acceptodds
Under review as a conference paper at ICLR 2027

RefFlow: Training-Free Reference-Guided Image Editing via Residual Instance Transport

Abstract

Recent training-free image editing methods with pretrained generative models enable flexible text-driven manipulation, but text guidance alone cannot reproduce the appearance of a specific visual instance. Moreover, even with a visual reference, it remains challenging to simultaneously achieve task diversity and reference fidelity. We introduce RefFlow, a training-free framework for reference-guided image editing via residual instance transport. RefFlow augments semantic editing with a reference-induced velocity residual, implemented through Relationally Consistent Reference Value Transport that adaptively transfers reference appearance via target-adaptive attention retrieval and relational refinement. This design preserves the target editing trajectory while enabling instance-specific visual control without any additional training. RefFlow is compatible with various pretrained models and supports diverse tasks, including object replacement, typography editing, texture transfer, and multi-reference composition. Experiments on RefEdit-Eval and DreamBooth demonstrate improved reference fidelity, editing success, and source preservation over existing training-free and trained baselines.

Primary area:
generative models
Keywords:
Reference-guided image editing, Training-free image editing, Flow matching

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_LDaJOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work introduces REFFLOW, a training-free framework for image editing based on residual instance transport. Built upon a frozen flow-based text-to-image (T2I) model, REFFLOW designs a reference-induced velocity residual module and a Relationally Consistent Reference Value Transport mechanism. By optimizing intra-image feature correspondences and performing spatially gated residual correction, it achieves high-fidelity reference-guided editing without inversion or fine-tuning. The framework natively supports complex tasks such as multi-reference composition, typography editing, and texture transfer.

Critical strengths
  1. This work identifies drawbacks in two major branches of training-free editing methods: conventional text-guided approaches fail to perform instance-level fine-grained editing, while existing reference-guided methods merely conduct naive attention feature replacement without optimizing cross-image correspondences, which leads to structural inconsistency and degraded fidelity.
  2. This work proposes the reference-induced velocity residual to enable lightweight, controllable reference appearance correction while preserving the original text-guided semantic editing trajectory and keeping the backbone model fully frozen.
  3. This work designs the Relationally Consistent Reference Value Transport mechanism to overcome the limitations of conventional independent token-wise matching. It optimizes cross-image correspondences under intra-image feature relational constraints and mitigates detail artifacts caused by structural misalignment.
  4. This work for the first time realizes training-free multi-reference image editing using a completely frozen flow-based T2I model, extending the scope to complex editing scenarios including typography editing and texture transfer.
Critical weaknesses
  1. RefEdit‑Eval is a self‑curated benchmark. The paper only presents aggregated overall results without per‑task score for typography editing, texture transfer, object editing and style transfer. Per‑category metrics would help readers better understand where REFFLOW excels and where it still has shortcomings.
  2. The paper lacks quantitative experiments targeting multi‑reference scenarios.
  3. While failure cases are briefly discussed in Appendix I, the analysis remains qualitative. Adding quantitative statistics for different failure categories would enrich the discussion of the method’s limitations.
Questions for authors
  1. RefEdit‑Eval is a self‑curated benchmark. In the paper, only aggregated overall results are reported, without separate per-task scores for typography editing, texture transfer, object editing, and style transfer. Could you please provide per-category metrics to help readers better understand the strengths and remaining limitations of REFFLOW across these individual tasks?
  2. The current work does not include quantitative experiments under multi-reference scenarios. Are there any quantitative results compared to other methods?
  3. Although failure cases are briefly discussed qualitatively in Appendix I, no quantitative statistics are provided for distinct failure categories. Could you add quantitative statistics for different failure types to further support and enrich the discussion on the limitations of your method?
5 comments
AuthorsOpenReview ↗

Response to W1/Q1: Task-wise Quantitative Evaluation

We thank the reviewer for this constructive suggestion. To provide a more fine-grained assessment beyond the aggregated results, we have added a task-wise quantitative evaluation on RefEdit-Eval in the updated manuscript.

Specifically, Appendix F (Table 4, p. 17) reports five evaluation metrics (CLIP-FG, DINO-FG, CLIP-T, PSNR-BG, and LPIPS-BG) separately across five editing categories: Object Placement, Object Replacement, Typography Editing, Texture Transfer, and Style Transfer, following the same evaluation settings as Table 1.

Table 4: Task-wise quantitative results on RefEdit-Eval. Performance of RefFlow across five editing categories. The best result in each metric is bolded. (Appendix F, p. 17 of the updated manuscript.)

