Consistent Story Generation: Unlocking the Potential of Zigzag Sampling

W1. Qualitative results fail to clearly demonstrate consistency advantages over baselines: For instance, in the image generated in Figure 4, the appearance and shape proportions of the sheep people also change significantly; The color and appearance of the horse have changed significantly. Annotate key subject traits to highlight consistent preservation and contrast with baselines’ inconsistencies. If instability occurs, analyze causes (e.g., sampling noise) and propose mitigations.

[Response]
We thank the reviewer for this critical observation. We agree that while our method significantly improves consistency, the inherent stochasticity of diffusion models can still lead to minor variations. The primary goal of our qualitative examples was to show a superior balance between subject consistency and prompt fidelity compared to baselines, which often fail dramatically on one of these two axes.
However, we recognize that the presentation of these results can be substantially improved. To make our method's advantages clearer, we will implement the reviewer's excellent suggestion.
Annotated Figures: We will revise Figure 4 (and Figures 6, 8, 9) to include annotations with call-out boxes and arrows. These annotations will explicitly highlight the key subject traits that are consistently preserved by our method (e.g., the satyr's specific horn shape and reed pipe, the centaur's bow and torso markings) and contrast them with the clear inconsistencies in the baseline results (e.g., changes in subject identity, failure to include key objects).
Analysis of Instability: We will add a sentence to the qualitative comparison section acknowledging that minor variations can still occur. We will attribute this to the exploration phase of the zigzag sampling, which is essential for generating diverse scene compositions, and frame it as a deliberate trade-off between rigid identity replication and creative diversity. This will better contextualize the results.

W2. Computational overhead lacks quantitative analysis: Decomposing each generation step into three sub-steps introduces additional computation, but the paper only mentions "computational cost" without metrics (e.g., inference time per image). This limits evaluation of its applicability to real-time scenarios.

[Response]
Inference time comparison: (seconds per image).
SDXL[50steps] 7.44s , Ours [50 steps] 21.33s.
FLUX[28steps] 1.50s , Ours [28 steps] 5.5 s

W3. Architectural adaptation analysis is insufficient: While AZS integrates more smoothly with FLUX than SDXL, the paper does not analyze why. For instance, can the feature injection strategy of SDXL be further optimized? This may limit the universality demonstration of the method on different models

[Response]
This is an insightful question, and the architectural differences between FLUX and SDXL likely contribute to FLUX's better compatibility with our method.
FLUX's Architectural Advantages: FLUX is a transformer-based model that utilizes a unified self-attention mechanism with modality-specific projection layers. This design allows FLUX to fuse semantic information across visual and textual modalities without relying on explicit cross-attention, leveraging the transformer's strength in modeling complex dependencies. In our method, this means that the newly injected subject visual tokens, text tokens, and current generated visual tokens are jointly considered within FLUX's unified attention layers. This simultaneous interaction likely enables FLUX to better balance subject consistency and prompt following, leading to improved performance.
SDXL's Architectural Considerations: In contrast, SDXL is a U-Net-based model that combines convolutional and transformer blocks. Its transformer blocks contain separate cross-attention layers for text information and self-attention layers for visual information flow between patches. When we inject subject visual tokens into SDXL's self-attention layers, the interaction between these subject tokens and the text tokens (handled by separate cross-attention layers) might be less integrated. This separation could make it more challenging for SDXL to optimally balance text following and subject consistency, as the model needs to reconcile information from distinct interaction pathways.

We believe there is still space to further improve the feature injection strategy in SDXL, and this is a promising direction for future work. Based on our analysis, we have identified several potential optimization strategies:

