We thank the reviewer for the thoughtful feedback and will address your concerns concisely.

Q1. Our contribution is not “building a dataset and training Style LoRAs.”

The core contribution is a two-stage, style–consistency disentanglement that yields a plug-and-play, style-agnostic Consistency-LoRA for image-to-image stylization. To our knowledge, we are the first to explicitly target stylization consistency (content/structure preservation across styles) as a stand-alone, reusable module that attaches only to the condition branch and works with any Flux style LoRA without modifying or retraining those style LoRAs. While collecting data and training style LoRAs are common practices (e.g., CSGO), our task differs: we focus on improving i2i consistency under style LoRA usage, not on proposing another style-transfer pipeline.

Q2. Evaluation scale and diversity

The five test styles are outside the training set, and the benchmark contains 100 images, so each reported metric is computed over 500 images (100 × 5). This scale is sufficient to reduce randomness and demonstrate effectiveness. “Curated” does not imply poor diversity; when constructing the benchmark we intentionally covered varied categories (people, animals, buildings, vehicles, products, indoor/outdoor scenes) to ensure diverse content and structures.

Q3. Originality of the two-stage disentangled training

Our strategy is motivated by first principles:

• Stage I (style): train per-style LoRAs and then freeze them.
• Stage II (consistency): train a style-agnostic Consistency-LoRA with the style LoRAs frozen and randomly sampled, encouraging content/structure preservation independent of any single style.

Ablations in Table 2 and Figure 3 show strong generalization to styles unseen during training. Our approach explicitly decouples style learning from consistency learning: in Stage II, we freeze all style LoRAs and randomly sample them while training a single Consistency-LoRA mounted only on the condition branch; causal attention, conditional token mapping, and feature reuse provide efficient conditioning. Style-swap and zero-style ablations indicate that the Consistency-LoRA does not encode a specific style, and performance on unseen styles does not degrade. In contrast, joint training tends to average styles, which weakens content/structure fidelity and reduces generalization.

We surveyed extensive related work and found no prior paper that adopts this two-stage design—freezing per-style LoRAs and training a separate style-agnostic Consistency-LoRA mounted only on the condition branch—to improve i2i stylization consistency across seen and unseen styles.

Q4. More diverse styles / additional evidence

Beyond Figure 2 and the quantitative results in the supplement, we will add more styles and scenes to further stress-test diversity. We also note that the model has been downloaded thousands of times on Hugging Face, indicating practical utility and community interest; this community feedback complements our quantitative analyses.

We thank the reviewer for the thoughtful feedback and will address your concerns comprehensively.

1) Decoupled training vs. joint training

Our goal is to build a plug-and-play, style-agnostic consistency module that—once trained—works with any community Flux style LoRA without retraining. Decoupled training follows a first-principles design:

• Stage I learns style in independent style LoRAs (frozen thereafter).
• Stage II learns consistency only while style modules remain frozen and randomized, which forces the Consistency-LoRA to encode style-agnostic content/structure preservation.

A joint scheme inevitably entangles consistency with specific styles and must be re-optimized when styles change, which defeats the plug-and-play objective. Our ablations confirm that decoupling yields stronger generalization to unseen styles and inputs, whereas joint training tends to overfit the seen style set.

2) LoRA “novelty” (design and objective)

LoRA is a parameter-efficient fine-tuning tool, and many methods implement it similarly. The novelty in our work lies in how LoRA is trained and placed:

• We train on paired images to learn consistency (rather than learning a specific style or subject from exemplars).
• We introduce a two-stage decoupled pipeline that first learns style LoRAs, then learns a Consistency-LoRA with styles frozen and sampled across steps.
• We mount the Consistency-LoRA only on the condition branch, avoiding interference with the style LoRAs at inference time and preserving each style module’s identity while enforcing content/structure stability.

3) Feature reuse is more than KV caching

While the core mechanism leverages KV cache, the efficiency gains arise from the combined design of Conditional Token Mapping and Feature Reuse. In practice, GPU memory and latency overheads remain under 10% over standard text-to-image for our image-to-image stylization, which is substantially more efficient than OmniControl and EasyControl in comparable settings.

4) Ablation settings (Table 2, Figure 5)

All ablation experiments use exactly the same training setup and hyperparameters as the main model (training steps, batch size, optimizer, learning rate, etc.); there is no underfitting caused by configuration, so our ablations are fully fair and sound. The underfitting-like behavior observed is more likely due to joint training, which tends to learn an averaged style.

5) Why causal attention over bidirectional attention

• Cacheability: with a causal mask on the conditional path, conditional keys/values remain static across diffusion steps, enabling KV caching and lowering latency/memory. Under bidirectional attention (naïve concatenation), conditional tokens attend back to evolving denoising tokens, so their states change each step and cache reuse breaks.
• Reduced cross-condition interference: masking restricts condition↔condition attention while allowing denoiser←condition aggregation, stabilizing control when multiple conditions are present.
• Empirical: we observe faster inference and lower memory with causal attention, while maintaining or improving quality relative to bidirectional attention.

We thank the reviewer for the thoughtful questions and will address your concerns comprehensively.

1) Ablation

Using text + Consistency-LoRA alone already yields high-consistency stylization (see Figure 6). Figure 3 and Table 3 can also be viewed as an ablation of Style LoRA: OmniConsistency can be combined with both seen and unseen style LoRA modules to achieve high-quality image stylization consistency, effectively preserving the semantics, structure, and fine details of the original image. In Table 3, the FID and CMMD across 10 styles (5 seen / 5 unseen) show no degradation on unseen styles, supporting that the Consistency-LoRA is style-agnostic rather than tied to any specific trained style.

