Intermediate Text Representation Guided Text-to-Image Generation for Enhancing One-and-Only Alignment

Conference: ECCV 2026
arXiv: 2606.30262
Project Page: https://soyoun-won.github.io/one-and-only-ir-guidance/
Code: None
Area: Diffusion Models
Keywords: Text-to-Image Generation, Concept Association Bias, Intermediate Representation Guidance, Information Theory, Training-free Method

TL;DR

This paper demonstrates through information-theoretic analysis that the intermediate layers of text encoders retain more conceptual mutual information than the final embeddings. It proposes IR-guided diffusion, which injects intermediate hidden states into the conditioning signals during the early stages of denoising to recover attributes suppressed by strong priors. Requiring no training or external models, it improves VQAScore by up to 19.1 percentage points on OAO-AttackBench while maintaining generation quality and human preference.

Background & Motivation

Text-to-image (T2I) diffusion models have made remarkable progress in generating realistic images. However, when the text prompt conflicts with strong visual priors learned from the training data, models often ignore explicit textual instructions, defaulting to generating the typical appearance stored in their "memory." This paper terms this phenomenon concept association bias: models bind certain visual attributes so tightly to specific concepts that explicit overrides in the text are suppressed. Although this bias is not inherently harmful—precisely enabling models to generate rich images from sparse descriptions—it poses a fundamental limitation when users require creative or counterfactual generation.

Concept association bias is most severe in "One-and-Only" (OAO) entities—those that possess a single unique, canonical form, such as the Earth, the Mona Lisa, or the Eiffel Tower. For example, given the prompt "Earth is a purple planet," SD3.0 consistently fails to override its learned blue-green Earth prior, either completely ignoring the color attribute or failing to correctly associate purple with Earth.

The core insight of this paper stems from a key observation: simple concepts emerge in early layers of the text encoder, but certain concepts are lost in the final embedding. The authors formalize this intuition through an information-theoretic analysis: the text encoder forms a Markov chain \(Y \rightarrow c_h \rightarrow c_f\). By the Data Processing Inequality, it follows that \(I(Y; c_h) \geq I(Y; c_f)\), meaning that the mutual information between the intermediate hidden states and the input text is no less than that of the final embedding. Combined with existing findings that "object-level concepts emerge in early layers and remain stable afterwards," it can be further derived that \(I(b; x_t^h \mid A) \geq I(b; x_t^f \mid A)\)—conditioning on intermediate representations enhances attribute-specific mutual information. Based on this, they propose IR-guided diffusion, an entirely inference-time method that requires no training, optimization, or external models.

Core Idea: Injecting the hidden states from the intermediate layers of the text encoder as a "supplementary signal" into the denoising conditioning. This leverages word-level conceptual information retained in the intermediate layers but compressed in the final embedding, recovering the text-image alignment suppressed by strong priors without altering the model architecture.

Method

Overall Architecture

The core problem addressed by IR-guided diffusion is that when generating OAO entities, T2I models fail to override learned strong visual priors with explicit attribute specifications in the prompt due to concept association bias. The overall approach of the method is: extract the hidden state \(c_h\) from the intermediate layer of the text encoder, integrate it via additive injection with the final embedding \(c_f\) to form the IR-guided embedding \(\tilde{c} = c_f + \lambda c_h\), use \(\tilde{c}\) as a conditioning signal during the early denoising stage to recover suppressed attribute information, and automatically switch back to \(c_f\) using an adaptive scheduler once the image structure stabilizes to refine the details. The entire process runs purely at inference time, without involving any training, fine-tuning, or external models.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Text Prompt Y"] --> B["Text Encoder Forward Pass"]
    B --> C["Extract Intermediate Representation c_h<br/>(k-th Layer Hidden State)"]
    B --> D["Final Embedding c_f<br/>(L-th Layer Output)"]
    C --> E["Post-processing: LayerNorm(c_h)<br/>Distribution Alignment, Retains Word-level Distinctiveness"]
    D --> F["Construct IR-guided Embedding<br/>c̃ = c_f + λ·c_h"]
    E --> F
    F --> G["Early Denoising Steps t > t*<br/>Condition on c̃, Inject Suppressed Concepts"]
    D --> H["Late Denoising Steps t ≤ t*<br/>Condition on c_f, Refine Details"]
    G --> I["Adaptive Scheduler<br/>Monitor ΔD_t to Find Switch Point t*"]
    I --> H
    H --> J["Generate Image"]