Prompt-Aware Dynamic Injection: Currently, we cache a static set of top-k visual tokens representing the subject's identity. A more advanced strategy would be to dynamically select or weight which subject tokens to inject based on the current scene's prompt. For example, for a prompt like "a dog running in a field," tokens corresponding to the dog's body and legs would be prioritized. For "a close-up portrait of a dog," face and fur texture tokens would be more relevant. This would create a more direct and semantic link between the textual guidance and the visual injection.
Spatio-Temporal Adaptive Injection: Our current method injects features into the mid and upper layers of the U-Net during each zig step. This could be optimized by making the injection adaptive:
- Spatially (Layer-wise): Inject high-level structural features (e.g., face shape) into deeper layers of the U-Net and more fine-grained texture details (e.g., fur pattern) into earlier layers.
- Temporally (Step-wise): Apply stronger visual injection during the initial diffusion timesteps to firmly establish the subject's identity, and then gradually reduce its strength in later steps to give more control to the text prompt for refining details and scene composition.
Cross-Attention Modulation: A more sophisticated approach would be to create a more direct bridge between the visual identity and text guidance pathways. Instead of only injecting tokens into the self-attention layers, the cached subject features could be used to modulate the output of the cross-attention layers. This would guide the model to interpret the text prompt in the context of the desired subject's appearance, potentially resolving the architectural separation more effectively.
We are excited by these possibilities and thank the reviewer for prompting this valuable discussion. We will add a paragraph outlining these future research directions to the conclusion or limitations section of our final manuscript to acknowledge this potential for improvement and guide future work in this area.

W4.Typography issues in Figures 8 and 9: Uneven letter spacing in left-side text causes readability problems, requiring formatting corrections.

[Response]
We thank the reviewer for spotting this formatting error. This was an oversight in the script used to generate the figures. We will regenerate Figures 8 and 9 for the final manuscript, ensuring that all text is formatted with proper letter spacing and is fully legible.

W1.1 Fig. 3 does not contain the labels zig,zag, and generation in the for sub-images.

[Response]
We will update Figure 3(c) to explicitly label the "zig," "zag," and "generation" sub-steps in the timeline, making the diagram a much clearer illustration of our method as described in the text.

W1.2 In Sec.3.2 the zig, zag and generation steps are defined multiple times, each time emphasizing a different aspect.

[Response]
In Sec.3.2 We first introduce the foundational zigzag sampling mechanism, outlining the general role each step plays (e.g., zig for exploration, zag for refinement, generation for final output). Following this, we explain our motivation for introducing an asymmetric information flow within this pipeline to better balance subject consistency and prompt following. We will revise our current writing to improve clarity.

W1.3 Inconsistent Abbreviations

[Response]
Thank you for pointing out this oversight. Consistency is key for clarity. We will explicitly define "SSC (Subject Semantic Control)" upon its first mention and ensure it is used consistently throughout the paper and in Figure 3(d). The different methods in Table 3 follows below definition (line 279-284):
"Asymmetry" (Our Method): Visual injection (ZVS) only in zig step. Text guidance (APZI with full prompt) in zig and generation steps. Null prompt in zag step. "Zig-gen": Visual injection (ZVS) in zig and generation steps. Text guidance (full prompt) in zig and generation steps. Null prompt in zag step. "Zig-zag": Visual injection (ZVS) in zig and zag steps. Text guidance (full prompt) in zig and generation steps. Null prompt in zag step. "All" (Fully Symmetric): Visual injection (ZVS) in zig, zag, and generation steps. Text guidance (full prompt) in zig and generation steps. Null prompt in zag step. We will clarify this distinction in the revised paper, making it explicit that these ablations examine the impact of where visual conditioning is applied within our established asymmetric zigzag framework, not variations of the zigzag process itself.

W2 Missing Implementation Details.

[Response]
Number of Subject Tokens (k): As shown in our ablation study (Table 4, page 9), we experimented with different values for the hyperparameter k. We found that setting k=0.2 (i.e., caching the top 20% of visual tokens based on attention scores) achieves the best balance between subject consistency and text alignment. We will highlight this specific value in our main Implementation Details section (Appendix C) to make it more prominent.

Inference Time Overhead: Inference time comparison: (seconds per image). SDXL[50steps] 7.44s , Ours [50 steps] 21.33s. FLUX[28steps] 1.50s , Ours [28 steps] 5.5 s.

Making code publicly available: We will include a commitment in the paper to release our source code and detailed instructions upon publication to facilitate reproducibility.

W3 Discussion of Failure Cases.

[Response]
This is an excellent suggestion that will add valuable nuance to our paper. We will perform a more detailed analysis of these cases. We will add a new subsection in the Appendix titled "Analysis of Failure Modes and Limitations," which will be referenced from the main text. In this section, we will discuss:
The Faun (Fig. 6): We will analyze how semantic changes in prompts (e.g., from an active "dancing" pose to a passive "resting" one) can sometimes lead to the model dropping secondary attributes (like the wings), which might not have been encoded as a core part of the subject's identity.
The Rainbow (Fig. 6): We will discuss how our method successfully preserves the core identity ("sticker of a vibrant rainbow"), but can show minor inconsistencies in secondary elements (clouds) which is not specifically mentioned in the prompt. This is a known challenge in generative models, and we will frame our discussion in this context.
This analysis will provide a more balanced perspective on the method's strengths and current boundaries.

