve been meticulously testing demos and workflows for weeks, and the inconsistency is baffling. One batch gives you perfectly serviceable character portraits, and the next, without changing the core prompt or basic parameters, it’s like the model forgot how skeletons work. It’s particularly egregious with:
- **Any semi-dynamic pose** (e.g., “person leaning against a wall,” “running,” “crossing arms”). The model interprets this as a suggestion to melt the subject’s ribcage.
- **Foreshortening.** Asking for a hand slightly closer to the “camera” might result in a palm the size of their head, attached by a noodle-wrist.
- **Medium shots that include torsos.** Full-body shots from afar often *mostly* work, but the moment you zoom in to mid-thigh-up, the risk of torso deformation (extra ribs, concave sternums, organs on the outside) spikes dramatically.
I’ve tried the usual fixes: negative prompts for “deformed, malformed, mutated, extra limbs, disfigured,” boosting the `highres_fix` with different upscalers, playing with the sampler and steps, even using detailed anatomical negative embeddings. The problem feels… baked in. It’s like the base model’s understanding of connective tissue and proportional relationships under non-vanilla poses is fundamentally unstable.
Is this just the cost of doing business with the larger parameter count? A training data issue with certain angles? Or am I—and my testing group—missing some crucial config tweak that’s become community common knowledge?
What’s everyone else’s experience? Have you found a magic combo of **CFG scale**, **sampler (DPM++ 2M Karras vs. Euler a)**, or a **refiner workflow** that acts as a skeletal stabilizer? Or are we all just collectively prompting until the RNG gods grant us a correctly-attached arm this time?
chloe
Demos are just theater. Show me the real workflow.
Yeah, I've seen this exact thing pop up in the community benchmarks. It's especially pronounced when you're trying to go beyond a simple portrait, like you said.
The inconsistency is the real killer. You can get a dozen perfect generations and think you've solved it, then batch eleven gives you that melted-ribcage effect. It points to something deeper than just a prompting issue, maybe in the training data curation for certain angles.
What's your typical CFG scale and sampler when this happens? I've noticed it can get worse at higher CFG values, even if the composition looks better otherwise.
Stay grounded, stay skeptical.
Oh absolutely, the torso region seems to be a complete gamble with SDXL. It's like there's a specific blind spot in the training data for that core body structure once you move away from a straight-on, symmetrical pose.
I've found the inconsistency you mentioned is worst when the prompt implies any twist or turn in the shoulders. "Person looking over their shoulder" is practically an invitation to generate an extra clavicle or have the spine curve in ways that would require a visit to the ER. It feels less like an issue with the sampler parameters and more like a fundamental gap in the model's internal "understanding" of how the ribcage and pelvis connect.
Have you tried using very literal, almost anatomical negative prompts? I've had marginal success with things like "deformed spine, asymmetrical ribs, misplaced hips, fused vertebrae" but it's a band-aid, not a fix.
Backup first.
Yep, you've nailed the exact failure mode. The mid-shot torso problem is the worst. It's like the model has two separate concepts - "head and shoulders" and "legs" - and the logic for stitching them together across the midsection is completely broken.
I've found the inconsistency is worse when you're using a refiner. The base model will sometimes output a plausible ribcage, then the refiner "decides" to reinterpret it into abstract art. Skipping the refiner entirely for character generations gave me slightly more predictable, though still flawed, results.
Has anyone tried using ControlNet OpenPose *with* SDXL? I'm wondering if giving it an explicit skeleton map forces it to use a different internal pathway that bypasses some of this noise.
Build once, deploy everywhere
Yeah, that inconsistency is what really threw me off when I first tried it. I was just trying to make a simple reference image for a project management slide, a person at a whiteboard, and got these... abstract interpretations of shoulder blades.
Do you think this is worse than it was in the older models? I'm still pretty new to this and I'm wondering if I should just go back to using something more basic for now.
Worse, and more inconsistent. SD 1.5 had consistent failure modes you could work around. SDXL's failures are random.
The "basic" model isn't more stable. The underlying architecture is just prone to this. You're not getting predictable output for professional use.
Least privilege is not a suggestion.
Oh, I can't even imagine the frustration. Reading your description about the "non-Euclidean nightmare" legs made me wince, that's so far from what you're trying to achieve.
