The guidence_1 in the code repository?

[GpenAI28]

We are concerned about the authors' artificial introduction of guidance_1 into the text-to-image model. Specifically, the authors' code below shows that the first denoising step of the text-to-image model manually sets curr_guidance to 1.0 to additionally increase sample diversity.

guidance_1 = torch.full([1], 1.0, device=device, dtype=torch.float32).expand(1)

if guidance is not None and skip_first_cfg and i == 0:
curr_guidance = guidance_1
else:
curr_guidance = guidance

However, the base FLUX model sets guidance (curr_guidance) to 3.5 throughout the entire sample process, which is clearly unfair. In particular, the authors have not stated or explained this special configuration in the paper. We question the authors' experimental results and look forward to their response.

[GaParmar]

Hi,

Thank you for the interest in our paper. We will update the arxiv soon with additional details and new baselines that we have added.

Cheers!
Gaurav

[GpenAI28]

Hi,

Thank you for the interest in our paper. We will update the arxiv soon with additional details and new baselines that we have added.

Cheers! Gaurav

We note that the Shielded Diffusion paper reports a Vendi Score of 3.749 on SD3-M, whereas the score you reported for Shielded Diffusion on OpenReview is 3.144—significantly lower than the value in the original paper. Since Shielded Diffusion, Diversity Flow, and Interval Guidance have not released their text-to-image code, we are concerned about the validity of the results reported by the authors. Furthermore, the authors’ method is based on inference-time scaling, while these baselines are training-free. Were the comparisons conducted under the same NFE (Number of Function Evaluations)? We kindly ask the authors to clarify this point as soon as possible. If possible, we warmly encourage the authors to release the code for the baselines and evaluation—we would be very happy to help verify the results.

[GpenAI28]

Hi,

Thank you for the interest in our paper. We will update the arxiv soon with additional details and new baselines that we have added.

Cheers! Gaurav

In fact, we have conducted experiments ourselves. Under the standard FLUX setting—specifically with CFG guidance set to 3.5—it is impossible to achieve the qualitative diversity results shown in the authors' paper. This is because the authors' method essentially performs pruning for efficiency and does not alter the inherent diversity ceiling of FLUX. For instance, even when generating 100 images from the same prompt using the default FLUX pipeline, we cannot reproduce the level of diversity qualitatively demonstrated in the paper. However, such diversity can indeed be obtained by manually reducing the CFG guidance during early sampling steps. Since the authors’ code uses guidance_1 by default (which effectively lowers early-stage guidance), we believe the authors should clearly disclose this detail in their paper.

[GaParmar]

Shielded Diffusion reports a Vendi score of 3.749 for SD3-M on imagenet dataset and image resolution 256x256.
We use full 768x768 image resolution and a different dataset (GenEval). The difference in score might be because of this.
We are working on releasing the evaluation scripts and baselines.
The baselines shown in Figure 6 are indeed not doing inference time scaling as those methods are not designed for test time scaling. The baselines in Figure 8, on the other hand shows comparison with other methods that do inference scaling.
We are also actively working on updating the paper with the details.
Thanks again for all the detailed comments, we really appreciate it!

Regards,
Gaurav

[GpenAI28]

Shielded Diffusion reports a Vendi score of 3.749 for SD3-M on imagenet dataset and image resolution 256x256. We use full 768x768 image resolution and a different dataset (GenEval). The difference in score might be because of this. We are working on releasing the evaluation scripts and baselines. The baselines shown in Figure 6 are indeed not doing inference time scaling as those methods are not designed for test time scaling. The baselines in Figure 8, on the other hand shows comparison with other methods that do inference scaling. We are also actively working on updating the paper with the details. Thanks again for all the detailed comments, we really appreciate it!

Regards, Gaurav

Thank you for the clarification, which largely addresses our concerns. We greatly appreciate your willingness to publicly release the baseline and evaluation code, which is a significant contribution to research on diversity in diffusion models. We are eagerly looking forward to the code release, as fair comparison and evaluation are of great importance. Thank you again for your efforts, and we will continue to follow your work closely. We’re also wondering when the code might be released, because we’re very excited about it!