smolvlm2_anti_aesthetics_7b

Position: Universal Aesthetic Alignment Narrows Artistic Expression

This repository contains the model weights presented in the paper Position: Universal Aesthetic Alignment Narrows Artistic Expression.

The model is trained on the VisionReward dataset to explore the bias introduced by universal aesthetic alignment in generative models. This specific checkpoint is a 7B parameter version optimized for faster speed.

• Project Page: Anti-aesthetics Website
• Code: GitHub Repository (AAS)

Sample Usage

import torch
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from PIL import Image

# Load the model and processor
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "weathon/smolvlm2_anti_aesthetics_7b", dtype="auto", device_map="cuda"
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-4B-Instruct")

# Example rating logic
def rate_single_image(image, dimension, guideline):
    messages = [
        { 
            "role": "user",
            "content": [
                {
                    "type":"text",
                    "text": f"Please rate this image for its {dimension} quality. Use this guideline {guideline}. Response a single number.",
                },
                {
                    "type": "image", 
                    "image": image.resize((512, 512))
                }
            ],
        }
    ]

    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
        padding=True
    ).to(model.device)

    id_of_interest = processor.tokenizer.convert_tokens_to_ids(["0", "1", "2"])
    
    with torch.no_grad():
        logits = model(**inputs).logits
    
    # Extract probabilities for the rating tokens
    logits = logits[:, -1, id_of_interest] 
    prob = torch.softmax(logits, dim=-1)
    
    # Calculate weighted score
    score = torch.dot(prob[0], torch.tensor([0, 1, 2], device=prob.device).bfloat16())
    return float(score)

Citation

@article{guo2025aesthetic,
  title={Aesthetic Alignment Risks Assimilation: How Image Generation and Reward Models Reinforce Beauty Bias and Ideological "Censorship"},
  author={Guo, Wenqi Marshall and Qian, Qingyun and Hasan, Khalad and Du, Shan},
  journal={arXiv preprint arXiv:2512.11883},
  year={2025}
}