acceptodds
Under review as a conference paper at ICLR 2027

Unlocking Training-Free Continuous Attribute Control in Text-to-Image Diffusion Transformers via M-Space

Abstract

Text-to-image diffusion transformers (DiTs) have shown impressive capabilities in generating high-quality images, yet lack straightforward mechanisms to continuously adjust the strength of individual semantic attributes during image synthesis. To achieve such control, existing approaches often rely on training additional low-rank control adapters or optimizing semantic directions in textual embedding space, incurring extra computational costs. In this paper, we investigate whether the internal conditioning representations of pre-trained text-to-image DiTs can be directly used for continuous attribute control. Our analysis identifies a compact global modulation conditioning space, termed M-space, in which a shared conditioning representation is used to produce layer-wise modulation parameters in text-to-image DiTs. Attribute directions in M-space can be obtained by simply averaging representation differences across diverse pairs of text prompts that differ only in the target attribute. Linear traversal along these directions enables smooth and bidirectional attribute control, while their composition enables multiple attributes to be manipulated simultaneously. Furthermore, we introduce Self-Attention Map Blending (SAMB) and Attention-Guided Modulation Routing (AGMR) to retain attribute-irrelevant content during editing and enable instance- level attribute control, respectively. Extensive qualitative and quantitative experiments on text-to-image DiTs demonstrate that, without requiring any additional training, our method achieves performance comparable to existing state-of-the-art approaches in continuous attribute control.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Continuous attribute control, image editing, text-to-image generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_eJCjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for continuous attribute control in the modulation space (M-space) of a pretrained DiT. It builds attribute direction vectors from the differences between paired text representations and adds them during generation. This allows continuous attribute control without training. The paper also introduces SAMB and AGMR to preserve the image’s spatial structure and keep non-target regions unchanged during attribute control. The experiments show a good balance between control strength and content preservation.

Critical strengths
  1. The method is simple and requires no additional training. It uses the existing M-space of a pretrained DiT and averages the differences between paired text representations to build reusable attribute directions. There is no need to train an adapter for each attribute or optimize model parameters. This provides a low-cost approach to continuous attribute control.

  2. The proposed modules achieve a good balance between control strength and content preservation. In the reported experiments, the method outperforms the compared baselines on three attribute response metrics and the LPIPS and DINO metrics for content preservation. The ablation studies on SAMB and AGMR also support their ability to preserve non-target regions and spatial structure.

Critical weaknesses
  1. The process for building attribute direction vectors lacks key details and ablation studies. For example, how many prompt pairs are used? How are these prompts chosen? How do the authors ensure that averaging removes unrelated factors and keeps only the target semantic direction, without introducing other biases?

  2. The method is designed around the properties of M-space. It therefore depends on architectures such as FLUX and SD3, which take pooled conditional embeddings as input. Newer generative models, such as Qwen-Image, no longer use these embeddings, and their AdaLN is conditioned only on timestep embeddings. The proposed method therefore cannot be applied to these models.

Questions for authors
  1. How is the range of the attribute control coefficient λ determined? How is the “maximum effective control strength for each method” in Section 4.1 measured? For the proposed method, how does the amount of visual change relate to λ? Is the relationship roughly linear?

  2. In SAMB, how is the strength coefficient η set for the source and edit branches?

3 comments
AuthorsOpenReview ↗

Response to Reviewer eJCj (Part 1/2): Weaknesses 1-2 & Question 1

Response to Critical Weakness 1.

We thank the reviewer for requesting these details. We use N = 100 prompt pairs per attribute. Within each pair, only the target-attribute phrase changes, while the subject, non-target attributes, and context remain identical. For example, an age-control pair is “a young man in a park” and “an elderly man in a park.” Across pairs, we vary subject descriptions, non-target attributes, and contexts to reduce dependence on any particular prompt.

This construction is motivated by prior work on linear semantic representations and prompt ensembling. The Linear Representation Hypothesis and the Geometry of Large Language Models (ICML 2024) formalizes concept directions using counterfactual word pairs and estimates them by averaging representation differences across pairs. In vision-language applications, StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery (ICCV 2021) constructs text directions from differences between template-averaged embeddings to reduce sensitivity to wording. These studies provide motivation and empirical precedents for our direction construction strategy.

