Every text-to-image model ships with a document that most art criticism ignores: the model card. It is not a press release. It is a curatorial statement about whose body counts as the baseline, written by the people who assembled the training data and set the filters. Read it the way you would read a museum’s collection policy, because that is what it is.

The Stable Diffusion v1-4 model card states plainly that the model “was trained on subsets of LAION-2B(en), which consists of images that are primarily limited to English descriptions,” and that “white and western cultures are often set as the default.” That is not an inference from a screenshot. It is the developers’ own sentence, published on the model card. The same card notes that “the ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.”

This matters for anyone making queer and gender-critical work in Berlin’s alternative art ecologies, because the model card is a named decision point. It tells you what was scraped, what was filtered, and what the developers already know fails. The question is not whether the model “wants” to render a particular body. The question is what the documentation commits to, and what it leaves unstated.

Three mechanisms, not one

Criticism that treats every failed render as a “refusal” produces bad documentation. The v1-4 card describes at least three distinct mechanisms that shape what comes out.

Stated limitations. The card lists: “Faces and people in general may not be generated properly.” It also states the model “cannot render legible text” and “does not perform well on more difficult tasks which involve compositionality.” These are failures, not refusals. They are the model’s own admitted incompetence at rendering the human form with fidelity.

The safety checker. The v1-4 card describes a Safety Checker that “works by checking model outputs against known hard-coded NSFW concepts,” with the concepts “intentionally hidden to reduce the likelihood of reverse-engineering this filter.” The checker “compares the class probability of harmful concepts in the embedding space of the CLIPTextModel after generation of the images.” This is a filter, not a limitation. It blocks after generation.

Platform policy. The Stable Diffusion Online acceptable-use policy (last updated April 21, 2026) states: “You may not use Stable Diffusion Online to generate, edit, upload, request, or distribute sexual, pornographic, nude, fetish, erotic, sexually explicit, adult, or NSFW content.” That is a flat prohibition at the platform layer. It is not the same as the model card’s stated limitations, and it is not the same as the safety checker. It is a third mechanism, and it is the one most likely to affect queer and trans artists working with nudity, kink, or non-normative embodiment.

Conflating these three produces criticism that cannot be checked. Separating them produces a log that can be cited.

The training data is documented, and so is its filtering

The v1-4 card states the model “was trained on a large-scale dataset LAION-5B which contains adult material and is not fit for product use without additional safety mechanisms and considerations.” It adds: “No additional measures were used to deduplicate the dataset.”

The v2 model card states the model “was trained on a subset of the large-scale dataset LAION-5B, which contains adult, violent and sexual content. To partially mitigate this, we have filtered the dataset using LAION’s NSFW detector.” The v2 card also states that “Stable Diffusion v2 mirrors and exacerbates biases to such a degree that viewer discretion must be advised irrespective of the input or its intent.”

Stability AI’s safety page states that “in versions of Stable Diffusion developed exclusively by Stability AI, we apply robust filters on training data to remove unsafe images.” It also states that “in our hosted applications and APIs, we apply robust filters on prompts and outputs to screen for unsafe content that violates our platform terms of service.”

What none of these documents say is that queer or trans bodies are a named category. The omission is the point. The filters operate on “NSFW” and “unsafe” as undefined categories, and the concepts are “intentionally hidden.” An artist cannot know from the documentation whether a given prompt will be blocked, distorted, or rendered. The only way to know is to log it.

The gap between the safety page and the platform policy

Stability’s safety page uses the language of “robust filters on prompts and outputs” and “unsafe content that violates our platform terms of service.” The Stable Diffusion Online acceptable-use policy uses the language of a flat prohibition on “nude” and “fetish” content. These are different documents with different scopes, and the gap between them is itself evidence.

The Stable Diffusion Online policy describes a third-party service that “may use models in the nano-banana series and seed-dream series” for image generation. It is not Stability AI’s own policy. But it is the policy that governs a widely used web interface, and it states that the service “may review prompts, uploads, and outputs using automated systems and manual review” and “may block requests, remove content, suspend access, or terminate accounts.”

For an artist, this means the boundary of the renderable is set at multiple layers: the model’s stated limitations, the safety checker’s hidden concepts, and the platform’s acceptable-use policy. A documentation practice that records only the final image misses two of the three.