Key Designs

1. Information-Theoretic Analysis: Why Intermediate Layers Retain More Concept Information

To address the fundamental question of "why concepts are discarded by the final embedding," the authors provide a rigorous formalization from an information-theoretic perspective. Given a text prompt \(Y = \{A, b\}\) (where \(A\) is an OAO object and \(b\) is a specified attribute), concept loss is characterized by the near-indistinguishability in distribution between images generated conditioned on \((A, b)\) versus those conditioned on \(A\) alone, i.e., \(D_{\mathrm{KL}}(p_\theta(x_t \mid A, b) \;\|\; p_\theta(x_t \mid A)) \approx 0\), which is equivalent to the conditional mutual information \(I_\theta(x_t; b \mid A) \approx 0\).

The text encoder processes prompt \(Y\) through a series of transformer blocks, each being a deterministic function of its input, thus forming a Markov chain \(Y \rightarrow c_h \rightarrow c_f\), where \(c_f = h(c_h)\). From the Data Processing Inequality (DPI), it directly follows that:

\[I(Y; c_h) \geq I(Y; c_f)\]

Subsequent processing layers can only preserve or discard information, and cannot create it out of thin air—the amount of information in the intermediate representation is strictly no less than that in the final embedding. Furthermore, previous studies [toker2024diffusion] show that object-level concepts emerge in early layers and remain stable thereafter, i.e., \(I(A; x_t^h) \approx I(A; x_t^f)\). Using the chain rule of mutual information \(I(Y; x_t) = I(A; x_t) + I(b; x_t \mid A)\) combined with the two preceding properties yields the core conclusion:

\[I(b; x_t^h \mid A) \geq I(b; x_t^f \mid A)\]

This means that conditioning on the intermediate representation can enhance "attribute-specific" mutual information—forming the theoretical foundation of the proposed method's effectiveness. The supplementary material provides the full proof of Lemma 1 (developed based on conditional independence and the chain rule). This analysis also explains why they choose to extract \(c_h\) right after the 1st transformer block (\(k=1\)): layer-wise and token-wise cosine similarity analysis reveals that early layers retain word-level distinctiveness, while deeper layers gradually converge to sentence-level semantics, and OAO alignment requires precisely word-level signals.

2. IR-guided Embedding Construction and Injection: Integrating Intermediate Representations into Denoising

To address the puzzle of "how to effectively integrate intermediate representations into the denoising process," the authors design a simple additive injection scheme. The core operation is adding the intermediate representation as a supplementary signal to the final embedding:

\[\tilde{c} = c_f + \lambda c_h\]

where \(\lambda \in [0, 1]\) controls the injection strength (fixed to \(\lambda = 0.2\) in experiments). Prior to injection, the final LayerNorm of the text encoder is applied to \(c_h\) with the goal of aligning the distributions of \(c_h\) and \(c_f\). Experimental analysis shows that LayerNorm reduces the Wasserstein distance (WD) from 0.20 to 0.05 while preserving token-wise fluctuation in cosine similarity (in contrast, Mean-Std alignment, despite reducing WD to 0.03, flattens token-wise similarity—smoothing out word-level distinctiveness). This is precisely why LayerNorm is chosen as the post-processing step: OAO alignment requires word-level signals, and only LayerNorm balances distribution matching and word-level information preservation.

For multi-encoder architectures (such as SD3.0 which has three text encoders: OpenCLIP ViT-G, CLIP ViT-L, and T5-XXL), intermediate representations are extracted independently from each encoder, assembled according to the native concatenation scheme, and then added to the concatenated final embedding:

\[\tilde{c} = \bigl[c_f^{(1)} \;\|\; \cdots \;\|\; c_f^{(N)}\bigr] + \lambda \bigl[c_h^{(1)} \;\|\; \cdots \;\|\; c_h^{(N)}\bigr]\]

where \([\cdot \|\cdot]\) denotes concatenation, and the superscripts index different encoders. This preserves the native conditioning interface of each architecture without requiring any structural modifications.

The authors also systematically compare four alternative post-processing strategies: (1) None—directly replacing \(c_f\) with \(c_h\), introducing severe distribution mismatch; (2) Mean-Std—normalizing \(c_h\) to the sample-wise statistics of \(c_f\), resulting in loss of word-level distinctiveness; (3) Projection—projecting \(c_h\) onto the direction of \(c_f\), which retains only the aligned component; (4) Residual—using the orthogonal residual after projection, but discarding the stable semantics in \(c_f\). Experiments show that additive injection consistently outperforms or matches these alternative schemes across all metrics.

