Diffusion Adversarial Post-Training for One-Step Video Generation

Video Comparisons

We would like to clarify that existing research on diffusion acceleration has primarily focused on the image domain. To the best of our knowledge, no prior studies have proposed high-resolution, one-step video generation methods, and consequently, no suitable baselines exist for comparison. Therefore, in our submission, we evaluate our method using the image generation task, for which comparable prior works are available.

We acknowledge the reviewer’s perspective. Following the reviewer’s suggestion, we have included a comparison with the consistency distillation baseline and our base model. The evaluation follows the public VBench protocol. Specifically, we generate 5 videos for each of VBench's 946 prompts and report the VBench scores in the table below.

As the table shows, APT significantly outperforms the baseline consistency distillation 1NFE, 2NFE, and 4NFE. We observe that the consistency 1NFE baseline generates lower-quality results (e.g., blurry videos), and these issues persist even when increasing to 4NFE. This result verify the effectiveness of APT post-training compared to the baseline.

Even when compared to our base model at 50NFE (25 steps + CFG)—despite this being an unfair comparison—our APT 1NFE achieves a comparable total score (82.00 vs. 82.15) and even performs better on some metrics.

We will add the table and the analysis in the revised paper.

Related Works

We appreciate the reviewer pointing out the related works. We will add all of them in the revised paper.

• SF-V and OSV are 1-step image-to-video generation. We initially did not include them because our work focuses on text-to-video generation. Both of these works are based on UNet based Stable Video Diffusion and can generate up to 1024x576 for 14 total frames. In comparison, our works can generate 1280x720 of total 48 frames (2s 24fps) from text prompts.

• PCM is an adversarial consistency distillation approach very similar to Hyper and our paper has conducted extensive comparison with Hyper. We notice that the paper has also experimented with distillation on AnimateDiff along with AnimateLCM and AnimateDiff-Lightning. We will add this paper to our related works.

• InstantPortrait focuses on 1-step image editing tasks with contributions primarily addressing task-specific problems, which may not be directly relevant to our problem.

Comparison Against Other Distillation Methods

The reviewer suggested a comparison against ADD, LADD, and UFOGen. In fact, both ADD and LADD are already included in our submitted paper. Specifically, the ADD method is referred to as SDXL-Turbo and the LADD method as SD3-Turbo, following the naming conventions used by Stability AI [1,2]. We apologize for any confusion and will clarify this in the revised version. We did not include UFOGen in our comparisons, as it was shown to performce worse than DMD2 in the DMD2 paper. Since our paper includes DMD2 as a baseline, we believe this provides a sufficient point of comparison.

VBench Results

We have provided VBench metrics for the rebuttal. Please refer to the rebuttal response to Reviewer afjC. These results will be included in the revised version of the paper.

Other Clarifications on Our Claims

About Ln42-Ln45: We have acknowledged in Ln70-Ln76 that UFOGen is closest to our work which only applies adversarial training on real data. We also elaborated our differences compared to UFOGen. Specifically, our discriminator design is closer to the traditional GAN, and our APT model can surpass the teacher model in some criteria which validates the proposed adversarial post-training.

About Ln56-Ln57: We have elaborated in Appendix B how the traditional COCO FID and PFID metrics in Table 5 and 6 do not fully capture the model performance and thus opt for human evaluation. We leave the exploration for better automated metrics to future works. Our claim is based on Table 1 and 4, where our APT model shows improved visual fidelity compared to the original diffusion model.

Comparison with Consistency Baseline

We respectfully point out that consistency distillation (CD, including methods such as LCM) has been extensively studied in prior works (e.g., DMD, DMD2, Lightning, Hyper-SD, LADD), which consistently show that CD struggles to produce sharp results in a single step. For instance, see Figure 3 in Lightning. This has also been shown in Figure 9 of our submission which illustrates that as training progresses, the initial CD model generates noticeably blurry outputs in one step. Therefore, in our work, we choose to compare directly against state-of-the-art methods rather than the CD baseline.

However, following the reviewer's suggestion, we have included the CD-initialized model as a baseline. We compute metrics: FID, PFID, and CLIP scores, on both the COCO-10K and COCO-5K benchmarks. As shown in the table below, the results are consistent with those previously reported in the paper. We will incorporate them in the revised paper.

In addition, we have included a video comparison between the CD and APT models on the VBench benchmark. Please refer to our response to Reviewer afjC for the detailed results.

More visualization comparison between CD and APT is provided in this external link.

Differences to Existing Methods

We would like to elaborate on the differences between our methods and existing methods (e.g. DMD CVPR24):

1. Simplicity: DMD combines variational score distillation (VSD) and rectified flow (RF) objectives, as VSD alone can lead to mode collapse while RF can cause sharpness issues. DMD2 changes to use both VSD and adversarial objectives, but this setup requires training 3 networks: the student, the VSD negative score model, and the adversarial discriminator. In contrast, our method adopts a simpler approach, using only an adversarial objective with CD initialization. This streamlined design allows the model to adapt quickly—in just 350 iterations (see Figure 9 in the paper).

2. Post-training vs. distillation quality: Existing methods (such as DMD, DMD2, Lightning, Hyper, ADD, and LADD) distill the teacher results, where the teacher model is the quality upper-bound. Our method is adversarial post-training (APT) against real data, and we demonstrate that it is able to surpass the teacher diffusion model in some criteria. The discussion is provided in Appendix H.

3. First large-scale video generation with 1NFE: Existing methods like DMD and LADD require precomputing teacher noise-sample pairs, which can be computationally expensive—especially for high-resolution video tasks. In contrast, our method trains directly on real data, avoiding these overheads. This enables us to be the first to demonstrate one-step generation using a large-scale T2V model, achieving 1280×720 resolution at 24 fps for 2-second videos.

4. Closer to traditional GANs: Existing adversarial distillation methods generally follow the DiffusionGAN approach, which may appear similar to GANs but differs in key ways: 1) noise corruption is added to the discriminator inputs (e.g., LADD, UFO-GAN, Lightning), and 2) it freezes the discriminator backbone and only trains the generator (e.g., ADD, LADD). In contrast, our method more closely resembles classical GANs: the discriminator is fully trainable, and no noise corruption is applied to its inputs. Additionally, we introduce an approximate R1 loss, inspired by traditional R1 regularization, which significantly improves training stability. We hypothesize that this design contributes to the reduction of visual artifacts in our one-step generation, as compared to existing works (LADD, Lightning, etc.), as illustrated in Figure 3.

We hope the above points clarify the differences from related work and substantiate the contributions of our method.