A documentation method

The useful unit is not the image. It is the log. For each prompt, record:

• The exact prompt text, including language.
• The model version and checkpoint (v1-4, v2.1, SDXL, SD 3.5, or a fine-tune).
• The platform or API used, and its acceptable-use policy version.
• Whether the output was blocked, distorted, or rendered.
• The model card’s stated limitations that apply (faces, compositionality, legible text).

The model card supplies the reference against which the log is read. If the card says “faces and people in general may not be generated properly,” and your log shows a pattern of distortion for a particular kind of body, you have a documented finding, not an anecdote.

This is not a claim that the model refuses to render queer or trans bodies. The retrieved documents do not name those categories. It is a claim that the model card and the acceptable-use policy are quotable documents that set the conditions under which any body is rendered, and that those conditions are skewed toward a default that the developers themselves describe as white and western.

What the marketing says, and what the card says

Stability’s Stable Image page describes Stable Diffusion 3.5 as offering “market-leading prompt adherence” and “diverse outputs” that “represent a diversity of people and scenes around the world.” That is marketing copy. It is not a measurement. The model cards for v1-4 and v2 are the documents that state limitations and biases in the developers’ own words. When the marketing and the model card diverge, the model card is the document that can be cited.

The v2 card states that the model “was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.” That sentence is a disclaimer. It is also a curatorial statement: the model does not claim to represent people accurately. It claims to generate images. The difference matters for anyone using it to depict a body that is already underrepresented in the training data.

The art-critical stakes

The model card is a curatorial document. It names the training data, the filters, and the stated limitations. It makes a claim about whose body counts as the baseline, and it does so in language that can be quoted. Reading it as art criticism rather than as technical documentation is the point.

For artists working in Berlin’s alternative spaces, the model card is also a reminder that gatekeeping is not only institutional. It is embedded in the tools. The Senatsverwaltung funding guidelines, the Ausstellungshonorar standards, the jury scoring sheets—these are documents that decide who gets shown. The model card and the acceptable-use policy are documents that decide what gets rendered. Both are decision points. Both can be cited.

The question is not whether the model is neutral. The developers already answered that. The question is what you do with the answer.

FAQ

Does the Stable Diffusion model card say it refuses to render queer or trans bodies?
No. The retrieved model cards and policies do not name queer or trans bodies as a category. The v1-4 card states that “white and western cultures are often set as the default” and that “faces and people in general may not be generated properly.” The Stable Diffusion Online policy prohibits “nude” and “fetish” content without specifying whose bodies are affected.

What is the difference between a limitation and a refusal?
A limitation is a stated failure: the v1-4 card says the model “cannot render legible text” and “does not perform well on more difficult tasks which involve compositionality.” A refusal is a filter: the v1-4 card describes a Safety Checker that blocks outputs based on hidden NSFW concepts. The Stable Diffusion Online policy describes a platform-level prohibition enforced by automated and manual review. These are different mechanisms and should be documented separately.

Where can I find the model card?
The Stable Diffusion v1-4 model card is published on Hugging Face at the CompVis/stable-diffusion-v1-4 repository. The v2 model card is available in the Stability AI GitHub repository. Stability AI’s safety page and terms of service are on stability.ai. The Stable Diffusion Online acceptable-use policy is on stablediffusionweb.com.

Does the model card tell me what the safety checker blocks?
No. The v1-4 card states that the NSFW concepts are “intentionally hidden to reduce the likelihood of reverse-engineering this filter.” The only way to document what is blocked is to log prompts and outputs systematically.

Is the Stable Diffusion Online policy the same as Stability AI’s policy?
No. The Stable Diffusion Online acceptable-use policy describes a third-party service that “may use models in the nano-banana series and seed-dream series” for image generation. Stability AI’s safety page describes filters on its own hosted applications and APIs. They are separate documents with separate scopes.

What should I record in a prompt log?
Record the exact prompt text, the model version and checkpoint, the platform or API used, whether the output was blocked, distorted, or rendered, and which stated limitations from the model card apply. The model card is the reference document; the log is the evidence.

Noa Vex writes on queer and gender-critical art criticism, focusing on the documents that decide who gets shown and what gets rendered.