3. Adaptive Scheduling: When to Stop IR Guidance

Although the intermediate representation carries richer information, using \(\tilde{c}\) throughout the entire denoising process introduces visual artifacts. Diffusion models determine global structures in early steps and refine details in later steps, meaning IR guidance is only needed during the structural formation phase. The core challenge is: how to automatically determine the switch point \(t^*\) without relying on hand-crafted hyperparameters? The optimal number of guidance steps varies across different datasets and architectures (ablation opens show that SD2.1 performs best with k=5 on OAO-AttackBench, but with k=1 on Whoops), making manual presets ungeneralizable.

The authors design an adaptive scheduler based on structural acceleration. First, the clean image is estimated at each timestep using the Tweedie formula:

\[\hat{x}_{0|t} = \frac{x_t - \sqrt{1-\bar{\alpha}_t}\,\epsilon_\theta(x_t, t)}{\sqrt{\bar{\alpha}_t}}\]

Next, the amount of structural change between consecutive timestep estimates is computed as \(D_t = \|\hat{x}_{0|t} - \hat{x}_{0|t-1}\|_2^2\), and its first-order difference is defined as the structural acceleration:

\[\Delta D_t = D_t - D_{t-1}\]

\(\Delta D_t > 0\) indicates that the structure is still changing rapidly, whereas \(\Delta D_t \leq 0\) indicates that the layout has begun to stabilize. The switch point is defined as the first timestep where the acceleration becomes non-positive:

\[t^* = \min\bigl\{t \;\big|\; \Delta D_t \leq 0\bigr\}\]

The complete denoising process is:

\[x_{t-1} = \begin{cases} \text{denoise}(x_t, \tilde{c}) & \text{if } t > t^*,\\[2.0pt] \text{denoise}(x_t, c_f) & \text{if } t \leq t^*. \end{cases}\]

Ablation studies confirm that adaptive scheduling outperforms or matches the optimal fixed scheduling across all data-architecture combinations, avoiding manual searches for optimal steps in each setting. The computational overhead of calculating \(D_t\) is negligible compared to denoising itself (supplementary results show a single-A100 inference time increase from 4.80s to 5.07s, ~5.6% increase, with peak VRAM remaining completely unchanged).

An Illustrative Example

Take "Octagon-shaped Earth" as an illustrative example to walk through the process. The text prompt \(Y = \{\text{Earth}, \text{octagon-shaped}\}\) is fed into three text encoders of SD3.0.

Forward Pass & Extraction: Each encoder simultaneously outputs the final embeddings \(c_f^{(1)}, c_f^{(2)}, c_f^{(3)}\) and the 1st-layer hidden states \(c_h^{(1)}, c_h^{(2)}, c_h^{(3)}\). In \(c_f\), the association between Earth and "spherical" is extremely strong—the "octagon" attribute information was gradually compressed through the layered abstraction of the 24-layer CLIP ViT-G encoder. However, token-wise cosine similarity analysis in \(c_h\) reveals that the "octagon" token still retains significant distinctiveness compared to other tokens.

Embedding Construction: The corresponding encoder's final LayerNorm is applied to each \(c_h^{(i)}\) to align its distribution with \(c_f^{(i)}\) (reducing WD from ~0.2 to ~0.05). They are then concatenated in a native layout to construct \(\tilde{c} = [c_f^{(1)}\|c_f^{(2)}\|c_f^{(3)}] + 0.2 \cdot [c_h^{(1)}\|c_h^{(2)}\|c_h^{(3)}]\).

Denoising & Scheduling: Starting from pure noise \(x_T\), the first few steps are conditioned on \(\tilde{c}\). The adaptive scheduler calculates \(\Delta D_t\) in real time: in early denoising, \(\Delta D_t\) remains continuously positive (global shape—spherical vs. octagonal—is evolving rapidly). Around 30%-50% of the total steps, \(\Delta D_t\) becomes non-positive for the first time (overall layout is locked, and the Earth's shape is determined), marking \(t^*\). After this, the system switches back to the standard \(c_f\) condition to complete detail generation such as continental textures and ocean colors. In the final image, the Earth presents an octagonal outline instead of the default spherical shape—IR guidance successfully injects the "octagon-shaped" attribute during the structural formation phase, while \(c_f\) ensures surface features are preserved during later steps. Notably, analysis shows that the baseline's estimated image at the very beginning of denoising (T-6) is already close to the final output (prematurely converging to the spherical prior), whereas the IR-guided estimated image at the same stage has more noise and a completely different structure—IR guidance effectively delays premature structural convergence, allowing room for attribute exploration.