Specifically, we compute each pair’s M-space representation difference and average these differences. Matching the descriptions within each pair helps reduce contributions from shared content, while averaging across diverse pairs attenuates wording- and context-dependent variations that are inconsistent across pairs. This is intended to emphasize the shared target-attribute direction; it does not guarantee the removal of all unrelated factors or systematic biases.

To assess sensitivity to the number and selection of prompt pairs, we conducted an ablation over 25 attributes. For each attribute, we construct a reference direction from 10,000 prompt pairs, repeatedly sample subsets of different sizes, and compute the cosine similarity between each subset-derived direction and the reference. We average over repeated subsets within each attribute and then give all attributes equal weight. At N = 100, the mean cosine similarity is approximately 0.992, indicating stable direction estimation across sampled subsets of the evaluated prompt pool. Details are provided in Appendix L.

Response to Critical Weakness 2.

We thank the reviewer for highlighting this architectural dependency. Our current method requires a pooled text condition that contributes to a globally shared modulation representation in the DiT. Both FLUX and SD3.5 provide this interface, allowing us to construct attribute directions in their respective M-spaces without additional training.

We agree that our method cannot be directly applied to architectures without this text-conditioned modulation pathway. Z-Image and Qwen-Image do not provide the corresponding interface required by our method. Consequently, our current training-free formulation cannot be directly applied to these models without modifying the method or introducing additional model-specific adaptations.

More broadly, our findings highlight shared text-conditioned modulation as a useful interface for continuous semantic control. We hope this perspective can inform future DiT architecture design by motivating explicit, accessible conditioning interfaces for controllable generation.

Response to Question 1.

We determine the effective control range for each method by progressively varying its control coefficient and jointly assessing target-attribute response and image quality. For our method, CLIP, SigLIP, and BLIP attribute responses all peak at λ = 18. However, more than 80% of the generated images at this scale exhibit hand abnormalities. We therefore set the maximum effective control strength to λ = 16 to balance attribute response and image quality. Each comparison method is calibrated using the same procedure.

By construction, the displacement in M-space is directly proportional to λ. The resulting visual changes, however, pass through a nonlinear generation process. As shown in Appendix L, the mean CLIP, SigLIP, and BLIP attribute responses exhibit a roughly linear increasing trend over much of the effective control range. At excessively large λ values, image degradation, including noise and other artifacts, contributes to the subsequent decline in these scores.

AuthorsOpenReview ↗

Response to Reviewer eJCj (Part 2/2): Question 2

Response to Question 2.

We set η = 0.6: the edited branch blends its own self-attention map with the source map using weights of 0.6 and 0.4, respectively, while the source branch remains unchanged.

The ablation from Appendix N is reproduced below. Smaller η strengthens source-map blending, improving content preservation at the expense of attribute response. We select η = 0.6 to balance attribute response, continuity, and content preservation.

E: extrapolation; C: continuity.

ηE-CLIP ↑E-SigLIP ↑E-BLIP ↑C-CLIP ↑C-SigLIP ↑C-BLIP ↑LPIPS-Alex ↓LPIPS-VGG ↓DINO ↓ID ↓
1.0 (no SAMB)0.26220.07880.51690.27350.22070.25240.27060.30660.16000.5035
0.80.25760.07730.50620.28610.23020.31440.15040.19280.08890.4065
0.70.25380.07680.46160.27920.22260.27390.11640.15990.06810.3609
0.6 (default)0.25060.07580.45770.27360.21170.23710.09710.14070.05020.3322
0.50.24660.07410.37490.26130.20920.22490.08430.12750.04540.2744
0.40.24190.07060.33850.24610.20120.20850.07090.11980.03880.2428

Thank you for the response. My concerns regarding the construction of attribute direction vectors and the strategies for selecting certain hyperparameters have been adequately addressed. I decide to maintain the original score.

Reviewer_So9FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free method continuous attribute control in text to image diffusion transformers using a compact global modulation conditioning space. It further introduces Self-Attention Map Blending (SAMB) to retain attribute-irrelevant content during editing and Attention-Guided Modulation Routing (AGMR) to enable instance level attribute control. Experiments are conducted on Flux.1-dev and Stable Diffusion 3.5 Large across a diverse set of attributes.

Critical strengths
  • The method is simple and training-free. A concept direction is a mean difference over prompt pairs and can be reused across prompts and seeds.

  • The comparison of P-, M-, and M+ space (Fig. 2) gives a useful empirical study at where DiT text conditioning is most editable.

