ASemConsist : Adaptive Semantic Feature Control for Training-Free Identity-Consistent Generation

Shin Seong Kim1*, Minjung Shin1*, Hyunin Cho1, Youngjung Uh† 1
1Yonsei University
*Equal contribution.
†Corresponding author.
arXiv Code (coming soon)

TL;DR: We enable FLUX to generate subject identity-consistent images across diverse text prompts without additional training.

Abstract

Recent text-to-image diffusion models have significantly improved visual quality and text alignment. However, generating a sequence of images while preserving consistent character identity across diverse scene descriptions remains challenging. Existing methods often struggle with a trade-off between maintaining identity consistency and ensuring per-image prompt alignment.

In this paper, we introduce a novel framework, ASemconsist, that addresses this challenge through selective text embedding modification, enabling explicit semantic control over character identity without sacrificing prompt alignment. Furthermore, based on our analysis of padding embeddings in FLUX, we propose a semantic control strategy that repurposes padding embeddings as semantic containers. Additionally, we introduce an adaptive feature-sharing strategy that automatically evaluates textual ambiguity and applies constraints only to the ambiguous identity prompt.

Finally, we propose a unified evaluation protocol, the Consistency Quality Score (CQS), which integrates identity preservation and per-image text alignment into a single comprehensive metric, explicitly capturing performance imbalances between the two metrics. Our framework achieves state-of-the-art performance, effectively overcoming prior trade-offs.

Comparison with existing methods

Qualitative results

Quantitative results

Ours achieves state-of-the-art performance on the CQShar, balancing per-image alignment and identity consistency. We highlight the best score in light red and the second-best in yellow.

Table

How dose it work?

Selective text-embedding modification

We selectively modify text embeddings to control identity and per-image semantics, amplifying identity-consistent features while suppressing irrelevant ones. Additionally, based on our analysis, we repurpose FLUX padding embeddings—dominated by dummy semantics—as semantic containers by injecting meaningful per-image information. These strategies enable stable identity preservation without sacrificing per-image prompt alignment.

Adaptive feature-sharing

We first automatically evaluate identity ambiguity by analyzing feature cohesion across transformer blocks and timesteps.

When ambiguity is high, we selectively apply residual feature sharing from a cached identity embedding. This adaptive mechanism reinforces identity consistency without sacrificing visual diversity or prompt fidelity.

Balanced Evaluation: Consistency Quality Score (CQS)

We propose the Consistency Quality Score (CQS), a unified metric that jointly evaluates the balance between identity preservation and per-image text alignment.

By measuring both criteria simultaneously within a single score, CQS enables a comprehensive and reliable ordering of identity-consistent generation methods.

Given an identity prompt pid and per-image prompts [p1, …, pk], we generate a set of images X = {x1, …, xk}, where each image xi reflects its per-image prompt p_i and identity prompt pid. We define two alignment scores:

Definition of t_i and a_i

where ti measures alignment with both the identity and per-image prompts, and ai measures alignment with the per-image prompt alone. Identity consistency across the generated images is quantified using DreamSim:

Definition of d_i

DreamSim scores are transformed using (1−·) and scaled via min–max normalization to match the range of VQA scores.

To explicitly account for the imbalance between identity consistency and per-image text alignment, we define the alignment gap Δi = ai − ti, which measures how much the per-image alignment deviates from the combined identity–prompt alignment.

We summarize the dataset-level tendency via mean positive and negative gaps,

Definition of Delta plus and minus

and define per-sample reward and penalty terms that interpolate between dataset-level statistics and instance-level deviations:

Definition of delta_i minus and plus

where λ ∈ [0,1] balances dataset-level and instance-level adjustments. The adjusted identity score is then given by

Definition of adjusted identity score d_i star

where μ, τ ≥ 0 control the strength of penalties and rewards.

The final per-sample score is computed as a harmonic mean:

Definition of CQS harmonic mean

By integrating identity consistency and per-image text alignment into a single metric and explicitly penalizing their imbalance, CQShar provides a balanced, comprehensive criterion for evaluating and ordering identity-consistent generation methods.

BibTeX

@article{shin2025asemconsist,
      title={ASemConsist: Adaptive Semantic Feature Control for Training-Free Identity-Consistent Generation},
      author={Shin, Minjung and Cho, Hyunin and Uh, Youngjung and others},
      journal={arXiv preprint arXiv:2512.23245},
      year={2025}
    }