Loss & Training

IR-guided diffusion is an entirely training-free method. It has no loss functions, no gradient updates, and no learnable parameters; all operations are executed at inference time. There are only two key hyperparameters: (1) intermediate layer extraction position \(k=1\) (all experiments extract right after the 1st transformer block; ablation shows early layers achieve higher VQAScores, consistent with theoretical predictions that word-level information is richest in early layers); (2) injection strength \(\lambda = 0.2\) (ablation shows performance is stable within \(\lambda \in [0.1, 0.5]\), whereas excessively large \(\lambda\) introduces distribution shift and degrades performance). The computational overhead is minimal: on an A100, SD3.0 native inference takes 4.80s per image, while IR-guided inference takes 5.07s (an increase of ~5.6%), with peak VRAM remaining completely unchanged (18.76 GB).

Key Experimental Results

Main Results

The table below summarizes the core results of IR-guidance on four benchmarks. The substantial gains in VQAScore reflect improved fine-grained attribute alignment, while the mild changes in CLIPScore indicate that global semantic quality remains stable.

Benchmark	Backbone	Metric	Baseline	IR-guidance	Gain
OAO-AttackBench	SD3.0	VQA	0.625	0.707	+0.082
OAO-AttackBench	SD3.0	CLIP	0.260	0.272	+0.012
OAO-AttackBench	SD3.0	HPSv2.1	0.263	0.271	+0.008
OAO-AttackBench (Landmarks-Material)	SD3.0	VQA	0.525	0.716	+0.191
OAO-AttackBench	SD2.1	VQA	0.596	0.621	+0.025
Whoops	SD3.0	VQA	0.778	0.802	+0.024
Whoops	SD3.0	CLIP	0.291	0.296	+0.005
Whoops	SD3.0	HPSv2.1	0.295	0.299	+0.004
Gecko(R)	SD3.0	VQA	0.866	0.871	+0.005
Gecko(S)	SD3.0	VQA	0.744	0.748	+0.004

Ablation Study

Configuration	VQAScore (OAO-AttackBench, SD3.0)	Description
Full model (Additive injection + LayerNorm + Adaptive scheduling, λ=0.2)	0.707	Full method
w/o Adaptive scheduling, fixed k=5	~0.695	Adaptive scheduling outperforms all fixed-step sizes; optimal fixed value varies with dataset/architecture
w/o Adaptive scheduling, fixed k=10	~0.688	k=1 is optimal on Whoops, different from OAO—no universally optimal fixed value
Injection strength λ=0.1	~0.705	λ∈[0.1, 0.5] consistently outperforms baseline (0.625); the method is insensitive to λ
Injection strength λ=1.0	~0.685	Excessive injection leads to distribution shift, degrading performance but still outperforming baseline
Post-processing replaced by None (using \(c_h\) directly)	0.509 (SD2.1)	Severe distribution mismatch, leading to the largest drop
Post-processing replaced by Mean-Std	0.617 (SD2.1)	Word-level distinctiveness is smoothed, resulting in significant drop
Post-processing replaced by Projection	0.627 (SD2.1)	Retains only components aligned with \(c_f\), discarding orthogonal info
Post-processing replaced by Residual	0.617 (SD2.1)	Discards the stable semantics of \(c_f\)
w/o IR-guidance (Baseline)	0.625	Only uses \(c_f\); concept association bias remains unresolved

