Questions Regarding Benchmark Comparisons, FakeClue Dataset Quality, and Request for FakeVLM Model Weights

[JoyHanson-34]

Dear Siwei Wen,

I have a few questions regarding your work, and I would appreciate your insights:

1. Have you compared your method with more widely recognized benchmarks or training-based methods, such as the GenImage test set or AIGC Benchmark?

2. In Table 2 of the paper, the results show significantly better performance than GPT-4O, with FakeVLM also performing far better than models like Qwen2-VL, DeepSeek, and IntelVL. The last three LMMS are labeled using FakeClue, which raises some concerns about the quality of the FakeClue dataset. I examined the FakeClue test set and encountered severe quality issues. For example, for images from the Chameleon dataset like:

   • chameleon/fake/762482d5-d381-4586-8964-c3d567e233cf.jpg
   • chameleon/fake/f1159a40-3475-4ffc-bd87-7ae9b73e925a.jpg
   • chameleon/fake/939a1cf4-ba98-4570-abc7-a7a3f23c6b08.jpg
   • chameleon/fake/71ec1b22-945b-4a6c-bfb7-36cb4e70bfa4.jpg
   • chameleon/fake/99cd97e4-6c8f-4fe9-a80a-9b8d5f0971e8.jpg
   • chameleon/fake/8f05fe0f-3b17-442b-9b8b-cde419445110.jpg

The explanations provided for these images offer very little information, and despite varying image contents, the reasoning is often generic: "Despite its seemingly authentic appearance, certain features, such as disproportionate textures or odd lighting, hint that this image was generated by AI." For example, for the image chameleon/fake/492b2e59-5da3-4218-8ff0-ea6283c8bae3.jpg, the explanation merely repeats the content description, which clearly seems like a stitched-together response from multiple LMMs. In contrast, for images from the GenImage dataset, the explanations are overly verbose, sometimes exceeding 1500 characters for a single image. This leads me to a new concern: Given the questionable quality of the annotations, is it meaningful to calculate ROUGE_L and CSS scores under such conditions?

3. Finally, I would like to request the release of the FakeVLM model weights to help clarify these issues. Due to limited computational resources, I might not be able to fully replicate the original training accuracy.

Thank you for your attention to these matters, and I look forward to your responses.

Best regards
P.S. Thank you for open-sourcing such an interesting piece of work!

[lingcco]

1. Your suggestion is highly valuable. We will incorporate more benchmarks and training-based methods for comparative analysis in future work.
2. Regarding the Chameleon dataset, since the fake images in this dataset are often difficult for even humans to detect obvious artifacts, the phenomenon you mentioned is actually the result of our specifically designed prompting strategy for Chameleon to ensure accuracy. Concerning the dataset quality, we initially attempted human annotation but found that different annotators could provide completely different judgments on the same image, especially regarding the presence of artifacts. Moreover, the high cost of human labeling made large-scale expansion infeasible. Therefore, we developed a labeling pipeline based on multiple large models to construct the dataset. Naturally, using large models for labeling can introduce some sample-specific anomalies due to autoregressive language models characteristics, and we are actively exploring further optimization strategies.
3. We may take some time to release the weights. Now, our data is all open source. If you need to use them, you can get the weights by training yourself.
   Once again, we sincerely appreciate your valuable feedback, which provides important guidance for improving our future work.

[JoyHanson-34]

Thank you very much for your detailed reply!

After reading your explanation, I still have a few concerns that I would like to discuss further:

First, you mentioned in the paper that you are "using a multi-LMM labeling strategy based on category priors." However, the LMMs used for labeling — Qwen2-VL, InternVL, and DeepSeek — perform very poorly on FakeClue, with accuracy close to random guessing.
This raises a critical question for me: If these LMMs themselves cannot reliably determine whether an image is real or fake, can the explanations they provide for their judgments be considered trustworthy?
In my opinion, the "reasoning" generated by these LMMs might not stem from a real understanding of why an image is fake, but rather be the result of mechanical responses based on the prompt.

Second, regarding dataset quality, in the GenImage dataset, the explanations are often excessively lengthy and largely repetitive content descriptions. Meanwhile, some non-LMM-based methods have already achieved over 98.5% accuracy on GenImage, making it a relatively easy dataset.
Thus, I believe that the true advantage of LMM-based methods would be better demonstrated on more challenging datasets like Chameleon. However, the current quality of the Chameleon annotations is somewhat disappointing 😭.

Finally, from a practical application perspective, when dealing with simple datasets like GenImage, human evaluators generally do not even need reasons because the AI-generated traces are so obvious. Only when facing difficult cases would humans truly require reasoning assistance.

Thank you again for your valuable work and for taking the time to reply.
I look forward to hearing your thoughts again. Wishing you all the best in your work and life!