Critical weaknesses
  • The paper claims to be “the first training-free approach for continuous semantic attribute control in text-to-image DiTs”. However, FluxSpace (Dalva et al., 2024) is also training-free and includes a control that edits the pooled text representation, which is closely related to the P- and M-space editing studied here. The authors should either qualify this claim or explain how their approach differs. FluxSpace appears only in Appendix I but is excluded from Table 1.

  • The novelty of the method is incremental as SAMB is similar to the self-attention injection used in Prompt-to-Prompt (Hertz et al., ICLR 2023) and MasaCtrl (Cao et al., ICCV 2023) and AGMR’s cross-attention masking resembles Prompt-to-Prompt’s local blending.

  • Missing related works. SAEdit (Kamenetsky et al., 2025), All-in-One Slider (Ye et al., CVPR 2026), CompSlider (Zhu et al., ICCV 2025), SliderSpace (Gandikota et al., ICCV 2025), and Prompt Sliders (Sridhar et al., ECCVW 2024). These should be discussed and compared. SAEdit in particular also offers continuous token-level control and targets the same instance-level goal as AGMR.

  • The disentanglement gains come from SAMB not from M-space as Table 2 shows that without SAMB, LPIPS-Alex rises from 0.097 to 0.271 and DINO from 0.050 to 0.16. SAMB could be added to any baseline.

  • The quantitative gains in Table 1 are marginal. SliderEdit achieves the best continuity on all three metrics. SliderEdit and Kontinuous Kontext both preserve identity better (ArcFace ID distance 0.284 vs. 0.332). The extrapolation improvements are small (e.g., CLIP 0.2506 vs. 0.2490 for SliderEdit).

  • Qualitative results in Figure 6 and in Figs 9, 10, 11 do not show continuous control but change abruptly. This matches the weak continuity scores in Table 1.

Questions for authors
  • How was the “maximum effective control strength” chosen for each method, and how sensitive are the metrics to that choice?

  • What is the value of η, and how many prompt pairs N are used per attribute? How do the results vary with N and with the choice of prompts?

  • Since FLUX uses CLIP only for the pooled embedding, how does Text Slider’s effect on FLUX differ from a P-space edit?

  • There are no quantitative results on SD3.5 and for real-image editing, as well as a quantitative evaluation of AGMR’s leakage to non-target subjects.

  • What is the inference-time overhead of running the parallel source branch?

  • There is no user study on continuity and attribute fidelity, given the small differences in the reported metrics.

6 comments
AuthorsOpenReview ↗

Response to Reviewer So9F (Part 1/3): Weaknesses 1–4

Response to Critical Weaknesses 1 and 3.

We thank the reviewer for raising concerns about the scope of our claim and the incomplete discussion of related work. We acknowledge that we missed some closely related works highlighted by the reviewer and that our claim of being the first training-free approach for continuous semantic attribute control in text-to-image DiTs was not sufficiently substantiated. We have therefore removed this claim and revised the Introduction and Related Work to discuss FreeSliders, FluxSpace, SAEdit, All-in-One Slider, CompSlider, SliderSpace, and Prompt Sliders. We have also added quantitative comparisons with relevant baselines in the Quantitative Results section.

We further clarify the distinction from FluxSpace. FluxSpace performs fine-grained and coarse control by interpolating orthogonal attention outputs and pooled text conditions, respectively; the latter corresponds to P-space in our paper. In contrast, we construct attribute directions by averaging representation differences across contrastive prompt pairs that differ only in the target attribute in the DiT's shared text-conditioned modulation space. We systematically compare P-space, M-space, and M+-space, and introduce SAMB and AGMR for content preservation and instance-level attribute control, respectively.

For SAEdit, we explicitly discuss its instance-level editing capability and include it in the additional quantitative comparison. These experiments evaluate target-attribute response, continuity, content preservation, and attribute leakage on the non-target person in two-person scenes. Detailed results are provided in Appendix K.

Response to Critical Weakness 2.

We thank the reviewer for highlighting these connections. We acknowledge that attention-based content preservation and semantic region localization have precedents in Prompt-to-Prompt and MasaCtrl. We clarify below how the mechanisms of SAMB and AGMR differ from those of these prior approaches.

(1) SAMB: Self-attention map blending for content preservation.