Key Findings

• VQAScore gains are much larger than CLIPScore: This indicates that IR guidance primarily improves fine-grained attribute binding rather than global semantics. The Landmarks-Material subcategory (e.g., "Machu Picchu made of clay") shows the largest improvement (+19.1 pp), as these material attributes conflict most severely with the canonical shape of the OAO entity, causing the worst information compression. The stability of the CLIPScore also indicates that the method does not degrade global semantic quality.
• Early layers (k=1) perform best and diminish layer-by-layer: On OAO-AttackBench, the VQAScore extracted from the 1st layer is the highest and monotonically decreases as depth increases. This aligns with the theoretical prediction that "word-level information resides in early layers and is compressed into sentence-level semantics in deep layers." On Whoops, the inter-layer variation is smaller, reflecting that the prior rigidity of everyday objects is less severe than that of OAO entities.
• IR guidance delays premature structural convergence: Analysis shows that the baseline's estimated \(\hat{x}_0\) at the very early denoising stage (T-6) is already close to the final output (spherical Earth), whereas the IR-guided structure at the same stage is significantly different and noisier. The mean squared difference curve reveals that the IR-guided denoising exhibits two local maxima after the transition point, demonstrating that the denoising trajectory has been readjusted—IR guidance grants space for attribute exploration, preventing the model from locking prematurely onto learned prior structures.
• KID and human preference remain unaffected: IR-guidance on the three OAO-AttackBench categories achieves KID scores comparable to or slightly better than the baseline (SD3.0 Celestial drops from 0.0576 to 0.0483, Landmarks from 0.0583 to 0.0500). HPSv2/HPSv2.1 ratings improve slightly on both OAO-AttackBench and Whoops, demonstrating that tighter text alignment simultaneously yields better visual aesthetics.
• Robust performance on Gecko proves generalizability: On the general compositional benchmarks Gecko(R) and Gecko(S), IR-guidance slightly outperforms the baseline across all four metrics (CLIP/VQA/HPSv2/HPSv2.1). This demonstrates that the method does not degrade quality on standard prompts but is mildly beneficial—even in the absence of explicit concept conflict, the extra information from intermediate representations still fosters more faithful text alignment.

Highlights & Insights

• A complete logical chain closed by information theory and empirical validation: From formalizing concept loss (KL divergence \(\approx\) 0) to mutual information decomposition (\(I(Y;x_t)=I(A;x_t)+I(b;x_t|A)\)), and deriving the information superiority of intermediate representations using the Data Processing Inequality (\(I(b;x_t^h|A) \geq I(b;x_t^f|A)\)), the theoretical claims are thoroughly validated across four benchmarks. This is far more elegant and convincing than purely empirical methods—each design choice has a firm theoretical backing.
• "Cost-free" inference-time method design: Requiring no training, fine-tuning, optimization, external models, or predefined bias sets, it simply leverages the existing intermediate representations inside the model, with a meager ~5.6% computational overhead and zero VRAM increase. This paradigm of "mining discarded signals within the model" is transferable to other encoder-decoder architectures: any generative model using an encoder can inspect whether its intermediate layers retain information lost in the final output and recover it using a similar injection strategy.
• Adaptive scheduling as a general denoising monitoring component: Estimating the clean image via the Tweedie formula and monitoring structural acceleration is, in essence, an online indicator for detecting whether the denoising trajectory is starting to converge. This design is not limited to T2I but is applicable to any inference-time intervention scenario for diffusion models—such as adaptive scaling for classifier-free guidance, or choosing the optimal intervention step for concept editing. It can easily serve as a general-purpose plug-and-play module.
• OAO-AttackBench fills an evaluation gap: Existing benchmarks (Whoops / Gecko / T2I-CompBench) focus on compositionality or common-sense violations of everyday objects, whereas OAO-AttackBench specifically target the core visual identity of entities that have "one and only" canonical form (spanning three types: celestial bodies, landmarks, and artworks)—representing the most extreme test bed for concept association bias. The dataset of 504 counterfactual prompts (meticulously categorized by shape, material, pattern, style, and genre) is systematically designed, highly extensible, and ready to serve as a standard testing platform for concept alignment studies.

Limitations & Future Work

• Context overflow issue: The authors admit that IR guidance occasionally propagates visual features of the target attribute to the background (e.g., in "Big Ben made of clay", the background might display clay textures too) rather than restricting it strictly to the target object. This is because the objective of the method is "text alignment" rather than "object editing"—ensuring the global image is faithful to the text without a reference image. Pinpoint object-level editing belongs to the domain of text-to-image editing; marrying IR guidance with editing tools (e.g., initiating structural layout with IR guidance, then applying localized editing) stands as an appealing future direction.
• Limitation inherited from the backbone model's capacity ceiling: IR guidance cannot transcend the cognitive boundaries of the backbone model. For example, on complex human pose generation, even if IR guidance improves text alignment, the anatomical inaccuracies intrinsic to SD3.0 still persist. This imposes a "ceiling effect"—if the backbone model fundamentally lacks representation for a concept, the intermediate layers will not contain the relevant signals either.
• Inapplicability to closed-source models: The paper tests Nano Banana Pro (Gemini series) on OAO-AttackBench and finds that even frontier closed-source models still suffer from concept association bias (VQAScore of 0.819, but failing on celestial bodies shape overrides and pattern-object association). However, IR guidance requires white-box access to intermediate layer representations, which cannot be directly applied to black-box APIs. Exploring how to adapt this core idea to black-box configurations (e.g., simulating intermediate layer injection via specialized prompt engineering techniques or leveraging API output logit/likelihood patterns to infer when a concept is being suppressed) represents a valuable research direction.
• Simplifying the injection mechanism: The additive injection \(\tilde{c} = c_f + \lambda c_h\) is direct and effective, but theoretically, \(c_h - c_f\) (the discrepancy between the intermediate and final representations) directly represents the lossy compressed information. Injecting this discrepancy rather than the raw intermediate representation might be more efficient and controllable. Furthermore, practicing injection in the cross-attention map dimension rather than the embedding space is also highly worth investigating.