Minor Note on Figure 4 (Centaur):
Thank you for pointing this out. We acknowledge that the difficulty in generating specific composite subjects, such as centaurs, appears to be an inherent limitation of the underlying text-to-image base models themselves, as we observed consistent challenges even when using the vanilla base model without our proposed method. We will revise Figure 4 to include a more representative example where the base model performs adequately, and we will add a short discussion in the limitations section or an appendix to address these inherent shortcomings of the base models in generating certain complex or hybrid subjects. This will clarify that such generation failures are not a specific limitation of our method but rather a challenge inherited from the foundational models.

Q1. In which layers is the zigzag pattern applied?

[Response]
This is detailed in Appendix C (line 344-353) and will be made clearer. For SDXL, visual token caching and injection are applied only in the mid and upper layers of the U-Net. For FLUX, we cache tokens from the early text-image interaction layers and inject them into all subsequent layers.

Q2. How much longer does it take compared to simple text-guided diffusion?

[Response]
See above responses to W2.

Q3. How many subject-specific tokens are used?

[Response]
See above responses to W2.

The authors, thank you for your elaboration and the clarifications. I will keep my slightly positive score.

W1.While computational overhead is acknowledged, deeper analysis of inference time vs. baselines would strengthen practicality claims.

[Response]
Inference time comparison: (seconds per image).
SDXL[50steps] 7.44s , Ours [50 steps] 21.33s.
FLUX[28steps] 1.50s , Ours [28 steps] 5.5 s

W2. The role of the "zag" step (inverse denoising) could be elaborated further—why is a null prompt optimal here?.

[Response]
The "zig" step is designed for exploration and initial injection of strong subject-specific visual information. The "zag" step, performing inverse denoising, acts as a refinement and propagation phase for the latent representation. Our intuition is that after injecting visual identity cues in the zig step, applying a strong textual prompt or further visual conditioning in the zag step would cause immediate overfitting to the scene's textual description or over-constrain the subject's appearance too early, potentially disrupting the delicate balance needed for flexible storytelling. A null prompt in the zag step allows the visually grounded subject information to propagate through the denoising process without being pulled strongly by any specific text or new visual cues. This enables the subject identity to "settle" into the latent representation in a more generalized manner, preparing it for the final "generation" step which then focuses purely on aligning with the detailed scene prompt. This prevents the model from overly focusing on the current scene's textual details for subject definition, thus promoting better consistency across different scenes while preserving diversity.
We conduct an additional ablation study to empirically support the optimality of the null prompt in the zag step, as also suggested by the other reviewer.
Asymmetry + Zag-Text": Our "Asymmetry" method, but with the full textual prompt also applied during the zag step, instead of a null prompt.
Asymmetry + zag-text: clip-I 0.9098, clip-T 0.8841, DreamSim 0.2251
Asymmetry + zag-null (ours): clip-I 0.923, clip-T 0.8946, DreamSim 0.1789
The results above show that adding text in zag step leads to significant decreases in subject consistent evaluation metrics including clip-I and DreamSim scores. This demonstrates that the null prompt is indeed optimal for preserving a consistent subject identity across the story. We will include these results and discussion in the revised paper.

W3. The zigzag concept builds on prior work (e.g., [51]); the paper could better delineate its unique adaptation for asymmetry.

[Response]
We agree. While [51] (Golden Noise for Diffusion Models) explores the functional roles of zig (exploration), zag (refinement), and generation steps in general, our work uniquely adapts and leverages this structure for asymmetric information flow aimed specifically at the subject consistency challenge in visual storytelling. We will explicitly delineate our unique adaptation in the revised "Related Work" and "Method" sections:
Prior Work ([51]): Introduced the zigzag decomposition and identified the general roles of zig (exploration), zag (refinement), and generation.
Our Unique Adaptation: We capitalize on these identified roles by proposing a novel, asymmetric strategy for when and how to inject different types of semantic information:

1. Zig (Exploration): Dedicated to initial, strong subject identity grounding via visual token injection (ZVS). This sets the consistent subject foundation.
2. Zag (Refinement/Propagation): Utilizes a null prompt (APZI) to allow the subject identity to propagate and refine in the latent space without being constrained by scene-specific text or further visual conditioning, preventing overfitting.
3. Generation (Final Alignment): Focuses solely on text alignment with the full narrative prompt, relying on the already-consistent subject from previous steps.