Task categoryCLIP-FG ↑DINO-FG ↑CLIP-T ↑PSNR-BG ↑ (dB)LPIPS-BG ↓
Object Placement0.9210.7790.34828.770.105
Object Replacement0.9200.8220.32420.710.199
Typography Editing0.9480.6960.37523.460.234
Texture Transfer0.8210.2240.34526.370.084
Style Transfer0.8840.7730.35722.030.412

The results reveal distinct performance characteristics across tasks. RefFlow achieves strong reference fidelity in typography editing and object placement/replacement, while object placement also exhibits favorable background preservation. Meanwhile, texture transfer shows relatively lower DINO-FG, and style transfer exhibits higher LPIPS-BG, highlighting task-dependent challenges in reference fidelity and source preservation.

These additional results provide a more detailed assessment of RefFlow's strengths and remaining limitations across different editing categories.

AuthorsOpenReview ↗

Response to W2/Q2: Quantitative Multi-Reference Evaluation

We thank the reviewer for highlighting the need for quantitative multi-reference evaluation. To address this concern, we have conducted additional quantitative experiments covering two to five reference images, comparing RefFlow against two representative multi-reference editing models, DreamOmni2 and Qwen-Image-Edit-2511, on the same editing cases.

The new results are reported in Table 2 (p. 9), covering reference fidelity (CLIP-FG and DINO-FG), text alignment (CLIP-T), background preservation (PSNR-BG and LPIPS-BG), and editing success (Human-SR). RefFlow consistently achieves better aggregate performance than both baselines across these metrics, demonstrating its effectiveness in multi-reference editing while preserving unrelated source content.

Table 2: Quantitative comparisons on multi-reference editing. Results are averaged across editing cases with two to five reference images. The best and second-best results are bolded and underlined, respectively. (p. 9 of the updated manuscript.)

MethodCLIP-FG ↑DINO-FG ↑CLIP-T ↑PSNR-BG ↑ (dB)LPIPS-BG ↓Human-SR ↑ (%)
DreamOmni20.7190.1340.316‾\underline{0.316}15.36‾\underline{15.36}0.277‾\underline{0.277}16.7
Qwen-Image-Edit-25110.758‾\underline{0.758}0.272‾\underline{0.272}0.30914.130.34340.0‾\underline{40.0}
RefFlow (Ours)0.9070.8030.36926.700.06586.7

Additional side-by-side qualitative comparisons are provided in Figure 11 (p. 19).

AuthorsOpenReview ↗

Response to W3/Q3: Quantitative Failure Analysis

We thank the reviewer for this constructive suggestion. Building upon the qualitative failure analysis in the original submission, we have added a quantitative breakdown of failure categories in Appendix J (Table 7, p. 24).

Table 7: Failure statistics of RefFlow. Frequency of different failure types across 40 unsuccessful editing cases. Multiple labels may be assigned to each case. (Appendix J, p. 24 of the updated manuscript.)

Failure categoryCountIncidence (%)
Edit Omission717.5
Reference Appearance Mismatch2255.0
Prompt Following Failure922.5
Source Preservation Failure1230.0
Rendering Artifacts1127.5

Specifically, we analyze 40 unsuccessful editing cases identified through human evaluation and classify them into five failure categories: Edit Omission (17.5%), Reference Appearance Mismatch (55.0%), Prompt Following Failure (22.5%), Source Preservation Failure (30.0%), and Rendering Artifacts (27.5%). Since a single editing result may exhibit multiple failure types, we adopt multi-label annotation.

The analysis shows that reference appearance mismatch is the most frequent failure type, indicating that preserving fine-grained instance-specific details remains a primary challenge. Together with the qualitative example in Figure 16 (p. 24), these additional statistics provide a more systematic characterization of RefFlow's remaining limitations.

We sincerely thank the reviewer for their positive assessment of our work and constructive feedback. We appreciate their recognition of RefFlow's contributions, and believe that the additional task-wise evaluations, multi-reference comparisons, and quantitative failure analysis have further strengthened the experimental evidence and clarified the strengths and limitations of our approach.

Reviewer_JnU2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RefFlow, a training-free framework for reference-guided image editing. It combines reference value transport, relational correspondence refinement, and spatially gated residual editing to transfer instance-specific appearance from one or multiple reference images while preserving the source image structure.

Critical strengths

1.The method is technically well presented, with a clear formulation of reference-induced residual editing and relational correspondence refinement.