---

2) Differences from EasyControl

Both our method and EasyControl are built on Diffusion Transformer-based controllable generation, and we both use causal attention and positional-encoding interpolation. Concurrent work OmniControl V2 likewise adopts positional-encoding interpolation. These components are increasingly standard in DiT-based controllable pipelines; our contribution lies in two-stage decoupling and a style-agnostic consistency module that is plug-and-play with style LoRAs. We cite EasyControl in the main text (concurrent work), but there are clear differences:

• Task focus. EasyControl studies DiT controllable generation with layout/subject images as conditions; it does not present a style-transfer task in the paper. Our goal is a general consistency plugin that stabilizes content/structure across any image × any style LoRA.
• Two-stage decoupling and data requirement. Recent attempts train Ghibli style transfer with EasyControl on paired data, or build Simpsons style transfer via OmniControl—but these require collecting paired data and retraining a LoRA for each new style. In contrast, our method is plug-and-play with community Flux style models and does not require per-style retraining of the consistency module.
• Efficiency. With additional inference-time optimizations, our method shows a ~40% reduction in inference time compared to EasyControl and uses less GPU memory.

---

3) Comparison with In-Context LoRA

In-Context LoRA targets image generation and multi-frame consistency, but does not provide image-conditioning (upload an image and obtain a consistent stylized output). The settings differ substantially, so a direct comparison is not applicable.

---

4) Other Questions

• “Does Stage I Style LoRA carry content information, and is Stage II just a correction?”
  We mitigate potential content leakage in Stage I via a rolling loading strategy and by ensuring broad category coverage when collecting training data. Stage II’s Consistency-LoRA is not a mere correction; the two-stage training paradigm and decoupling design are crucial to learning style-agnostic consistency.

• “What happens if the Style LoRA is removed at inference?”
  At inference, we can completely remove the Style LoRA and use a text prompt to specify the style, still producing highly consistent results (see Figure 6).

Weaknesses

1. “This paper is more of an engineering application; generalization and natural-scene cases are insufficient.”
   Response. Our contribution is not an engineering workaround but an architectural solution to the long-standing consistency problem in stylization. OmniConsistency introduces a plug-and-play, style-agnostic Consistency-LoRA that preserves semantics, structure, and fine details for any image × any style. As evidence, Figure 3 already includes group portraits, animals, architecture, vehicles, and products, demonstrating broad coverage of natural scenes. We added additional natural-scene results (e.g., street scenes, indoor scenes, and landscapes) in the supplementary.

2. “Is the Consistency-LoRA truly style-agnostic? How do you know it doesn’t encode style?”
   Response. Two experiments directly support the style-agnostic claim:

   • Figure 3 (Seen vs. Unseen styles). OmniConsistency works equally well with seen and unseen style LoRAs, producing high-quality stylization while consistently preserving content and structure.
   • Table 3 (FID & CMMD across 10 styles: 5 seen / 5 unseen). Scores on unseen styles do not degrade, indicating that the Consistency-LoRA does not entangle style and generalizes beyond the styles used during training.

---

Questions

1. How is the model’s generalization verified? Does it essentially depend on the capabilities of trained LoRAs?
   Response. We evaluate on styles and inputs that are unseen during training. Both Figure 3 and Table 3 show consistent performance on unseen styles and unseen images. OmniConsistency does not rely on any particular style LoRA’s capacity; instead, the Consistency-LoRA decouples style from consistency, providing stable content/structure preservation regardless of which style LoRA is used.

2. What is the significant difference from directly training LoRAs? How do you ensure the Consistency-LoRA carries no style information (how is decoupling guaranteed)?
   Response. Our goal is not to train more style LoRAs. We propose a general consistency plugin that can be combined with any Flux style LoRA to improve image-to-image consistency.

   • Two-stage decoupling.
     • Stage I: Train a bank of style LoRAs (22 styles). These LoRAs are then frozen.
     • Stage II: Freeze the base model and all style LoRAs, and train only the Consistency-LoRA. During training, we randomly sample styles and inputs while optimizing content/structure-preservation objectives. The Consistency-LoRA cannot encode a fixed style, and instead learns style-agnostic consistency.
   • Empirical evidence. Figure 3 (seen/unseen styles) and Table 3 (unseen FID/CMMD not degrading) show that the Consistency-LoRA remains style-agnostic in practice.

3. Does it support generalization to plugins?
   Response. Yes. OmniConsistency is plug-and-play and is compatible with commonly used control plugins such as IP-Adapter (as shown in Figure 6) and EasyControl. We will add more combined results and analyses in the revision/supplementary.

4. The comparison with many recent style transfer methods is lacking.
   Response. The task setting differs: classical “style transfer” takes a single style reference image plus a content image; our setting assumes style is provided by a style LoRA (trained on a set of images) and focuses on consistency enhancement under that usage mode. Within our setting, we already compare against seven strong baselines, including a GPT-4o-based assessment, which is sufficient to demonstrate effectiveness and superiority. To further bridge the gap, we will add “bridging” comparisons in the supplementary, adapting recent style-transfer methods to a comparable setup (e.g., few-shot reference to train a small style LoRA, then applying Consistency-LoRA).

5. How does it support arbitrary styles?
   Response. Because the Consistency-LoRA is style-agnostic and only enforces content/structure consistency, it can be paired with any style module (e.g., any style LoRA, including newly trained few-shot/one-shot styles). Results in Figure 3 and Table 3 show no degradation for unseen styles.

6. Is the content scoring reliable? Validation is missing.
   Response. We use CLIP-Image for content/structure alignment, a widely used metric for consistency/fidelity, and include a GPT-4o evaluation with fully reproducible prompts and settings (provided in the supplementary).