acceptodds
Withdrawn from ICLR 2027

DBFP: PLUG-AND-PLAY DUAL-BAND FUSION FIXED-POINT SAMPLING FOR SUBJECT-DRIVEN GENERATION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Subject-driven generation must preserve a reference subject’s identity while executing a text edit—two objectives that compete under a single sampling condition. Even strong backbones leave this tension unresolved: following the prompt more closely drifts identity, while clinging to the reference weakens instruction following. Evaluation compounds the difficulty by scoring the two axes separately, without a compact criterion for their joint balance. To this end, we present Dual-Band Fusion Fixed-Point (DBFP), a training-free plug-and-play sampler. Given only a reference image and an edit prompt , it builds a structure prompt that preserves subject attributes while absorbing salient edit keywords, pairs it with the original instruction, fuses the two guided velocities by early-stage frequency-band mixing in Fourier space, and refines the mixed step with a short Picard loop—without touching weights or embeddings. We further propose the Subject-Text Harmonic Score (STHS) as a compact measure of the coupled identity–edit balance. Experiments on multiple subject-driven generation backbones show that DBFP improves this balance and, under the same experimental protocol, yields stronger overall subject-driven generation performance as a dropin sampling plug-in.

Primary area:
generative models
Keywords:
subject-driven generation, training-free sampling, plug-and-play, flow matching, identity preservation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Nra1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the trade-off between identity preserving and text-guided editing in subject-driven image generation and proposes DBFP, a training-free sampling method. It first construct a structural prompt compatible with the target edit, fuses different frequency components of the structure and edit branches’ velocity fields during early sampling, and refines the updates through a small number of Picard iterations. The authors also introduce the Subject-Text Harmonic Score (STHS), which combine identity similarity and text–image similarity into a unified score.

Critical strengths

The paper proposes a training-free, plug-and-play framework that achieves some improvements across multiple backbone models.

Critical weaknesses
  1. Poor paper writing and unclear motivation for the method design. The paper is poorly written, particularly in the Introduction section, where the motivation behind the method is unclear. For example, why should the structure branch contribute more to low frequencies, while the edit branch contributes more to high frequencies? Identity information can also include high-frequency details such as textures and logos, whereas edits involving pose or viewpoint may require changes to low-frequency structure. Besides, The metric design is also insufficiently motivated. Why aggregate the two identity cosine similarities using a geometric mean rather than a simple product? Likewise, why combine the resulting subject-fidelity score with CLIP-T using a harmonic mean? The authors do not clear state the motivation for these design in main paper. Moreover, DreamBench++ [1] already ranks methods by the product of concept preservation (CP) and prompt following (PF) scores. The paper neither compares its aggregation scheme with this simpler product-based approach nor adequately discusses the distinction and additional contribution.

  2. Novelty and contribution are limited, which may not meet the bar for ICLR. In particular, the metric appears to offer only a minor contribution, given that prior work such as DreamBench++[1] already combines subject preservation and prompt following through a product-based score. The paper’s main contribution therefore rests primarily on the proposed sampling method, which has unclear motivation.

  3. Insufficient ablations to establish the independent contributions of the sampling components. Compared with the baseline, DBFP simultaneously introduces a structural prompt, dual-condition predictions, frequency-domain fusion, and Picard refinement. However, the existing ablations do not separately evaluate the contributions of prompt construction and frequency-domain fusion. Consequently, the independent contribution of the core sampling design remains insufficiently established.

[1] DreamBench++: A Human-Aligned Benchmark for Personalized Image Generation. ICLR 2025.

Questions for authors

Address the concerns in Critical Weaknesses section.

Reviewer_NbLeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies subject-driven generation (personalization) and addresses the trade-off between preserving subject identity and following a text edit. To address this, it proposes DBFP, a training-free, plug-and-play sampler. Given a reference image and an edit prompt, DBFP constructs a structure prompt, fuses the two guided velocities in Fourier space, and refines the mixed step with a short Picard loop. Experiments are conducted on a variety of backbones.

Critical strengths
  • Framing personalization as a frequency allocation problem between identity and editing is an interesting perspective.
  • The method is training-free and works in plug-and-play manner.
  • Evaluation includes many recent backbones.
Critical weaknesses
  • Although this is a personalization work, the method is not compared against any personalization methods (at least DreamBooth, Textual Inversion, and Custom Diffusion). References to well known personalization works that tackle the same problem are also missing (for example DisenBooth, Cones, TextBoost, ELITE, BLIP-Diffusion, DTI, and InstantBooth). The paper should include these references and compare against at least some of them.
  • Since the method is built on a frequency perspective, including FreeU as a comparison would strengthen the performance claim.
  • The method assumes that identity is low frequency and edits are high frequency. This fits the lighting and style edits that dominate the test set, but it does not fit structural edits such as pose or viewpoint, where the edit itself is low frequency. The paper handles these cases by adding pose words to the structure prompt, not by the frequency split. In the other direction, fine identity cues such as facial geometry are high frequency and are therefore taken from the edit branch, which is not discussed.
Questions for authors
  • How much computation and wall clock overhead does the method add at inference? The Picard loop in particular seems likely to add a considerable amount. A table reporting the wall clock time of each stage of the method would strengthen the paper's practicality.
  • How good is the caption quality from the VLM? It would be helpful if the authors reported the sensitivity of the captioning to the choice of VLM, as well as the effect of the rule based Align step.
Reviewer_icPuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DBFP, a training-free sampling framework for subject-driven image generation that aims to balance identity preservation and text-guided editing. It constructs two conditioning prompts, combines their predicted velocities through frequency-band mixing, and applies Picard iteration to refine the sampling trajectory. The authors additionally introduce STHS, a harmonic score combining subject fidelity and text alignment, and demonstrate improvements across eight pretrained generative backbones.