2.The training-free setting is practically appealing, as it operates on a frozen T2I backbone without task-specific fine-tuning.

3.The paper evaluates both reference fidelity and source preservation, and reports promising results over several training-free and trained baselines.

Critical weaknesses

1.The core mechanism is close to prior matching-aware reference value injection, such as DreamMatcher. Moreover, TF-TI2I already studies training-free multi-reference generation with frozen T2I models. RefFlow’s main difference appears to be source-preserving editing with residual velocity composition and relational refinement. The paper should better clarify whether this is a substantial methodological advance rather than an incremental extension, and the “first method” claim seems too strong.

2.The paper claims task diversity and multi-reference capability, but results are aggregated across tasks and multi-reference editing is mostly qualitative. In addition, the comparison protocol is insufficiently specified, and the ablations do not isolate the residual velocity formulation from alternative reference-conditioning and attention-transport designs.

Questions for authors

1.Can the authors provide quantitative results across different task categories and multi-reference editing settings, rather than only reporting aggregated results and qualitative examples?

2.How are RefEdit-Eval cases, masks, prompts, and source-reference pairs constructed?

3.What prompts and inference settings are used for Qwen-Image-Edit-2511 and other baselines?

4.Can the authors ablate residual composition, refinement sensitivity, automatic masks, and multi-reference fusion?

8 comments
AuthorsOpenReview ↗

Response to W1: Methodological Novelty and Distinction from Prior Work

We thank the reviewer for this important comment. While prior work has explored matching-aware reference injection and training-free multi-reference conditioning, RefFlow differs in its reference transport formulation and source-preserving editing mechanism.

1. Distinction from DreamMatcher. DreamMatcher [1] employs dense matching and cycle-consistency filtering for reference value warping. In contrast, RefFlow explicitly optimizes soft correspondences under within-image relational constraints (Eq. 9) to maintain structural consistency across target regions. More fundamentally, RefFlow introduces reference appearance as a spatially gated velocity residual over the original semantic editing direction (Eq. 3), explicitly separating instance-specific reference guidance from source-preserving editing.

2. Distinction from TF-TI2I. Although TF-TI2I [2] supports training-free multi-image inputs, its experiments primarily focus on compositional generation and appearance/style blending (Sec. 5.1, Figs. 5–7 of [2]). This emphasis is consistent with its implicit visual-context mechanism, which aggregates reference information into contextual text tokens for semantic conditioning.

In contrast, RefFlow supports a broader range of targeted, source-preserving editing tasks, including not only style and texture transfer, but also explicit object insertion and replacement, typography editing, and multi-reference instance composition. Our explicit reference value transport and spatially gated velocity residuals enable localized transfer of instance-specific appearance while preserving unrelated source content. We have clarified this distinction in Section 2.2 (Related Work, p. 3) of the updated manuscript.

3. Clarification of the novelty claim. To avoid overclaiming, we have revised the statement in Introduction (p. 2). The claim is now limited to achieving region-controlled multi-reference image editing on a completely frozen T2I model, covering object, typography, and texture edits. To our knowledge, this setting is not provided by TF-TI2I or other methods mentioned in our work.

References

[1] Jisu Nam et al. DreamMatcher: Appearance Matching Self-Attention for Semantically-Consistent Text-to-Image Personalization. CVPR, pp. 8100–8110, 2024.

[2] Teng-Fang Hsiao et al. TF-TI2I: Training-Free Text-and-Image-to-Image Generation via Multi-Modal Implicit-Context Learning in Text-to-Image Models. ICCV, pp. 18377–18387, 2025.

AuthorsOpenReview ↗

Response to W2/Q1: Task-wise and Multi-reference Quantitative Evaluations

We thank the reviewer for the constructive suggestion regarding task diversity and multi-reference evaluation. We have uploaded an updated manuscript incorporating additional experiments and clarifications. All section, figure, table, and page references below refer to this revised version. To provide a more comprehensive assessment beyond the aggregated results and qualitative examples, we have added the following quantitative evaluations.

1. Task-wise quantitative evaluation. We additionally report category-wise results on RefEdit-Eval, covering five editing categories: Object Placement, Object Replacement, Typography Editing, Texture Transfer, and Style Transfer. The evaluation follows the same settings as the main experiments in Table 1. The new results are provided in Appendix F (Table 4, p. 17), reporting reference fidelity, text alignment, and background preservation across these categories. This provides a more fine-grained assessment of RefFlow's performance across different editing tasks.

