Why AI generated images all have the same look
HumanSignal Team
Generate a hundred images across a dozen prompts and a pattern shows up that is hard to unsee: a narrow palette, soft and directionless lighting, a shallow depth cue, and a compositional balance that resolves the same way every time. The common explanation is that the training images were homogeneous to begin with. That explanation is incomplete, and the more useful one has a measurement attached. Much of the sameness is introduced after pretraining, during the alignment stage where a model is tuned toward human preference. This piece covers the mechanism, the published measurements, and what preference data has to look like to avoid it.
The first move is to stop treating this as an impression. Output homogeneity has measurable dimensions, and the relevant ones are distributional rather than per-image: how widely a model's outputs spread across the space of plausible responses to a prompt, and whether particular attributes have become overrepresented.
Work on large-scale reinforcement learning for diffusion models supplies a concrete number. Fine-tuning a base model against ImageReward, a reward model trained on human preference, moved statistical parity from 0.567 to 0.659 on a skintone diversity measure. On that metric 0 indicates a uniform distribution, so movement upward is degradation, and the tuned model was distributing outputs less evenly than the base model it started from. The reward model had no fairness term in it. Optimizing for aesthetic preference alone was sufficient to produce the regression.
That single result is worth holding onto because it makes the phenomenon concrete and directional. Preference tuning improved the thing it measured and degraded something it did not measure, in the same run.
The distinction between per-image quality and distributional quality is what makes the effect easy to miss in normal review. Every individual output can be defensible, well composed, and on brand, while the set of outputs has quietly lost most of its range. Reviewers evaluating one image at a time have no way to see it, because the property is only visible across many generations. Any process that samples a handful of outputs for approval will report that quality is holding, which is accurate and beside the point.
Two effects compound during alignment, and neither requires the training images to have been homogeneous.
A reward model assigns a scalar to an image. Optimizing against it moves the generator toward whatever raises that scalar, including properties that were never intended to be part of the objective. The authors of the diffusion work above describe this as an alignment tax, noting that a model trained with a reward concerned only with one aspect of images may learn to neglect sample quality or overall diversity. The behavior is not a defect in the reward model so much as a consequence of using one number to stand in for a judgement that had several parts.
The compression happens earlier than most teams realize. It happens when the preference data is collected, because a reviewer choosing between two images produces a single verdict that folds together composition, palette, typography, lighting, and prompt fidelity. Published work on designer-annotated preference data makes the point directly: one label collapses several design criteria into a single score and the per-criterion signal is lost. A reward model trained on those labels learns a blend, and a generator optimizing the blend converges on whatever configuration maximizes it, which is a narrower target than the range of images a designer would have accepted.
The second effect is overfitting to the reward itself. Research on balancing quality and diversity in human-aligned diffusion models found that reward fine-tuning methods improve visual appeal while leading to diversity collapse through overfitting and reward hacking. The generator locates regions of output space the reward model scores generously and stays there.
This is not confined to one optimizer. Later work examining reinforcement learning methods for text-to-image fine-tuning reports the same reward and diversity tradeoff across PPO and LOOP variants, which suggests the pressure is structural to the setup rather than an artifact of a particular algorithm.
Because the narrowing is introduced during alignment, base model capability is the wrong lever. A larger and better-pretrained model has more range available to it, and the same preference-tuning pressure applied to that model still selects a narrow region of it. The mechanism is indifferent to how much capability sits underneath.
This explains a pattern teams report when they upgrade. A newer base model arrives, early unaligned samples look noticeably more varied, and the variety disappears once the model has been tuned to the same preference data on the same objective. The upgrade delivered range and the alignment stage spent it. Attributing the outcome to the model rather than the tuning leads to another upgrade with the same result.
The mitigation reported in the diversity work is consistent with this reading. The proposed strategy uses the base model for the early diffusion steps and switches to the reward-tuned model only for the final steps, which preserves the base model's variation in the composition-forming stage and applies the reward where it affects finish. The fix operates on where the reward is allowed to act rather than on model scale. The same paper reports that jointly training against multiple rewards, rather than one, achieved balanced performance across human preference, fairness, and compositionality.