Related Work & Insights

• vs. Diffusion Lens [toker2024diffusion]: Diffusion Lens discovered that early layers contain simple concepts that fail to materialize in generated images. However, it stayed at the stage of passive observation without offering actionable solutions. This paper directly bridges the gap by translating the observation into a practical inference-time intervention and theoretically justifying "why early layers are better" (holding larger word-level mutual information), representing a crucial leap from "observation" to "utilization."
• vs. Attend-and-Excite [chefer2023attend] (and similar methods requiring priors): These approaches mandate specifying "which prompt tokens require boosted attention" at inference time (which demands prior knowledge of which concepts are being neglected). In contrast, the current method requires no predefined bias sets or token-level priors, recovering all suppressed concepts automatically—since the intermediate layers natively retain all compressed information, eliminating the need to manually identify "what was lost."
• vs. Kim et al. [kim2025text]: Both methods leverage internal representations of text encoders to improve alignment, but Kim et al. requires optimization for aligning cross-attention maps and self-attention maps, which inflates computational complexity and remains confined to single-encoder architectures. Under our method, the inference is training-free and fully compatible with SD3.0's multi-encoder architecture, hence broader in scope and substantially more efficient.
• vs. Projection-Space Editing [orgad2023editing]: Eliminates specific association biases through projection edits but depends on pre-identified "relations to edit." This work, focusing on "recovering ignored concepts," is highly complementary to "removing unwanted associations." One could first employ projection-space editing to purge social biases and then invoke IR guidance to restore suppressed descriptions—allowing the two approaches to play distinct but synergistic roles.

Rating

• Novelty: ⭐⭐⭐⭐ Translates the known phenomenon "intermediate representations are more informative than final embeddings" into a practical inference-time guidance method. While the singular concept is not earth-shattering, the systematic combination of information-theoretic formalization, multi-encoder compatibility, adaptive scheduling, and a clear focus on OAO entities forms a complete and revealing framework. The "structural acceleration" design in the adaptive scheduler is particularly elegant.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Four benchmarks (including the newly proposed OAO-AttackBench), two backbones (SD2.1/SD3.0), seven categories of metrics (VQA, CLIP, KID, HPSv2, HPSv2.1, CHAIR, structure-graph evaluation), and complete ablations (layer choice, injection strength, scheduling strategy, four post-processing types, compared with fixed scheduling, and computational overhead). The experimental design is rigorous and highly comprehensive.
• Writing Quality: ⭐⭐⭐⭐ The arguments in the information theory section are clear and come with full proofs (including the Lemma 1 derivation in the supplementary material). The method description is well-structured with complete formulations. However, some quantitative results are scattered across database tables (Table 1-4 + Table S4, etc.); a more consolidated layout would improve readability.
• Value: ⭐⭐⭐⭐ Possesses direct utility for T2I applications requiring precise text alignment (creative design, educational visualization, counterfactual reasoning, etc.). The training-free nature facilitates simple integration into existing pipelines—any inference service deploying Stable Diffusion can adopt it with zero training cost. OAO-AttackBench represents a new standard testing arena for concept association bias research. The primary bottleneck lies in the dependency on white-box access, yielding no direct benefits to closed-source API users yet.

Related Papers

• [CVPR 2026] Extending One-Step Image Generation from Class Labels to Text via Discriminative Text Representation
• [ICML 2026] Alignment-Guided Score Matching for Text-to-Image Alignment in Diffusion Models
• [ECCV 2026] The Illusion of High Utility in Safety Alignment of Text-to-Image Diffusion Models
• [CVPR 2025] A Comprehensive Study of Decoder-Only LLMs for Text-to-Image Generation
• [AAAI 2026] Unleashing the Potential of Large Language Models for Text-to-Image Generation through Autoregressive Representation Alignment