Table 4: Task-wise quantitative results on RefEdit-Eval. Performance of RefFlow across five editing categories. The best result in each metric is bolded. (Appendix F, p. 17 of the updated manuscript.)

Task categoryCLIP-FG ↑DINO-FG ↑CLIP-T ↑PSNR-BG ↑ (dB)LPIPS-BG ↓
Object Placement0.9210.7790.34828.770.105
Object Replacement0.9200.8220.32420.710.199
Typography Editing0.9480.6960.37523.460.234
Texture Transfer0.8210.2240.34526.370.084
Style Transfer0.8840.7730.35722.030.412

2. Multi-reference quantitative evaluation. To further substantiate the multi-reference capability previously demonstrated through qualitative examples, we conducted additional quantitative evaluations of RefFlow with multiple reference images (up to five). Furthermore, we evaluated two representative pretrained multimodal editing models, Qwen-Image-Edit-2511 and DreamOmni2, under the same evaluation protocol.

The new results are reported in Table 2 (p. 9), where RefFlow demonstrates clear advantages in multi-reference editing, particularly in reference fidelity, background preservation, and editing success.

Table 2: Quantitative comparisons on multi-reference editing. Results are averaged across editing cases with two to five reference images. The best and second-best results are bolded and underlined, respectively. (p. 9 of the updated manuscript.)

MethodCLIP-FG ↑DINO-FG ↑CLIP-T ↑PSNR-BG ↑ (dB)LPIPS-BG ↓Human-SR ↑ (%)
DreamOmni20.7190.1340.316‾\underline{0.316}15.36‾\underline{15.36}0.277‾\underline{0.277}16.7
Qwen-Image-Edit-25110.758‾\underline{0.758}0.272‾\underline{0.272}0.30914.130.34340.0‾\underline{40.0}
RefFlow (Ours)0.9070.8030.36926.700.06586.7

Additional side-by-side qualitative comparisons are provided in Figure 11 (p. 19). Together, these additional evaluations provide more detailed evidence for RefFlow's effectiveness across diverse editing tasks and multi-reference configurations.

AuthorsOpenReview ↗

Response to Q2: Benchmark Construction, Masks, and Prompts

We thank the reviewer for requesting further details on the evaluation data construction.

1. Source–reference pair construction. As described in Section 5.1 (p. 6), RefEdit-Eval uses real images from Unsplash and synthetic images generated by Qwen-Image-2512. To further explain how editing cases are constructed, we have added the pairing procedure to Section 5.1 (p. 6). Specifically, we first manually curate separate source and reference image pools. Candidate pairs are constructed through random sampling, followed by Gemma 4-assisted visual screening and final manual verification to ensure that each pair satisfies the intended task-specific editing requirements.

2. Mask construction. We use the off-the-shelf RMBG-2.0 model to precompute the masks used in our main evaluations. We have explicitly clarified this protocol in Appendix B (Implementation Details, p. 14) of the updated manuscript.

3. Prompt construction. As stated in Appendix B (p. 14), the source, reference, and target prompts are generated with the assistance of Gemma 4. To further clarify our protocol, we adopt a three-part descriptive format: the source prompt describes the original scene, the reference prompt specifies the reference appearance, and the target prompt describes the intended edited scene. Following a minimal-edit principle, only task-relevant content is modified, while unrelated source-scene details are preserved. All generated prompts are manually verified. An illustrative example is provided in Figure 7 (p. 16).

Reviewer_aWNvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RefFlow, a training-free framework for reference-guided image editing via residual instance transport. RefFlow augments semantic editing with a reference-induced velocity residual. The authors propose Relationally Consistent Reference Value Transport. It uses target-adaptive attention retrieval and refines correspondences using within-image feature relations. RefFlow supports object replacement, typography editing, texture transfer, and multi-reference composition. Experiments on RefEdit-Eval and DreamBooth demonstrate improved reference fidelity, editing success, and source preservation over existing training-free and trained baselines.

Critical strengths
  1. RefFlow adds instance-specific appearance control to FlowEdit without any additional training. It keeps the text-guided semantic editing velocity and adds a reference-induced velocity residual. This residual isolates the reference contribution and enables explicit control over its strength and spatial support. The framework supports object replacement, typography editing, and texture transfer. On RefEdit-Eval and DreamBooth, RefFlow shows higher reference fidelity than the compared methods.

  2. Relationally Consistent Reference Value Transport refines target-reference correspondence by preserving within-image feature relations. Without this refinement, independent matches may assign inconsistent reference details to related target regions. Table 3 shows that adding this refinement increases DINO-FG from 0.755 to 0.896 on DreamBooth.