SAMB is inspired by prior attention-control methods, including Prompt-to-Prompt and MasaCtrl. Prompt-to-Prompt uses self-attention replacement to preserve source structure. SAMB instead blends the source and edited self-attention maps, allowing both source structural relationships and edit-specific attention patterns to contribute to generation.

MasaCtrl constructs mutual self-attention using queries from the editing branch and keys and values from the source branch. SAMB blends the attention maps while retaining the editing branch’s value features, preserving source structural relationships without explicitly injecting source value representations.

(2) AGMR: Attention-guided region-specific modulation.

AGMR and Prompt-to-Prompt’s local blending both use cross-attention responses for semantic region localization. Their subsequent control mechanisms, however, differ.

Prompt-to-Prompt’s local blending uses attention-derived masks to spatially blend source and edited latents, restricting changes to selected regions. AGMR uses the target-subject mask to route source or attribute-edited modulation features to the corresponding image tokens within DiT blocks. This enables continuous attribute control for the selected subject through token-specific conditioning.

Response to Critical Weakness 4.

We agree that the substantial improvements in LPIPS-Alex and DINO in Table 2 mainly come from SAMB. These metrics primarily measure content preservation, and we have revised the wording to avoid attributing these gains entirely to M-space.

With SAMB disabled in both settings, M-space outperforms P-space across all evaluated metrics (below), supporting better attribute control and disentanglement relative to P-space. The full ablation study is presented in Section 4.5 (Ablation Studies).

E: extrapolation; C: continuity.

VariantSAMBE-CLIP ↑E-SigLIP ↑E-BLIP ↑C-CLIP ↑C-SigLIP ↑C-BLIP ↑LPIPS-Alex ↓LPIPS-VGG ↓DINO ↓ID ↓
P-spaceOff0.23890.05910.24720.20470.21370.16470.27930.31730.22100.5230
M-spaceOff0.26220.07880.51690.27350.22070.25240.27060.30660.16000.5035

We also applied SAMB to other comparison methods to assess whether its benefits extend beyond our framework. As shown in Table 4 of the paper, these improvements demonstrate that SAMB also effectively enhances content preservation in other continuous attribute-control methods. Although these baselines also benefit from SAMB, our complete method maintains a favorable overall trade-off across attribute response, continuity, and content preservation. These experiments are reported in Appendix I.

AuthorsOpenReview ↗

Response to Reviewer So9F (Part 2/3): Weaknesses 5-6 & Questions 1-3

Response to Critical Weaknesses 5 and 6.

We thank the reviewer for these observations. We agree that the extrapolation gains over SliderEdit are modest, that SliderEdit achieves the best continuity scores, and that SliderEdit and Kontinuous Kontext achieve lower identity distances. Our results support a competitive overall trade-off without additional training, rather than consistent superiority over training-based methods.

Compared with the training-free baseline FluxSpace, our method achieves better results across all reported metrics in Table 1. Compared with training-based methods, ours achieves the best extrapolation scores across all three vision–language models, together with the lowest LPIPS and DINO distances. Its contribution lies in achieving competitive continuous attribute control without training additional control modules or optimizing attribute directions.

Response to Question 1.

For each method, we selected the maximum effective control strength by testing candidate slider scales and evaluating the corresponding CLIP, SigLIP, and BLIP attribute responses alongside image quality. For our method, all three attribute-response metrics reach their maximum at λ = 18. However, more than 80% of the generated images exhibit hand artifacts at this scale. We therefore set the maximum slider scale to λ = 16 to balance attribute strength and image quality. The other methods were calibrated using the same procedure. Details of the slider-scale selection are provided in Appendix L.

Our evaluation protocol uses the largest control strength at which each method maintains acceptable image quality to measure its maximum effective attribute response. The remaining metrics are computed within the corresponding usable control range. Since numerical slider scales are not directly comparable across methods, applying the same calibration procedure to determine each method’s usable range provides a fairer comparison than imposing a common numerical limit.

Response to Question 2.

(1) SAMB blending weight. We set η = 0.6. The ablation study of this weight is provided in Appendix N.

(2) Number of prompt pairs. We use N = 100 prompt pairs per attribute. Following CLIP’s use of prompt ensembling to reduce sensitivity to wording, we construct a reference M-space direction for each attribute by averaging directions from 10,000 prompt pairs. For each of 25 attributes, we repeatedly sample subsets of different sizes and compute the cosine similarity between each subset-derived direction and its reference. We average over repeated subsets within each attribute and then give all attributes equal weight. At N = 100, the mean cosine similarity is approximately 0.992, indicating stable direction estimation across the sampled subsets. Details are provided in Appendix L.

