CREward: A Type-Specific Creativity Reward Model

This paper introduces CREward, the first type-specific creativity reward model that evaluates and generates creative images across geometry, material, and texture axes by leveraging human benchmarks and large vision-language models to enable applications in assessment, explanation, and guided generation.

The Gist

Imagine you are an art teacher trying to grade a class of students who are drawing chairs. In the past, you might have just given them a single score like "8 out of 10" for "creativity." But that's a bit vague. Did they get an 8 because the chair looked weird? Because it was made of jelly? Or because it had a cool, swirly pattern?

This paper introduces CREward, a new tool that acts like a super-smart art teacher who doesn't just give a single grade. Instead, it breaks "creativity" down into three specific, understandable ingredients: Geometry (the shape), Material (what it's made of), and Texture (the surface details).

Here is how the paper explains this system, using simple analogies:

1. The Problem: Creativity is Too Complicated

The authors argue that treating creativity as one big, blurry blob is "naive." Just like a chef needs to know if a dish is too salty, too sweet, or too spicy to fix it, AI designers need to know exactly what makes an image creative. Is the chair creative because it has a weird shape (Geometry), or because it looks like it's made of glass (Material)?

2. The Solution: A Three-Lens Camera

The researchers built a model called CREward that looks at images through three specific lenses:

• Geometry: The overall shape and curves. (Think: Is the chair a perfect cube, or does it twist like a corkscrew?)
• Material: How light bounces off it. (Think: Is it shiny metal, soft velvet, or squishy jelly?)
• Texture: The fine details on the surface. (Think: Is it smooth, rough, or covered in polka dots?)

3. How They Taught the AI (The "Human vs. Robot" Test)

To teach CREward, the team first asked real human design experts to look at pairs of images and vote on which one was more creative for each of the three categories. This was their "gold standard."

Then, they asked a very smart AI (called an LVLM, or Large Vision-Language Model) to do the same voting. They found something surprising: The AI agreed with the humans almost perfectly. In fact, for some categories, the AI agreed with humans even better than the humans agreed with each other!

Because the AI was so good at mimicking human taste, the researchers used the AI's votes to train CREward. This is like hiring a robot assistant to grade thousands of homework assignments so the teacher doesn't have to do it manually.

4. What CREward Can Do (The "Magic Tools")

Once trained, CREward becomes a Swiss Army knife for creative AI:

• The Scorecard: It can look at any image and give it three separate scores (one for shape, one for material, one for texture). This helps researchers see which AI models are actually good at making creative shapes versus just making pretty textures.
• The Filter: Imagine you have a bucket of 1,000 AI-generated chairs. CREward can instantly sort them, pulling out the top 10 that have the wildest shapes or the most interesting materials. This helps human designers find inspiration quickly without scrolling through endless boring images.
• The Slider: This is the coolest part. The researchers attached "sliders" to the AI generator. If you slide the "Geometry" bar up, the AI starts making chairs with stranger shapes. If you slide "Material" up, it starts making chairs out of weird stuff like water or gold. It's like having a remote control for creativity.
• The X-Ray: Using a technique called Grad-CAM, CREward can highlight exactly where in the image the creativity is happening. If a chair is creative because of its twisted legs, the tool will highlight the legs, showing you exactly what the AI is focusing on.

5. The Catch (Limitations)

The authors are honest about the flaws.

• Quality vs. Weirdness: CREward is great at spotting "newness," but it might sometimes think a blurry, nonsensical image is creative just because it's weird. It needs a partner to check if the image actually looks good (quality) before celebrating its creativity.
• The "Spillover" Effect: Sometimes, when you try to make the shape more creative, the material accidentally changes too. The authors call this "entanglement." It's like trying to turn up the volume on a radio, but the bass gets louder too. They plan to fix this in the future.

Summary

In short, CREward is a new system that stops treating creativity as a mystery. It breaks it down into Shape, Stuff, and Surface, uses a smart AI to learn what humans find creative, and then gives us tools to measure, filter, and control exactly how creative our AI-generated art becomes.

Technical Summary

Technical Summary: CREward – A Type-Specific Creativity Reward Model

Problem Statement

Creativity in visual design is a complex phenomenon that is often treated as a single, undifferentiated quantity in computational assessments. Existing approaches to evaluating creative generation frequently rely on costly human evaluations or repurpose generic vision metrics (e.g., FID, CLIPScore) that are not designed to capture creative qualities. Furthermore, current methods lack the ability to define creativity in a precise, quantifiable, and explainable manner across specific dimensions. The paper argues that treating creativity as a monolithic notion offers limited guidance for generative models and fails to align with human perception, which hierarchically processes distinct visual attributes.

Methodology

1. Type-Specific Definition

The authors ground creativity in the image formation pipeline, decomposing it into three interpretable "axes" or types:

• Geometry: The overall shape, form, and arrangement of surfaces, features, and curves.
• Material: The medium defining surface interaction with light (reflection, absorption, transmission).
• Texture: The visual detail of a surface, including color, patterns, and markings.

2. Human Benchmark (CREBench-Human)

To establish a ground truth, the authors constructed a benchmark dataset involving five objects (chair, car, handbag, bowl, vase).