Critical weaknesses
  1. Although RefFlow is training-free, it requires additional reference and target forward passes, feature caching, and iterative correspondence refinement during inference. These operations may introduce considerable runtime and memory overhead compared with conventional training-free editing methods.

  2. The mask setting in the main experiments is unclear. Both masks can be externally provided or automatically estimated (Appendix A). The paper does not state which option was used for Tables 1 and 2. This is important because the masks control reference retrieval and where the edit is applied.

  3. Section 4.1 states, “When λ = 0, RefFlow reduces to FlowEdit.” In Figure 5, CLIP-FG is about 0.93 at λ = 0. Table 2 reports 0.824 for Vanilla FlowEdit on DreamBooth. Figure 5 does not state its dataset or settings. This makes it hard to judge how much of the gain over FlowEdit comes from the reference-induced residual.

Questions for authors
  1. Please state whether the masks used for Tables 1 and 2 were externally provided or automatically estimated. If the masks were externally provided, please also show results with automatically estimated masks. This would show whether mask estimation changes the results.

  2. Please state the dataset and settings used for Figure 5. Please explain why its λ = 0 result differs from the Vanilla FlowEdit result in Table 2. If the settings differ, please also report a λ = 0 result under the same settings as Table 2.

3 comments
AuthorsOpenReview ↗

Response to Weakness1: Computational Efficiency and Multi-Reference Scalability

We thank the reviewer for highlighting the computational efficiency of RefFlow. We have uploaded an updated manuscript incorporating additional experiments and clarifications. All section, figure, table, and page references below refer to this revised version.

We acknowledge that additional forward passes, feature caching, and correspondence refinement introduce inference overhead, as discussed in Appendix D (Figure 8) and Appendix J (Limitations). However, we believe this overhead represents a reasonable trade-off for the substantial improvements in reference fidelity and editing success.

First, the additional computation does not pose a practical barrier to using RefFlow. As shown in Figure 8 (p. 16), RefFlow requires 22.50 s and 34.93 GiB per image, achieving 39.1% and 71.9% faster inference than DreamOmni2 and Qwen-Image-Edit-2511, respectively, with lower GPU memory consumption. Both models remain representative reference-guided editors evaluated in recent studies [1, 2], supporting the practical relevance of this comparison. Together with improved reference fidelity and editing success (Tables 1 and 3), these results demonstrate a favorable quality–efficiency trade-off.

Second, RefFlow enables efficient multi-reference editing, which is not natively supported by existing training-free baselines such as SpecRef and ReInversion. As formulated in Eq. (14), RefFlow combines independently computed, spatially gated reference corrections under shared target attention. Unlike reference-token concatenation, this formulation does not increase the target attention sequence length as the number of references grows.

Additional multi-reference efficiency experiments. To further substantiate this property, we have added runtime comparisons with multiple reference images in Appendix D (Figure 9, p. 17). The results show that RefFlow exhibits moderate runtime growth as the number of references increases, while maintaining a clear efficiency advantage over DreamOmni2 and Qwen-Image-Edit-2511. Notably, with four reference images, RefFlow is approximately 2.0× faster than DreamOmni2 and 4.4× faster than Qwen-Image-Edit-2511. These findings further demonstrate the practical efficiency and scalability of RefFlow in multi-reference editing scenarios.

References

[1] Jingxu Zhang, Daneul Kim, Yueming Pan, Dong Chen, Kai Qiu, Yang Liu, Yifan Yang, Qi Dai, Xiaoyan Sun, and Chong Luo. RCEdit-500K: Reference Completion for Image-Conditioned Image Editing. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 533–550, 2026.

[2] Xiaoye Liang, Zhiyuan Qu, Mingye Zou, Jiaxin Liu, Lai Jiang, Mai Xu, and Yiheng Zhu. VTEdit-Bench: A Comprehensive Benchmark for Multi-Reference Image Editing Models in Virtual Try-On. In Computer Vision – ECCV 2026, Lecture Notes in Computer Science, vol. 17055, pp. 658–675. Springer, 2026.

AuthorsOpenReview ↗

Response to W2/Q1: Clarification and Evaluation of Mask Estimation

We thank the reviewer for pointing out the need to clarify the mask settings. For both Table 1 (RefEdit-Eval) and Table 3 (DreamBooth), we used externally provided masks precomputed using model RMBG-2.0 [1, 2]. We have explicitly clarified this setting in Appendix B (p. 14) in the updated manuscript.