(3) Choice of prompts. We construct diverse prompt pairs in which only the target-attribute phrase changes, while the remaining description is held fixed within each pair. For example, an age-control pair is “a young man in a park” and “an elderly man in a park.” Across pairs, we vary the subject descriptions, non-target attributes, and contexts to reduce dependence on any single wording or scene. The subset-resampling experiment above evaluates sensitivity to the selection of prompt pairs from this pool and supports the stability of the averaged direction under these variations.

Response to Question 3.

Both methods influence FLUX through the pooled global text embedding produced by CLIP. The difference lies in how this embedding is modified.

For our P-space editing, we first extract the pooled text embedding and then directly move it along a precomputed, fixed attribute direction:

e′=e+αd,e' = e + \alpha d,

where the same direction dd is used for a given attribute across different prompts, and the control strength determines the travel distance α\alpha. This operation is training-free and is applied after text encoding.

In contrast, Text Slider first modifies the CLIP text encoder using a trained LoRA adapter and then uses the modified encoder to produce a new text embedding. Thus, both methods affect FLUX through the pooled CLIP representation, but P-space editing directly applies a fixed semantic direction to the embedding, whereas Text Slider changes the text encoder itself through a learned LoRA module.

AuthorsOpenReview ↗

Response to Reviewer So9F (Part 3/3): Questions 4-6

Response to Question 4.

We are currently computing quantitative results for SD3.5 and real-image editing. These evaluations are still in progress, and we will report the results once they are complete.

Under our default parameter settings, the measured mask leakage rate is 2.49%. Separately, we have added a quantitative evaluation of attribute leakage in two-person scenes. We compute the CLIP, SigLIP, and BLIP extrapolation scores on the non-target person using the target-attribute description; lower scores indicate less unintended attribute influence.

Our method achieves lower leakage scores than both compared baselines across all three metrics, supporting effective instance-level control with reduced attribute leakage to non-target subjects. The results are shown below.

DimensionMetricCACSAEditOurs
Target-attribute response ↑CLIP0.23130.22520.2444
SigLIP0.06530.05170.0764
BLIP0.25220.21980.3230
Continuity ↑CLIP0.17790.16700.2091
SigLIP0.17240.15070.2019
BLIP0.16050.14890.2085
Content-preservation distance ↓LPIPS-Alex0.27190.15040.0454
LPIPS-VGG0.30040.18220.0721
DINO0.12080.06020.0421
Face-ID0.36620.31990.2829
Leakage ↓CLIP0.06380.05240.0256
SigLIP0.00710.00590.0029
BLIP0.05590.03480.0193

Response to Question 5.

We measured the inference overhead on an NVIDIA L20 GPU using BF16 precision, 1024 × 1024 resolution, and 28 sampling steps. Our method jointly processes the source and edited branches with batch size 2 to produce one edited image, compared with batch size 1 for standard FLUX.1-dev generation.

As shown below, this configuration requires approximately 2.28 times the runtime, while peak GPU memory increases by 0.43 GiB (1.9%).

Latency covers denoising and decoding; peak memory is measured during denoising.

MethodBatch sizeLatency (s)Peak memory (GiB)
Standard FLUX.1-dev125.6422.62
Ours (source + edited)258.4823.05

Response to Question 6.

We agree that automatic metrics alone may not fully capture perceptual differences in continuity and attribute fidelity. We therefore added a user study to complement the quantitative evaluation.

We compare our method with seven baselines: Continuous Attribute Control, Text Slider, FluxSpace, SAEdit, Concept Slider, SliderEdit, and Kontinuous Kontext. For each evaluation item, participants select the method they consider best among the eight candidates. We collected a total of 720 votes across three perceptual aspects: extrapolation, which measures the strength and fidelity of the desired attribute change; continuity, which measures the smoothness of the change along the control trajectory; and disentanglement, which measures the preservation of attribute-irrelevant content.

Our method receives the highest human preference in all three aspects: 38.75% for extrapolation, 44.17% for continuity, and 32.08% for disentanglement. It also achieves the highest average preference, 38.33%, compared with 15.97% for the strongest baseline. Detailed results are provided in Appendix M.