If the compression starts at collection, that is where it can be prevented. Three properties matter.
Collect a score per criterion rather than one verdict per image, using the criteria your reviewers are already weighing implicitly. This keeps the signal separable, which means a reward model can be trained per criterion and combined deliberately rather than learning an average nobody specified. It is a change to the collection interface more than to the volume of work, and it is the difference between the pairwise collection pattern most teams start with and something a generator can learn range from.
Reviewer disagreement on subjective work is information about where legitimate variation lives. The designer preference research treats disagreement measured around a Krippendorff's alpha of 0.25 as brand-specific judgement rather than annotation error, and collapsing it to a consensus label discards exactly the variation you want the model to retain. Storing per-reviewer labels rather than an aggregate preserves the option to model the distribution instead of its mean, which matters because subjective quality is a data problem before it is a modeling one.
Given that optimizing appeal alone degraded a measured diversity metric, diversity has to be scored and constrained rather than assumed. That means choosing the dimensions that matter for your use case, measuring them on the base model to establish a starting point, and treating movement against that baseline as a regression during tuning. Evaluation has to be built for it, which is why evaluating text-to-image output systematically belongs in the loop, alongside the practice of scoring outputs that have no single right answer with a defined rubric.
If you are buying preference data rather than collecting it, four items belong in the specification. Name the criteria and require a score for each, rather than accepting a single preference verdict. Require per-reviewer labels to be delivered rather than aggregated. Specify reviewer qualification, since design judgement is the property you are paying for. State the diversity dimensions you will measure, so the sample is built to cover them rather than to be easy to label.
Those four turn a preference dataset from a source of one blended signal into something you can optimize against deliberately. Guidance on commissioning preference data for a design model covers the operational side of running that collection.
If your outputs have converged on one look, the constraint is usually the structure of the preference data rather than the model consuming it. HumanSignal Services runs preference collection with qualified designers, delivering per-criterion scores and per-reviewer labels built to a specification you set. Book a scoping conversation to talk through the criteria and coverage your case needs.
Both contribute, and the published measurements point at fine-tuning as the part teams can most readily address. In the diffusion work cited above, a base model showed better distributional spread than the same model after tuning against a single human-preference reward, so the narrowing was introduced during alignment rather than inherited from pretraining. Pretraining composition still matters, and it is harder to change after the fact.
Not on its own, because the mechanism is indifferent to base capability. A stronger model has more range available, and applying the same single-reward preference tuning still selects a narrow region of that range. What changes the outcome is how the reward is constructed and where in the generation process it is allowed to act.
Pick dimensions that matter for your use case and measure distributional spread across many generations rather than judging individual images. The work cited here used statistical parity on a demographic attribute, where 0 means a uniform distribution, and comparable measures exist for palette, composition, and subject variation. The important discipline is establishing the base model's figures first, so tuning can be evaluated against a real starting point.
The published evidence says yes, provided both are in the objective. Jointly training against multiple rewards achieved balanced performance across human preference, fairness, and compositionality in the diffusion work, and a separate result showed that restricting the reward-tuned model to the final diffusion steps preserved base-model variation. Both approaches share the property of not letting a single aesthetic reward govern the whole process.
Per-criterion scores rather than a single winner, per-reviewer labels rather than a consensus, stated reviewer qualification, and the diversity dimensions you intend to measure. The first two keep the signal separable enough for a reward model to be composed deliberately. The last one ensures the sample covers the variation you want preserved rather than only the cases that are easy to judge.
It makes the training signal richer at the cost of being harder to use naively. Averaging labels produces a cleaner scalar that has already discarded the variation you are trying to retain, so the cleanliness is bought with the information. Storing per-reviewer labels preserves the option of modeling the distribution, and you can always aggregate later, which is not true in reverse.