Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Yes—early users of Stability AI’s Stable Diffusion 3 Medium (SD3 Medium) found that ordinary prompts for people could produce fused limbs, malformed hands and feet, and bodies with incoherent anatomy. The model launched on June 12, 2024. These examples document a serious weakness in human rendering, but no published statistic shows how often it occurs, and the cause has not been reduced to one proven mechanism.
What happened after the June 2024 release
Stability AI released SD3 Medium on June 12, 2024, presenting it as an open text-to-image model for consumer PCs and laptops as well as enterprise GPUs. Within hours, users shared generations in which arms merged into torsos, hands and feet contained impossible structures, and posed or reclining figures collapsed into what online discussions called “appendage soup” or “body horror.” Ars Technica described the launch as a major step backward for human rendering compared with contemporary image models.
The failures were not limited to deliberately difficult prompts. Reports included ordinary requests for people, including figures lying on grass. That does not mean every SD3 Medium image was malformed; it means the released model could fail dramatically on a common image category that users expected it to handle.
Which model is being discussed?
SD3 Medium, not every SD3 variant
Stability AI announced the wider Stable Diffusion 3 family in February 2024 with sizes ranging from 800 million to 8 billion parameters. The controversy concerns the Medium release, which Ars Technica reported has 2 billion parameters. Stability AI described it as its “most advanced text-to-image open model yet.” Its weights were made available under the company’s Community License.
#1 Best Overall
Those details matter because performance and licensing can differ between model sizes and releases. Reports about SD3 Medium should not automatically be treated as a verdict on every SD3 checkpoint.
What users saw in the images
- Hands with fused, duplicated, or incorrectly attached fingers.
- Feet and lower legs with implausible joints or merged shapes.
- Arms and legs that blended into the torso or into another person.
- Figures lying down or holding a pose whose bones and limbs no longer aligned.
- Globally incoherent anatomy even when faces, clothing, lighting, or backgrounds looked plausible.
These are visual examples and user reports, not a controlled error measurement. There is no reliable published failure-rate statistic establishing the percentage of human generations that were malformed.
Why the anatomy may have broken
The aggressive-filtering hypothesis
The most discussed explanation in contemporaneous coverage was over-aggressive filtering of adult or NSFW material in the training data. Anatomy examples can include nudity, and removing too much such material could leave a model with fewer examples of bodies, proportions, and poses. If the remaining data overrepresented clothed, cropped, or otherwise limited views, the model could have less information from which to learn complete human structure.
Ars Technica presented this as a user and analyst hypothesis, not as a proven explanation. It is therefore too strong to say that an NSFW filter alone caused the SD3 Medium failures. Stable Diffusion 2.0 had also been criticized for human-rendering problems, while later versions improved, so training-data composition is only one possible part of a more complicated pipeline.
Stability AI’s own explanation
In its follow-up, the Stability team said SD3 Medium had “critical quality issues mainly related to body poses and words that were too rarely seen in the training set.” That statement points to two classes of problem: the model’s representation of pose and anatomy, and inadequate exposure to some language concepts during training. It does not identify a single filtering decision as the definitive cause.
What the available evidence cannot establish
- There is no systematic, published benchmark showing SD3 Medium’s human-anatomy error rate.
- The public examples do not reveal how prompts, seeds, samplers, resolutions, or post-processing were selected.
- Examples demonstrate that severe failures occur, but not how SD3 Medium compares numerically with every other image generator.
How Stability AI responded
On July 5, 2024, Stability AI acknowledged that the release had fallen short. The company wrote: “We acknowledge that our latest release, SD3 Medium, didn’t meet our community’s high expectations.” It said it was pursuing continuous improvement and added: “Before we released SD3 Medium, our initial testing indicated that it was, in most cases, a much better base model compared to SDXL, in terms of prompt adherence, diversity, detail, and overall quality.”
Rank #4
The second statement describes Stability AI’s initial internal testing, not an independently published head-to-head benchmark. It also helps explain the gap between the launch messaging and the community’s experience: a model can score well on broad qualities while still failing badly on a specific category such as human poses.
What can—and cannot—be concluded from comparisons
Readers often ask whether SD3 Medium is worse than SDXL, Midjourney, or DALL-E 3. The available reporting supports a narrower answer than a universal ranking.
| Comparison area | What is established for SD3 Medium | What is not established |
|---|---|---|
| Human anatomy | Early users and journalists documented severe failures involving hands, feet, limbs, and posed figures. | No controlled benchmark establishes a failure rate or a definitive ranking against SDXL, Midjourney, or DALL-E 3. |
| Prompt adherence, diversity, detail, and overall quality | Stability AI said its initial testing usually found SD3 Medium better than SDXL on these measures. | The cited coverage does not provide the test set, scoring method, or independent replication. |
| Typography and text rendering | Not established by the cited reports for a controlled comparison. | No defensible numerical or universal advantage can be claimed. |
| Hardware and hosting | The Medium model was aimed at consumer PCs and laptops as well as enterprise GPUs. | Exact hardware requirements and equivalent requirements for competing services are not stated here. |
| Openness and local use | Stability AI released the weights as an open model under its Community License. | “Open” does not by itself specify an identical license or local-running process for other generators. |
| Commercial use | In a 2024 Community License update, Stability AI said free commercial use applied to individuals and small businesses with annual revenue below USD $1 million, subject to the license terms. | That was a dated policy statement; check the current license before commercial deployment. |
Practical guidance for creators evaluating SD3 Medium
Test anatomy separately from visual style
Do not judge a checkpoint only by attractive landscapes, objects, or portraits. If people matter to your project, test hands, feet, full-body standing poses, seated poses, and reclining figures as separate categories. Save the exact prompt and generation settings so a result can be reproduced.
Inspect every human generation
A polished face or convincing background can conceal broken joints or duplicated fingers. Review the entire body at the intended output size before using an image in a design, illustration, or production pipeline.
Use another generator when anatomy is mission-critical
The reports establish a serious SD3 Medium weakness, not a universal failure of every image model. If accurate human structure is more important than running an open model locally, evaluate alternatives on the same prompts and judge the outputs yourself. The cited coverage does not support a blanket claim that any one alternative always wins.
Check the license at the time of use
Because the Community License terms and eligibility thresholds can change, read the current Stability AI terms before selling images, embedding the model in a product, or using it for a business.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →The accurate takeaway
Stable Diffusion 3 Medium was a 2-billion-parameter open model released on June 12, 2024, and its early public reception exposed conspicuous failures in human anatomy. The filtering theory is plausible but unproven; Stability AI instead acknowledged critical problems with body poses and rarely represented words. The strongest defensible conclusion is therefore specific: SD3 Medium could produce horrific mangled bodies often enough in ordinary testing to undermine confidence in human-image workflows, while the public record still lacks a measured failure rate or a single confirmed cause.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