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

This paper studies training-free continuous attribute control in text-to-image Diffusion Transformers (DiTs). The authors define three spaces: P-space, which contains pooled text embeddings; M-space, which contains global modulation conditions after an MLP; and M+-space, which consists of the actual modulation parameters used by individual layers. For a given attribute, the method collects prompt pairs that differ only in the target attribute and averages the differences between their representations to obtain an attribute direction. During inference, this direction is linearly traversed from the source conditioning, enabling single-attribute control, bidirectional control, and multi-attribute composition.

Critical strengths
  1. The research problem is well-motivated and practically valuable. Leveraging conditioning representations within a pretrained DiT to achieve reusable continuous control without updating model parameters is a clear and meaningful research direction.
  2. The method is simple, interpretable, and offers a degree of novelty. Averaging representation differences across prompt pairs is not particularly complex in itself, but systematically applying this approach to the shared global modulation condition in DiTs and comparing three candidate spaces constitutes a reasonable research design.
  3. The core feasibility is supported by comprehensive empirical evidence. Figure 3, Figure 5, and the Appendix demonstrate control over a wide range of attributes, including age, expression, hair, accessories, body shape, and style, as well as unidirectional, bidirectional, and multi-attribute compositional control.
Critical weaknesses
  1. The novelty claim needs further justification. The authors overlook closely related works, including FreeSliders and FluxSpace. They should discuss these works and clarify how their approach differs. While exploiting the shared global modulation condition and comparing P-, M-, and M+-spaces may constitute novel contributions, the broader claim of being the first training-free approach for continuous semantic attribute control in text-to-image DiTs may be insufficiently substantiated.
  2. Generalizability across DiT models. The authors provide some results on SD3.5 in addition to FLUX. It would be interesting to evaluate whether the method generalizes to more recent DiT models, such as Z-Image and Qwen-Image. More qualitative and quantitative results on these models would help better establish the generalizability of the proposed approach.
  3. Separating the contributions of M-space and SAMB. In Table 2, removing SAMB eliminates the method’s clear advantage in preservation metrics, particularly ID and DINO. The current evidence supports the effectiveness of the combined M-space and dual-branch attention preservation system, but does not establish that M-space itself is more disentangled.
Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer VXyj : Weaknesses 1–3

Response to Critical Weakness 1.

We thank the reviewer for pointing out these closely related works. We acknowledge that we overlooked these related works and that our claim of being the first training-free approach for continuous semantic attribute control in text-to-image DiTs was not sufficiently substantiated. We have therefore removed this claim and revised the Introduction and Related Work to explicitly discuss FreeSliders and FluxSpace, clarify how our approach differs, and more precisely define our contributions.

Response to Critical Weakness 2.

We thank the reviewer for highlighting the importance of generalizability across DiT models. Our current formulation requires a specific architectural interface: a pooled text condition that contributes to a globally shared modulation representation. Both FLUX and SD3.5 provide this interface, allowing us to construct attribute directions in their respective M-spaces without additional training. The SD3.5 results in Appendix D demonstrate applicability beyond FLUX to another DiT architecture with a similar conditioning pathway.

Z-Image and Qwen-Image do not expose the corresponding pooled-text-to-modulation pathway required by our method. Consequently, our current training-free formulation cannot be directly applied to these models without modifying the method or introducing additional model-specific adaptations.

More broadly, our findings highlight shared text-conditioned modulation as a useful interface for continuous semantic control. We hope this perspective can inform future DiT architecture design by motivating explicit, accessible conditioning interfaces for controllable generation.

Response to Critical Weakness 3.

We agree with the reviewer that the clear gains in content-preservation metrics, particularly ID and DINO, are achieved by the complete method, with SAMB playing the primary role. We have revised the wording in the manuscript to avoid attributing these gains entirely to M-space.

To isolate the contribution of M-space, we compare P-space and M-space with SAMB disabled in both settings. The results are shown below, together with those of our complete method. The full ablation study is presented in Section 4.5 (Ablation Studies) of the paper. Under the same setting, M-space achieves higher scores on all three attribute-response metrics and all three continuity metrics, while reducing all four content-preservation distances relative to P-space.

E: extrapolation; C: continuity.