• Data Generation: 5,000 creative images were generated per object using various diffusion models, from which 20 diverse images and 5 "least creative" normal images were manually selected per object.
• Annotation: Five design experts performed pairwise comparisons on 100 randomly sampled image pairs. Annotators judged which image was more creative for specific types (Geometry, Material, Texture) and for "Overall" creativity.
• Metric: Winning rates were computed based on pairwise preferences, normalized by the number of appearances.

3. Human-LVLM Correlation Analysis

The study investigated whether Large Vision-Language Models (LVLMs) could align with human creativity judgments.

• Models Tested: Gemini-2.5-Flash (closed-source) and Gemma-3-27B-it (open-weight).
• Procedure: LVLMs were prompted with the same instructions used for human annotators to rank image pairs.
• Finding: Both models showed notable correlation with human judgments. Gemini-2.5 achieved the highest alignment (exceeding inter-human correlation in some cases), while Gemma-3 performed comparably or better than humans for specific types like material and texture. This validated the use of LVLMs as a scalable alternative for labeling.

4. CREward Training

Leveraging the alignment found in the previous step, the authors trained CREward, a type-specific creativity reward model.

• Data Synthesis: A large dataset of creative images was synthesized using five diverse Text-to-Image (T2I) models (Hunyuan-DiT, PixArt, Kandinsky v3, SD3.5-Large, Flux-schnell) guided by LLM-generated prompts targeting specific creativity types.
• Labeling: Gemma-3 was used to generate preference labels for 5,000 image pairs across the five objects.
• Architecture: A lightweight regressor consisting of a frozen vision backbone and a 5-layer MLP head. The model predicts scalar scores for Geometry, Material, Texture, and Overall creativity.
• Training Objective: The model was trained as a binary classifier on triplets $(x_A, x_B, y)$, where $y$ indicates preference. The loss function minimizes the negative log-likelihood of the pairwise logistic score.
• Backbone Selection: Among several tested backbones (VGG16, CLIP, DreamSim, DINOv3, SigLIP), SigLIP (specifically the Gemma-3 vision encoder) achieved the highest test preference accuracy and was selected as the backbone for CREward.

Key Results

Alignment with Human Perception

Despite being trained solely on LVLM-generated labels (Gemma-3), CREward achieved strong alignment with human judgments on the human benchmark:

• Rank Correlation: CREward achieved a Spearman rank correlation of 0.59 for Geometry, 0.72 for Material, and 0.76 for Texture with human rankings.
• Comparison: CREward outperformed the open-source Gemma-3 and the baseline Surprise metric. Notably, for Texture creativity, CREward surpassed the closed-source Gemini-2.5 in rank correlation.
• Inter-Human vs. Model: The model's alignment often exceeded the inter-human agreement (e.g., 0.72 for Material vs. 0.63 inter-human), suggesting LVLM-derived supervision can effectively model creativity.

Type Influence on Overall Creativity

Analysis of the benchmark data revealed that Geometry has the highest correlation with "Overall" creativity across human annotators, LVLMs, and CREward (correlation ~0.80–0.96), whereas Texture shows the weakest alignment. This aligns with findings in human visual perception where coarse structural information is processed before fine-grained details.

Model Performance Evaluation

• Generative Models: When applied to evaluate five representative diffusion models, Hunyuan-DiT attained the highest creativity scores across all types, while Flux-schnell yielded the lowest, consistent with the behavior of distillation-based models that prioritize fidelity over diversity.
• Methodologies: CREward successfully benchmarked creative generation methods like ConceptLab and C3, showing ConceptLab's strength in material/texture and C3's improvements in geometry.

Applications Demonstrated

The paper outlines three primary applications enabled by CREward:

1. Creativity Assessment: Providing type-specific scores to compare generative models and generation methodologies (e.g., ConceptLab vs. baselines).
2. Creative Sample Acquisition: Filtering large pools of generations to surface the most creative examples for specific types. This was demonstrated in a human-AI co-creation workflow where an industrial designer used top-ranked AI samples for inspiration, reducing the search burden.
3. Guided Creative Generation: Using CREward scores to train LoRA-based sliders. These sliders allow users to selectively amplify Geometry, Material, or Texture creativity during the denoising process of a diffusion model. The paper notes that while some entanglement between sliders exists, they effectively guide generations toward desired types.
4. Explainable Creativity: Integrating CREward with XAI techniques like Grad-CAM to visualize which image regions contribute to specific creativity scores, offering interpretable insights into the generation process.

Significance and Claims

The authors claim this work as the first type-specific creativity reward model that spans geometry, material, and texture.

• Formalization: It formalizes creativity along three specific, interpretable dimensions grounded in 3D rendering, moving away from treating creativity as a monolithic quantity.
• Scalability: By demonstrating that LVLMs can align with human creativity judgments, the paper proposes a scalable alternative to costly human labeling for training reward models.
• Controllability: CREward enables controllable creative generation through LoRA sliders, allowing targeted manipulation of creative properties rather than relying on random or bias-prone outputs.
• Modesty: The authors acknowledge limitations, including the model's emphasis on novelty over value (potentially scoring semantically weak images highly if they are novel) and the inherent entanglement between creativity types (e.g., changing geometry may inadvertently affect texture). They position CREward as a "useful starting point" and an "imperfect metric" to inspire future work in measurable and interpretable visual creativity.