Additional evaluation of automatic mask estimation. As requested, we reran RefFlow on both benchmarks using the automatic attention-based mask estimation described in Appendix A, keeping all other inference settings unchanged. The results are reported in Appendix H.3 (Table 5, p. 21).

Automatic masks do change the results, mainly in reference fidelity: CLIP-FG decreases from 0.909 to 0.847 on RefEdit-Eval and from 0.948 to 0.910 on DreamBooth, while DINO-FG decreases only slightly (0.686 to 0.671 and 0.896 to 0.883). Background preservation is nearly unchanged on RefEdit-Eval, with a small increase in LPIPS-BG on DreamBooth (0.324 to 0.346). Human-SR decreases by only 1.6 and 1.2 percentage points (90.5 to 88.9 and 66.7 to 65.5), indicating that editing success is largely maintained.

Table 5: Effect of mask estimation on RefFlow. Comparison between externally provided masks (RMBG-2.0) and automatic attention-based mask estimation on RefEdit-Eval and DreamBooth. The best result within each dataset is bolded. (Appendix H.3, p. 21 of the updated manuscript.)

DatasetMask schemeCLIP-FG ↑DINO-FG ↑CLIP-T ↑PSNR-BG ↑ (dB)LPIPS-BG ↓Human-SR ↑ (%)
RefEdit-EvalExternal (RMBG-2.0)0.9090.6860.35125.370.17590.5
Automatic (Ours)0.8470.6710.33225.400.17788.9
DreamBoothExternal (RMBG-2.0)0.9480.8960.36620.030.32466.7
Automatic (Ours)0.9100.8830.36720.010.34665.5

This reduction is in line with our analysis of mask localization in Appendix H.2 (Figure 14), p. 21, where shifting the reference or placement mask lowers reference fidelity while leaving background preservation largely unaffected.

Taken together, we believe these complementary experiments provide a comprehensive assessment of the role of mask estimation, spatial gating, and localization accuracy in RefFlow.

References

[1] BRIA AI. RMBG-2.0: BRIA Background Removal v2.0. Hugging Face Model Card, 2024. https://huggingface.co/briaai/RMBG-2.0.

[2] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral Reference for High-Resolution Dichotomous Image Segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024.

AuthorsOpenReview ↗

Response to W3/Q2: Clarification of Ablation Settings and the lambda=0 Baseline

We thank the reviewer for pointing out this discrepancy. The original Figure 5 and Table 2 (now Table 3) were obtained under different evaluation settings, in two respects:

(1) Evaluation samples. The original Figure 5 was evaluated on a randomly sampled subset of 42 cases (25 from RefEdit-Eval and 17 from DreamBooth), whereas the Vanilla FlowEdit result in the original Table 2 was evaluated on the full DreamBooth benchmark (84 cases).

(2) Inference configurations. RefFlow and Vanilla FlowEdit use their respective fixed inference settings, with nmax⁡=23n_{\max}=23 for RefFlow and nmax⁡=24n_{\max}=24 for the official FLUX implementation of FlowEdit. Here, nmax⁡n_{\max} specifies the number of final editing steps selected from a 28-step sampling schedule. While both configurations are used for their respective methods in the main evaluations, verifying the λ=0\lambda=0 equivalence requires matched inference settings.

To address this concern, we have conducted two complementary evaluations:

1. Full-benchmark ablation. We have rerun the complete ablation on the full RefEdit-Eval (126 cases) and DreamBooth (84 cases) benchmarks, using the default RefFlow settings and varying only λ\lambda and the residual gate. The updated Figure 5 (p. 10) reports CLIP-FG averaged over all 210 cases. The evaluation protocol is now explicitly stated in Section 5.3 (p. 10) and the figure caption.

2. Verification of the λ=0\lambda=0 equivalence. We additionally performed a controlled comparison by matching nmax⁡n_{\max} and all other inference settings to Vanilla FlowEdit. The results are:

CLIP-FG ↑nmax⁡n_{\max}RefEdit-EvalDreamBooth
Vanilla FlowEdit (Tables 1 and 3)240.7460.824
RefFlow, λ=0\lambda=0 (matched)240.7460.824

Under matched settings, λ=0\lambda=0 reproduces the Vanilla FlowEdit CLIP-FG scores on both benchmarks, identical to the reported precision. This is consistent with Eq. (3): when λ=0\lambda=0, the reference-induced residual vanishes, and RefFlow reduces to Vanilla FlowEdit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.