acceptodds
Under review as a conference paper at ICLR 2027

PP-CFG: Polarity-Preserving Guidance for Controllable Few-Step Image and Video Generation

Abstract

Few-step diffusion distillation reduces sampling cost but often weakens adjustable classifier-free guidance (CFG) and runtime negative prompting. We introduce PP-CFG, an architecture that keeps positive and negative text conditions separate and combines their attention responses with opposite signs within a shared visual backbone. Its Scale-Time Polarity Mixer refines each condition using the guidance scale and denoising timestep, allowing prompt influence to vary throughout generation. The student retains a continuous scale input and selectable negative prompts with one visual backbone pass per step. We evaluate PP-CFG on Wan2.1 for video generation with distribution matching distillation (DMD), and on Qwen-Image with both latent consistency model (LCM) distillation and DMD. PP-CFG achieves 81.46 VBench in four steps and lower video teacher–student distances across guidance scales than step-distilled DMD. On NegElement-1K, our benchmark of 1,000 matched image and video prompt cases, it achieves higher target-removal rates than the corresponding teachers. These experiments support the method's generality across attention architectures, generation modalities, and distillation objectives while retaining competitive generation quality.

Primary area:
generative models
Keywords:
Classifier-free guidance, diffusion distillation, few-step generation, negative prompting, image and video generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jho2OpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes a new method for CFG distillation by using a mixer to encode guidance scale and timesteps and using attention manipulation for negative guidance.

Critical strengths
  • The paper proposes a new and effective method for distilling few-step models.
  • The proposed method considers both image and video models with different architectures.
  • The math behind the paper supports the method well.
  • The human evaluation supports the first claim about distillation
Critical weaknesses
  1. The negative guidance control deviates from how negative prompts are intended to be used. A negative prompt is usually not directly conflicting with the positive prompt. For example, one might ask for a positive prompt: house negative prompt: windows . However, in the examples shown in Figure 7, the positive prompt directly mentions the negative element: the positive prompt asks for a wooden table with sunglasses, while the negative prompt asks for no sunglasses. This raises the question if sunglasses do not appear, is the model following the negative prompt or just failed to follow the positive one? If the author can provide more experiments on a small set of non-conflicting prompts (like the NegGenBench used in VSF [1] or DCS-Bench from [2]. Some examples from [2]: ("A Christmas tree", "gift boxes") ("A home office desk at night", "desk lamp"). The author could also try a few de-biased examples like (“a machine learning researcher”, “male”).
  2. For the guidance part, the authors compared enough baseline traditional distillation methods like DMD and LCM. However, for the negative guidance part, the authors did not compare with training-free methods like NASA (which is what the PP-CFG is built on top of), VSF, and NAG.
  3. Following the last one, the novelty of this paper seems limited. It is just NASA with two condition transformations trained with LCM or DMD loss.
  4. The authors did not specify what distillation datasets were used besides “internal-Image-10k”. With no information about it, it is unclear if it overlaps with the testing datasets.

[1] https://arxiv.org/abs/2508.10931
[2] https://arxiv.org/pdf/2605.29390
[3] https://arxiv.org/abs/2505.21179

Questions for authors

Could the author provide more information about the training dataset?

Reviewer_51n6OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks if we make diffusion models much faster while retaining control over both what to generate and what to avoid. To do so, the authors keep positive and negative prompts separate inside the model, adapt their representations to the requested guidance strength and generation stage, and combine their effects within one shared visual computation. The project has a useful architectural idea and credible engineering, but its evaluation is stronger on generation quality than on its main claim: better controllability. It does not yet establish superiority over the most relevant state of the art.

Critical strengths

The problem the authors takle is important and their solution is rational and well conceived. Testing Wan and Qwen-Image, cross-attention and joint attention, and LCM and DMD is valuable.

Critical weaknesses

The evaluation do not support the claims and the contributions of the idea. The novelty over existing guidance components is not sufficiently isolated, and the negative-prompt benchmark does not establish that improved target absence represents better controllability.