This precise, purposeful asymmetry of information injection (visual vs. textual, and null vs. full prompt) at specific stages of the zigzag process is our distinct contribution, directly addressing the core problem of balancing consistency and fidelity, which [51] did not explore.

Q1. The method adds sub-steps per generation. Could the authors quantify the runtime overhead compared to baselines (e.g., steps/second)?

[Response]
See response to W1.

Q2. What is the intuition behind using a null prompt here? Does empirical evidence support this over other designs (e.g., weakened prompt)?

[Response]
See response to W2.

Q3. The top-k ablation (Table 4) shows minor metric variations. Are there failure cases where k significantly impacts output?

[Response]
While Table 4 indicates that our method is robust to the choice of 'k' within the tested range (0.2 to 0.8) and variations are minor, we can elaborate on potential edge cases and the rationale for our optimal 'k' selection:

Extremely Low 'k' (e.g., <0.1): If 'k' is too small, only a very limited number of visual tokens would be selected, potentially leading to insufficient subject information being injected. This could manifest as a degradation in subject consistency, as the model might not have enough cues to maintain the identity across different scenes.

Extremely High 'k' (e.g., >0.9): Conversely, if 'k' is too high, it would mean injecting a very large proportion of visual tokens, including those that might be irrelevant or background-related. This could lead to an "over-conditioning" effect, similar to what is observed in our "fully symmetric" sampling ablation, resulting in visual artifacts or reduced generative diversity. It could also force the model to overemphasize the visual input, potentially hindering its ability to follow the textual prompt faithfully (lower clip-T).

Empirical Sweet Spot: Our chosen value of k=0.2 represents an empirical sweet spot that achieves the best balance between subject consistency and text alignment. This suggests that selecting a moderate number of the most relevant subject tokens is key, rather than an exhaustive or minimal set. We will clarify this in the revised paper to provide more comprehensive insights into the implications of 'k' selection.

We will present qualitative examples illustrating these potential subtle impacts for extreme k values in the Appendix to provide a more comprehensive understanding beyond just the quantitative metrics.

Q4. The paper notes occasional SDXL failures. Are there architectural insights explaining FLUX’s better compatibility?

[Response] This is an insightful question, and the architectural differences between FLUX and SDXL likely contribute to FLUX's better compatibility with our method.

FLUX's Architectural Advantages: FLUX is a transformer-based model that utilizes a unified self-attention mechanism with modality-specific projection layers. This design allows FLUX to fuse semantic information across visual and textual modalities without relying on explicit cross-attention, leveraging the transformer's strength in modeling complex dependencies. In our method, this means that the newly injected subject visual tokens, text tokens, and current generated visual tokens are jointly considered within FLUX's unified attention layers. This simultaneous interaction likely enables FLUX to better balance subject consistency and prompt following, leading to improved performance.

SDXL's Architectural Considerations: In contrast, SDXL is a U-Net-based model that combines convolutional and transformer blocks. Its transformer blocks contain separate cross-attention layers for text information and self-attention layers for visual information flow between patches. When we inject subject visual tokens into SDXL's self-attention layers, the interaction between these subject tokens and the text tokens (handled by separate cross-attention layers) might be less integrated. This separation could make it more challenging for SDXL to optimally balance text alignment and subject consistency, as the model needs to reconcile information from distinct interaction pathways.

General Model Scale and Performance: Beyond architectural design, FLUX is generally a larger and more capable model for text-to-image generation compared to SDXL, which has been demonstrated in various experiments. The transformer architecture, in general, has proven more effective when scaling model size and dataset, contributing to FLUX's overall superior performance in text-to-image generation.

L1. Expand on potential misuse and mitigation strategies.