Critical strengths
  • The identity-preservation versus editability trade-off is an important and well-motivated problem in subject-driven image generation.
  • The proposed method is training-free and can be integrated into compatible pretrained generators without modifying their weights.
  • The paper reports consistent improvements across eight recent foundation models, including HiDream-O1, Qwen-Image-Edit, and FLUX.2, suggesting good transferability across different backbones.
Critical weaknesses

Despite the broad backbone coverage, I have two major concerns: (A) insufficient discussion of closely related work, which makes the novelty of some contributions unclear, and (B) insufficient experimental comparisons, which limit the evidence for the proposed method's effectiveness.

  1. Missing comparisons with existing methods. The experiments only compare DBFP against the official samplers of pretrained foundation models. They do not establish competitiveness with existing training-free subject-driven generation methods, such as FreeCus [1], LatentUnfold [2], Personalize Anything [3], and Diptych Prompting [4]. Comparisons with representative methods on a shared benchmark would strengthen the evaluation.
  2. Insufficient discussion of related work. The paper does not adequately differentiate its approach from Subject-Agnostic Guidance [5] and Decoupled Guidance [6], which also address the identity-preservation versus editability trade-off. Furthermore, RPO [7] introduces a weighted harmonic mean of reference-image fidelity and text fidelity, which substantially overlaps with the motivation behind STHS. The authors should clarify the novelty of STHS relative to this prior work. The discussion of evaluation protocols is also outdated, relying primarily on DreamBooth (2023) without considering more recent benchmarks and evaluation metrics, such as DreamBench++.
  3. Insufficient controlled experiments. The paper lacks comparisons against prompt enrichment alone, scalar dual-field guidance, and compute-matched sampling alternatives. These controls are important for determining whether the reported improvements arise from the proposed frequency-band fusion and Picard refinement, rather than additional conditioning information or inference computation.

[1] FreeCus: Free Lunch Subject-driven Customization in Diffusion Transformers, ICCV 2025

[2] Flux Already Knows – Activating Subject-Driven Image Generation without Training, arXiv 2025

[3] Personalize Anything for Free with Diffusion Transformer, AAAI 2026

[4] Large-Scale Text-to-Image Model with Inpainting is a Zero-Shot Subject-Driven Image Generator, CVPR 2025

[5] Improving Subject-Driven Image Synthesis with Subject-Agnostic Guidance, CVPR 2024

[6] Decoupled Guidance: Disentangling Subject and Context Pathways in Text-to-Image Personalization, arXiv 2026

[7] Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning, NeurIPS 2024

Questions for authors
  • Inconsistent notation for Picard iterations. In Section 4.4 and Figure 4, the normalized Picard residual equals 1 at (k=1) by definition, yet the authors state that residuals drop by (k=1). Meanwhile, Section 3.5 states that (K=1) corresponds to two evaluations of (\Phi). Could the authors clarify the relationship between (k) and (K), and explain how Figure 4 supports the choice of (K=1)?

  • Motivation for fixed-point refinement. The fixed-point equation appears to be a heuristic modification of the sampling dynamics rather than a requirement of forward ODE integration. In particular, evaluating the velocity at a predicted endpoint while keeping the timestep fixed does not necessarily improve numerical integration accuracy. Could the authors provide a stronger theoretical or empirical justification for this refinement, ideally including comparisons with standard higher-order ODE solvers under matched computational budgets?

  • Frequency bands and semantic disentanglement. Fourier linearity establishes that frequency-wise velocity mixing produces the corresponding state increment, but it does not establish that low-frequency components encode subject identity while high-frequency components encode editing instructions. Could the authors provide empirical evidence supporting this assumption, for example through frequency-band analysis, reversed-band assignments, or per-band interventions?

  • Reproducibility. Despite the implementation checklist, several important details remain unspecified, including the exact per-backbone sampling configurations, complete prompt-construction rules, evaluation split, and actual inference costs. Could the authors provide these details to enable independent reproduction?

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

Related papers

Discussion (0)

Sign in to comment.