  1. Competitors: The relevant competitors are not just fast generators such as DMD, LCM, or AVD. They include methods specifically designed to restore negative prompting in fast generators. The authors already cites these candidates: NASA, NAG, VSF, PCM.

  2. The fact that target absence is treated as removal success, do not consider that the scene may be degraded to much and the positive part of the prompt discarded entirely. I would also include ordinary negative-prompt cases where the unwanted element is not explicitly requested in the positive prompt. Otherwise, the benchmark primarily tests conflict resolution rather than the full range of everyday negative prompting.

  3. Section 4.4 reports target-absence rates across models but does not compare each model with the target negative prompt versus an empty negative prompt while holding the positive prompt, initial noise, guidance scale, and sampling configuration fixed. Consequently, higher target absence cannot be attributed specifically to negative prompting: it may also reflect a greater tendency to omit the target independently of negative conditioning.

  4. ablations: Table 3 evaluates removal of timestep conditioning, removal of the mixer, and concatenation instead of addition of scale/time tokens using VBench Total, Quality, and Semantic scores. These results support the full mixer’s generation quality under the evaluated setting, but do not establish how the ablated components affect guidance-scale responsiveness or negative-prompt effectiveness. The scale sweep and target-removal evaluation are not repeated for these variants, leaving their contributions to the two claimed control capabilities unresolved.

Questions for authors

On the same backbone and comparable distilled checkpoint, does PP-CFG outperform NAG, VSF, and NASA-style subtraction in suppression, preservation, quality, and latency?

Reviewer_xCwEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces PP-CFG, a method that preserves adjustable classifier-free guidance (CFG) and runtime negative prompting in few-step diffusion models while maintaining a shared visual backbone. A Scale-Time Polarity Mixer adapts positive and negative text conditions according to the guidance scale and denoising timestep, while polarity blocks subtract negative-prompt attention responses from positive-prompt responses.

Critical strengths
  • This paper addresses the loss of adjustable guidance and runtime negative prompting in few-step diffusion models, which is an important limitation for controllable generation.
  • This paper integrates positive and negative prompt pathways within a shared visual backbone, enabling both controls with one visual backbone pass per denoising step efficiently.
  • The Scale-Time Polarity Mixer adapts text conditions based on guidance scale and denoising timestep, allowing prompt influence to vary throughout generation.
  • This paper Introduces NegElement-1K, a benchmark of 1,000 matched image and video prompt cases designed to evaluate negative-prompt-based target removal.
Critical weaknesses
  • The Qwen models are trained with empty negative prompts. However, they are evaluated with target-specific negative prompts. It is unclear how the model learns effective negative-prompt control under this setup.
  • Controlled evaluation is insufficient. The negative-prompt experiments compare PP-CFG with the original teacher but lack a within-model ablation to isolate the proposed polarity mechanism's contribution.
  • The validation of the benchmark is limited. NegElement-1K relies on template-generated prompts and a single VLM judge, raising concerns about generalization to realistic scenarios and evaluation reliability.
  • The benchmark mainly measures target removal. However, it does not evaluate whether other desired scene elements are preserved, potentially overstating the effectiveness of negative prompting. In short, content-preservation evaluation is limited. -Also, comparison with related methods is insufficient. It lacks direct comparisons with attention-based negative-guidance methods, such as NASA and NAG, and the ablations do not isolate the contribution of polarity subtraction.
  • Finally, empirical validation in the user study is limited. The user study includes only 10 participants and a small number of prompts, while the FVD analysis uses rank-deficient covariance estimates without regularization, which may weaken the reliability of the quantitative conclusions.
Questions for authors

Please refer to the weaknesses.

Comments on the paper

1 comment

I am impressed by the author's detailed rebuttal, and my concerns are resolved. I changed my score from 2 to 4. The NegGenBench results solved my concern, and the new comparison with the training-free method is valid. I would not want to take points off for novelty concerns, given that the comparison with the base is added. The details of the dataset are also informative. I am also impressed by the rebuttal the authors had for other reviewers.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.