[Response] We acknowledge the importance of discussing potential misuse, especially for generative models. We briefly touch upon the positive societal impacts in Appendix A. We will expand on the "Broader Impacts" section (Appendix A) to include:

Potential Misuse: Explicitly mention the risk of generating misleading or harmful content, such as deepfake stories or perpetuating stereotypes, given the enhanced consistency capabilities. Mitigation Strategies: While our work focuses on foundational research, we can suggest general mitigation strategies applicable to text-to-image models. These include: Ethical Guidelines for Use, Watermarking/Provenance,Responsible Deployment

L2. Could the method handle longer story sequences.

[Response].
We propose using a sliding window technique for the prompt lists. This approach allows the model to maintain context and subject consistency across a large number of images by processing prompts in overlapping segments. We conducted experiments using this solution, and the results demonstrate that our method can effectively generate consistent subjects throughout long stories and also maintaining strong text alignment.
We will include these detailed results and a discussion of the sliding window technique in the revised version of our paper.

Thanks for raising all your concerns, below are our responses:

W1. The idea of incorporating inversion process during sampling cannot be considered new, for example as in [1]. Actually, it can be considered as a special case of test time scaling [2], with a different guidance.

[Response].
We appreciate you highlighting related works on zigzag sampling and test-time scaling. We agree that zigzag sampling, as a concept for decomposing denoising steps, has been explored (e.g., [1]). Our novelty lies not merely in incorporating an inversion process, but in the asymmetric design of our Zigzag Sampling with Asymmetric Prompts and Visual Sharing (AZS) and its specific application to the challenging problem of maintaining subject consistency across multiple images in visual storytelling while preserving text fidelity.

While "test-time scaling" [2] refers to a broader category of techniques that adjust model behavior during inference, our AZS is a concrete sampling strategy that orchestrates the flow of specific semantic information (visual subject cues and textual prompts) in an asymmetric manner across the zig, zag, and generation sub-steps. Previous zigzag methods primarily focused on improving general generation quality or text following. In contrast, our key innovation is the strategic and asymmetric injection of subject-specific visual information only in the "zig" (exploration) step, using a null prompt in the "zag" (refinement) step to prevent overfitting, and leveraging the full textual prompt in the "generation" step for narrative coherence. This precise control over information flow within the zigzag framework is what distinguishes our method and allows us to achieve a superior balance between subject consistency and prompt fidelity, which is a fundamental requirement for visual storytelling that prior methods struggle with. We will elaborate on this distinction in the revised introduction and method sections.

W2. The main framework heavily relies on 1Prompt1Story, which lessens the impact of the paper itself. Can this sampling strategy be applied to other consistency generation frameworks?

[Response]
We conducted an experiment using our asymmetric zigzag sampling technique without the one-prompt method.
clip-I: 0.9101, clip-T: 0.8876, DreamSim: 0.2113
These results still outperform both the Consistory and StoryDiffusion models, demonstrating the generalizability of our proposed sampling strategy. Moreover, when combined with the one-prompt technique, it achieves the best performance, highlighting the complementarity between the two methods.

W3. The method part is hard to follow even after reading ConsiStory and 1Prompt1Story. I am not sure which designs of the previous work are adopted and which are not. I suggest the author to add details for the token caching, and unify the terms such as "one prompt", "SSC" etc., which are not described in the method section. More precise, step-by-step details would improve reproducibility.

[Response]
Thanks for raising this issue and providing the suggestions.
Delineation of Adopted vs. Novel Designs: We will explicitly state which components are adapted from prior work (i.e., , the idea of using an identity-focused prompt to extract subject-relevant visual features from [27], and the prompt fusion/reweighting mechanism from [28] for text fidelity) and clearly highlight our novel contributions (Asymmetry Zigzag Sampling, including Zig Visual Sharing (ZVS) and Asymmetric Prompt Zigzag Inference (APZI)).
Unifying Terminology: We will standardize and clearly define all terms.
"One Prompt" refers to the concatenated prompt strategy introduced by 1Prompt1Story [28], which we integrate for maintaining narrative coherence across scenes.
"SSC" (Subject Semantic Control) refers to our proposed module (illustrated in Figure 3(d)) that performs the asymmetric visual guidance within the zigzag sampling process. Step-by-Step Details: We will provide a more precise, step-by-step algorithmic description ( with pseudocode) for the Asymmetric Zigzag Sampling process, detailing how ZVS and APZI are integrated into the zig, zag, and generation sub-steps. This will significantly improve reproducibility.
W4. The ablation is designed to show that each sampling process is necessary, but is not enough to answer all design choice questions. First I am curious how are the time step set in zig-gen and zig-zag, since they are 2-forward forward-backward, not aligned with conventional time steps. Second, and alternative zag process with SSC or one prompt should be given, to prove zag could fight identity overfitting Third, why adding gen step at last also improves clip-i?