VariantSAMBE-CLIP ↑E-SigLIP ↑E-BLIP ↑C-CLIP ↑C-SigLIP ↑C-BLIP ↑LPIPS-Alex ↓LPIPS-VGG ↓DINO ↓ID ↓
P-spaceOff0.23890.05910.24720.20470.2137‾\underline{0.2137}0.16470.27930.31730.22100.5230
M-spaceOff0.26220.07880.51690.2735‾\underline{0.2735}0.22070.25240.2706‾\underline{0.2706}0.3066‾\underline{0.3066}0.1600‾\underline{0.1600}0.5035‾\underline{0.5035}
Ours (M-space)On0.2506‾\underline{0.2506}0.0758‾\underline{0.0758}0.4577‾\underline{0.4577}0.27360.21170.2371‾\underline{0.2371}0.09710.14070.05020.3322

Thank you for the authors' response. My concerns have been alleviated, and I am therefore willing to increase my score. However, I still have one remaining concern: many of the latest text-to-image generation models no longer incorporate the pooled-text-to-modulation pathway required by the proposed method. Have you considered adapting the proposed method to more modern generative models?

AuthorsOpenReview ↗

Response to Follow-up Question

We thank the reviewer for the positive feedback.

For the latest text-to-image generation models that no longer incorporate the pooled-text-to-modulation pathway required by our method, one adaptation we are considering is to introduce a lightweight text-conditioned modulation branch implemented as a small MLP. We would keep the original model frozen and train only the newly introduced small MLP, aiming to establish a connection between text semantics and modulation features while preserving the original model’s generation capabilities. We could then investigate constructing attribute directions in the modulation space exposed by this learned interface to enable continuous attribute control. This extension would require additional branch training and has not yet been experimentally validated; it is a direction we plan to explore.

Reviewer_B3zwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds continuous attribute sliders (age, beard, smile, style) to FLUX.1-dev with no training. It takes the average difference between prompt pairs in the post-MLP global conditioning vector (M-Space) and adds a scaled copy on denoising steps 3 to 27. A parallel unedited branch supplies self-attention maps that get blended into the edited run (SAMB). Cross-attention plus a flood fill restricts the edit to one subject (AGMR). On Table 1 it gets the highest edit strength (0.46) and the lowest scene drift (0.10) against trained slider and instruction-editor baselines.

Critical strengths
  1. Zero-training sliders that beat trained ones: A new slider costs a few text-encoder calls instead of a LoRA training run. It still gives stronger edits and less than half the scene drift of Concept Sliders in Table 1, so per-attribute training looks unnecessary for this task.
  2. A useful finding about where to edit: Showing that the post-MLP M-Space vector carries clean, roughly linear attribute directions, and that steps 0 to 2 should stay untouched, is a simple result. Other people building DiT editing tools can reuse it.
Critical weaknesses
  1. Scene preservation comes from SAMB: Paper credits M-Space for the 53.5% drop in drift. Table 2 shows drift goes from 0.10 to 0.27 without SAMB, which is worse than the trained baselines. Running Concept Sliders and the other baselines with SAMB would show how much M-Space adds by itself.
  2. Compute cost is not reported: The source and edited branches run side by side for 25 steps, about 53 transformer passes against 28 for the baselines, and 81 for real photos. Wall-clock time and peak GPU memory on named hardware would show the real cost per edit.
  3. Multi-subject routing is only shown qualitatively: AGMR keeps the top 20% of tokens as candidates, so close-up subjects get clipped. It also cannot separate two people with the same noun, and the evidence is two example scenes. A small quantitative multi-subject test would back the per-instance claim.
Questions for authors
  1. With SAMB added to Concept Sliders, does the drift gap to your method stay the same?
  2. What are the latency and peak memory per edited image, compared with a plain 28-step FLUX run?
  3. What values did you use for the number of prompt pairs, the SAMB blend weight, and the maximum slider scale?
2 comments
AuthorsOpenReview ↗

Response to Reviewer B3zw (Part 1/2): Weaknesses 1–2 & Questions 1–2

Response to Critical Weakness 1 and Question 1

We thank the reviewer for raising this concern regarding the respective contributions of SAMB and M-space to content preservation. We agree that the 53.5% reduction in drift should be attributed to the complete method rather than entirely to M-space. SAMB plays an important role in preserving content, and we have revised the wording in the paper accordingly.