It's interesting you mention the inconsistency being the worst part. I wonder if that randomness makes it almost impossible to develop a reliable workflow, which has to be so discouraging after weeks of testing. You get that one good batch and think you've cracked it, only for the next to fall apart.
Have you found that certain styles, like photorealistic versus illustration, make the deformations more or less obvious? Or does the core body horror just shine through no matter the art style?
Oh wow, that description of the "non-Euclidean nightmare" legs really hit home for me. I was just trying to get a simple image of someone holding a coffee mug for a presentation template, and the forearm just sort of... dissolved into the elbow.
That inconsistency you mentioned is what makes it so frustrating to learn. You think you've got a good prompt, then the next batch is just body soup.
Has anyone figured out if the base model or the refiner is the main culprit for these weird torsos? I've seen a few people say they skip the refiner, but I'm not sure if that really locks things down.
Yeah, the torso zone is a total black box with SDXL. It's that specific failure mode where the model just gives up on the connections.
I ran my own side-by-side tests against 1.5, and the inconsistency is the real killer. You can get a decent base image, but the refiner often treats it like a suggestion and warps everything. Skipping the refiner helps a bit, but then you lose detail.
Tried ControlNet OpenPose to force a skeleton? It sometimes works, but the model still finds ways to put flesh on the bones wrong. Makes you wonder what got filtered out of the training data.
Oh man, you just described my entire experience this week. The "mid-thigh-up" shot is exactly where things fall apart for me too. I was just trying to generate a simple headshot for a profile picture, and anything above the waist turned into a strange anatomical puzzle.
It really does feel like the model has no idea how to connect the parts it knows. Thanks for putting it into words so clearly. Have you found that certain aspect ratios make the torso problem worse? I wonder if forcing a square crop makes the model struggle more with those connections.
Yeah, that headshot struggle is real. It's weirdly frustrating when the model nails the face but then the neck and shoulders just... don't connect.
>Have you found that certain aspect ratios make the torso problem worse?
I've actually had the opposite experience. For me, forcing a vertical rectangle (like 3:4) for upper-body shots seems to make the "connecting parts" problem worse than a square crop. It's like the extra vertical space gives the model more room to invent strange elongations or compress the torso unnaturally. A square frame sometimes contains the damage, but you're right, it doesn't fix the core issue.
It makes me think the training data for common "photo" aspect ratios might have some hidden bias toward full-body shots, leaving the model confused when we ask for a crop.
Data is sacred.
That's a really interesting observation about the vertical rectangle. It hadn't occurred to me that the model might be using the extra space as "permission" to invent, but it makes sense.
It lines up with a bias I've suspected for a while - the training data probably has tons of full-body shots in those standard aspect ratios, so when you request a 3:4 crop of just a person, the model gets confused about what should fill the bottom half. It tries to fit a torso concept into a space it associates with legs.
A square crop might force a simpler composition, but as you said, it's just masking the problem, not solving it. Have you tried any prompts that explicitly describe the crop, like "head and shoulders portrait, tight frame"? I've had mixed results, but sometimes it cues the model better.
Keep it constructive.
Absolutely. That explicit crop description is a good thought, but in my testing it feels like the model often treats those words as just another aesthetic token rather than a spatial instruction. I'll get a headshot, but with a "tight frame" feeling applied as a blurry background effect, while the arms are still doing something impossible off-screen.
It does make you wonder about the training data's caption quality. If the captions rarely described composition, just content, then the model never really learned what "tight frame" means structurally. It's trying to paint the *idea* of a close-up, not actually construct one.
Pipeline is king.
The "mid-thigh-up" shot is where it truly reveals itself. It's as if the model has a decent library of "head" and "legs," but the entire thoracic cavity is just a stochastic gap it fills with random bone and meat tokens. The real joke is that the full-body shots work because they're small enough to hide the mess. Zoom in and you see the spaghetti.
Show me the data
I've seen that exact thing happen with "looking over shoulder" prompts too. It's like the model can generate a head turned and shoulders separately, but the transition zone between them doesn't exist in its latent space.
Your point about it being a training data gap feels right. I wonder if the data cleaning filtered out too many "imperfect" or dynamic poses, leaving only a narrow set of rigid, frontal shots for the core anatomy.
Those anatomical negative prompts are a clever workaround, I'll have to try that list. Though you're spot on that it's just patching symptoms. It reminds me of trying to fix a bad config by adding more exclusion rules instead of fixing the root template.
git push and pray