W4.1 how are the time step set in zig-gen and zig-zag?.

[Response]
We apologize for the confusion caused by the "zig-gen" and "zig-zag" terminology in Table 3. This refers to where our proposed visual semantic injection is applied within the zigzag sampling sub-steps, not a different type of zigzag sampling. The core zigzag sampling mechanism (forward denoising for zig/generation, inverse denoising for zag, as defined in line 175) remains consistent across all these ablation settings. The "time step t" in these equations refers to the standard diffusion timestep.
"Asymmetry" (Our Method): Visual injection (ZVS) only in zig step. Text guidance (APZI with full prompt) in zig and generation steps. Null prompt in zag step.
"Zig-gen": Visual injection (ZVS) in zig and generation steps. Text guidance (full prompt) in zig and generation steps. Null prompt in zag step.
"Zig-zag": Visual injection (ZVS) in zig and zag steps. Text guidance (full prompt) in zig and generation steps. Null prompt in zag step.
"All" (Fully Symmetric): Visual injection (ZVS) in zig, zag, and generation steps. Text guidance (full prompt) in zig and generation steps. Null prompt in zag step.
We will clarify this distinction in the revised paper, making it explicit that these ablations examine the impact of where visual conditioning is applied within our established asymmetric zigzag framework, not variations of the zigzag process itself.

W4.2 alternative zag process with ssc or one prompt should be given ?

[Response]
This is a valuable suggestion. Our hypothesis is that using a null prompt in the zag step allows the subject-specific visual information injected in the zig step to propagate and "settle" into the latent representation without being immediately overwritten or overly constrained by new textual descriptions, thereby preventing overfitting to prompt specifics and improving generalizable subject consistency.
We have conducted the following experiment:
"Asymmetry + Zag-Text": Our "Asymmetry" method, but with the full textual prompt (+zag-text) (or even the "one prompt" / "SSC" guidance) also applied during the zag step, instead of a null prompt.
Asymmetry + zag-text clip-I :0.9098 clip-T :0.8841 DreamSim :0.2251.
Asymmetry + zag-null clip-I :0.923 clip-T :0.8946 DreamSim :0.1789.
The results above show that adding text in zag step leads to significant decreases in subject consistent evaluation metrics including clip-Ii and DreamSim score.
The ablation of zag with ssc is exactly the experiment named zig-zag in table 3 ( see above explanation of methods in table 3).

W4.3 why adding gen step at last also improve clip-i ?

[Response]
There might be a slight misunderstanding here. In our "Asymmetry" method, the "gen step" (generation step) always uses the complete textual prompt for guidance. The improvement in CLIP-I (subject consistency) is not solely because of "adding the gen step at last." Rather, it is due to the synergistic effect of the entire asymmetric zigzag sampling process:
Zig Step (Exploration + Initial Identity Grounding): Strong subject-specific visual cues are injected, firmly establishing the subject's identity early on.
Zag Step (Refinement + Propagation, Null Prompt): The latent space is refined without additional textual or visual interference. The null prompt here is critical; it allows the visually conditioned subject information from the zig step to propagate through the denoising process without being immediately pulled too strongly by scene-specific textual prompts, which could otherwise lead to "identity overfitting" or reduce scene diversity.
Generation Step (Final Alignment, Full Text Prompt): This step then focuses solely on aligning the image with the specific textual prompt, leveraging the consistent subject information that has been preserved and propagated from the zig and zag steps. Because the earlier zig and zag steps have already done the heavy lifting of grounding the subject consistently and preparing the latent space, the generation step can effectively refine the image according to the narrative without losing subject identity.

Therefore, the improvement in CLIP-I is a result of this carefully orchestrated asymmetric information flow. The "generation step" is where the narrative is finally realized with the subject consistently maintained due to the preceding zig and zag steps. The comparison in Table 3 specifically highlights that applying visual conditioning during the "generation" step (as in "Zig-gen") leads to poorer text alignment, while our "Asymmetry" method avoids this by focusing the "gen" step on text alignment alone after subject conditioning in the "zig" step.