Showing, Not Telling: How ITI-GEN Makes Diffusion Models Inclusive Without Retraining

Text-to-image models have a representation problem that is by now well documented. Ask Stable Diffusion for "a headshot of a person" and you receive a strikingly narrow slice of humanity. The model has internalized the statistical regularities of its training corpus, and those regularities encode the biases of whatever was scraped from the web. The question that matters for practitioners is not whether this bias exists but what to do about it without the expense and impracticality of curating balanced datasets and retraining billion-parameter models.

The paper ITI-GEN: Inclusive Text-to-Image Generation (Zhang et al., Carnegie Mellon University and Google, 2023) offers an answer that is both pragmatic and conceptually interesting. Rather than fighting the model with more elaborate text, it abandons text as the control signal for hard-to-describe attributes and uses reference images instead. The result is a method that steers a frozen diffusion model toward uniform attribute coverage without touching a single weight.

The Limits of Talking to the Model

The naive approach to inclusive generation is what the authors call hard prompt searching: enumerate the categories you want represented and append them to the prompt. Want diverse eyeglass representation? Generate half your images with "a headshot of a person with eyeglasses" and half with "without eyeglasses." This sounds reasonable, and for some attributes it works tolerably well. But it fails for two distinct reasons.

The first is linguistic ambiguity. Many visual attributes resist clean verbal description. Skin tone is the canonical example. The Fitzpatrick scale or the Monk scale partition skin tone into categories that are perfectly clear when you see them side by side, but the corresponding text descriptions are vague, culturally loaded, and inconsistently interpreted by a model trained on web captions. You cannot reliably summon "Monk skin tone 6" with a phrase.

The second is model misrepresentation. Even when a category can be named, the model may not synthesize it faithfully because the text token is poorly grounded in the training distribution. Negation is particularly fragile; "without eyeglasses" can still produce eyeglasses because diffusion models handle negation badly.

The authors' core observation is that the failure is not in the diffusion model's generative capacity but in the text interface itself. A picture conveys "this skin tone" or "this kind of eyeglasses" with a precision that no caption matches. So the natural question becomes: can reference images guide the construction of prompts that the model can actually act on?

Translating Visual Differences Into Prompt Embeddings

ITI-GEN's mechanism is elegant because it operates entirely in the shared vision-language embedding space that diffusion models already rely on, namely CLIP. The pipeline works as follows.

For each attribute of interest (say, eyeglasses), the user supplies a small set of reference images for each category: some images of people with eyeglasses, some without. The method then learns a set of prompt embeddings, one per category, by optimizing a training objective in the joint CLIP space. The key idea is directional: the difference between the learned prompt embeddings for two categories should align with the difference between the corresponding image embeddings. In other words, the method translates the visual contrast between "with glasses" and "without glasses" into a contrast between two learnable token vectors.

Several properties of this design deserve attention:

• No fine-tuning. The diffusion model and CLIP encoder stay frozen. Only a small set of prompt embeddings is learned, which is cheap and fast.
• Composability across attributes. Because each attribute gets its own learned embeddings, multiple attributes (skin tone, age, eyeglasses, and so on) can be combined, and the method targets uniform coverage across the joint set of categories.
• Inclusiveness by construction. At generation time, sampling is equalized over the learned category prompts. If you want a 50/50 split, you sample the two prompts equally. The uniformity is enforced by the sampling procedure rather than hoped for from the model.
• Plug-and-play compatibility. The learned prompts integrate with downstream tooling such as ControlNet and instruction-based editing models, since they are just embeddings prepended to the conditioning.

The headline empirical result is the shift from a roughly 92/8 imbalance in eyeglass representation to a balanced 50/50, achieved without any model surgery. The authors validate on human face datasets including CelebA and FairFace, and extend to scene attributes, demonstrating that the approach is not narrowly tuned to facial features.

What This Tells Us About Controllable Generation

The deeper contribution of ITI-GEN is not the specific debiasing result but the demonstration that reference images are a more expressive control interface than text for a meaningful class of attributes. This reframes the problem. Bias mitigation in generative models has traditionally been posed as either a data problem (collect balanced data) or a model problem (retrain or fine-tune for fairness). ITI-GEN reposes it as an interface problem: the failure was largely in how we were asking, not in what the model could do.

This connects to a broader trend in the field. Textual inversion and prompt tuning showed that small learnable embeddings can encode concepts a frozen model already implicitly knows. ITI-GEN applies the same philosophy to attribute control and fairness. The model already contains the capacity to render diverse skin tones and ages; what was missing was a reliable handle to access that capacity uniformly.

There are honest limitations to acknowledge. First, the method's fairness is only as good as the reference images. If your skin tone reference set is itself unbalanced or unrepresentative, the learned prompts inherit that problem; you have moved the curation burden from training data to reference images, which is a smaller burden but not zero. Second, disentanglement across correlated attributes is not guaranteed. Eyeglasses correlate with age in many datasets, and the directional alignment objective may bleed correlated features into the learned prompts. Third, the categorical framing of attributes is itself a modeling choice with normative weight. Partitioning skin tone or age into discrete bins, and enforcing uniformity over those bins, encodes a particular definition of fairness that may not match every deployment context. Uniform coverage is one fairness criterion, not the only one.

My Take

As I noted in a shorter post, the clever move here is sidestepping language entirely for attributes that text handles poorly. The strategic implication is what interests me most. If prompt embeddings learned from reference images can reliably steer a frozen model, then bias correction becomes a plug-in rather than model surgery.

I expect prompt embedding banks to become a standard interface for controllable generation over the next few years. Imagine "fairness presets" shipped as swappable embedding files: a skin tone balancer, an age balancer, a geographic diversity balancer, each a small artifact you load alongside a base checkpoint. This is far more maintainable than maintaining a zoo of fine-tuned checkpoints, and it separates the concern of fairness policy from the concern of model capability. A studio could audit, version, and swap its fairness presets independently of model upgrades.

The risk worth flagging is that this same mechanism is dual-use. A method that can equalize attribute distributions can also skew them deliberately. The same embedding bank that balances skin tone could be repurposed to homogenize it. The technical neutrality of the tool means governance has to live at the level of who curates and ships the embedding banks, not the level of the method itself.

Conclusion

ITI-GEN makes a focused, well-supported argument: for attributes that resist textual description, reference images provide a better control signal, and learned prompt embeddings can translate visual differences into actionable prompts that steer a frozen diffusion model toward uniform coverage. The practical payoff, achieving balanced representation without fine-tuning, is significant precisely because it is cheap and modular.

The open questions are about generalization and governance. How well does the directional alignment objective scale to many simultaneous attributes with strong correlations? Can the categorical, uniformity-based notion of fairness be relaxed to support continuous or context-dependent targets? And as embedding banks proliferate, who decides what "inclusive" means for a given deployment? The technical contribution is clear; the harder work of deciding how to use it well is just beginning.