To analyze the independent contribution of M-space, we conducted a controlled experiment with SAMB disabled in both P-space and M-space (Table 1). Compared with P-space, M-space achieves lower LPIPS-Alex, LPIPS-VGG, DINO, and identity distances. These results indicate that M-space itself provides better attribute control and disentanglement than P-space, and that its advantages do not come entirely from SAMB.

Moreover, after adding SAMB to Concept Sliders, the gap between Concept Sliders and our method indeed becomes substantially smaller, but it remains (Table 2). For LPIPS-Alex, before adding SAMB to Concept Sliders, Concept Sliders and our complete method obtain scores of 0.2087 and 0.0971, respectively, corresponding to a gap of 0.1116. With SAMB added, Concept Sliders improves to 0.1139, while our method remains at 0.0971, reducing the gap to 0.0168. Therefore, SAMB substantially improves content preservation, while the controlled comparison with P-space confirms that M-space also provides benefits in attribute control and disentanglement.

Supporting Tables

E: extrapolation; C: continuity.

Table 1. Controlled comparison of P-space and M-space with SAMB disabled in both settings.

SpaceSAMBE-CLIP ↑E-SigLIP ↑E-BLIP ↑C-CLIP ↑C-SigLIP ↑C-BLIP ↑LPIPS-Alex ↓LPIPS-VGG ↓DINO ↓ID ↓
P-spaceOff0.23890.05910.24720.20470.21370.16470.27930.31730.22100.5230
M-spaceOff0.26220.07880.51690.27350.22070.25240.27060.30660.16000.5035

Table 2. Comparison of Concept Sliders with and without SAMB and our method. Results for the other baselines augmented with SAMB are reported in Appendix I.

MethodSAMBE-CLIP ↑E-SigLIP ↑E-BLIP ↑C-CLIP ↑C-SigLIP ↑C-BLIP ↑LPIPS-Alex ↓LPIPS-VGG ↓DINO ↓ID ↓
CSOff0.24460.06490.36280.21780.21230.21970.20870.24110.10460.4115
CSOn0.23860.05750.30220.20870.17760.18890.11390.14810.08610.3825
OursOn0.25060.07580.45770.27360.21170.23710.09710.14070.05020.3322

LPIPS-Alex gap to CS. With SAMB disabled for CS, the gap to our complete method is 0.2087 - 0.0971 = 0.1116. After enabling SAMB for CS, it becomes 0.1139 - 0.0971 = 0.0168.

Response to Critical Weakness 2 and Question 2

We thank the reviewer for raising the concern about computational cost. We measured latency and peak GPU memory on an NVIDIA L20 GPU using BF16 precision, 1024 x 1024 resolution, and 28 sampling steps (Table 3). A standard FLUX.1-dev run with batch size 1 takes 25.64 seconds for denoising and decoding, with peak denoising memory of 22.62 GiB. Our method jointly processes the source and edited branches with batch size 2 to produce one edited image, taking 58.48 seconds with peak denoising memory of 23.05 GiB. Thus, our method requires approximately 2.28 times the latency, with a peak memory increase of 0.43 GiB (1.9%). We have added these measurements to the revised manuscript in Appendix J.

Table 3. Computational cost per output image. Latency covers denoising and decoding; memory is measured during denoising.

MethodBatch sizeLatency (s)Peak memory (GiB)
Standard FLUX.1-dev125.6422.62
Ours (source + edited)258.4823.05
AuthorsOpenReview ↗

Response to Reviewer B3zw (Part 2/2): Weaknesses 3 & Questions 3

Response to Critical Weakness 3.

We thank the reviewer for raising these concerns. As detailed in Appendix A, AGMR constructs a subject mask from image-to-target-text attention responses using two quantile thresholds. Under the default settings, the seed threshold (0.95) selects approximately the top 5% of image tokens by response as high-confidence starting points. The candidate threshold (0.80) selects approximately the top 20% of image tokens by response as the candidate region, within which these seeds can expand through eight-neighbor region growing; the largest connected component is retained as the final mask. Thus, the default candidate region can limit coverage of close-up subjects. In our experiments, the default 20% candidate ratio is sufficient for most two-person images. Both the seed and candidate quantile thresholds are adjustable and can be adapted to different subject sizes and scene layouts.

To provide quantitative evidence, we conducted a small-scale comparison with CAC and SAEdit in two-person scenes. The results are shown below.

DimensionMetricCACSAEditOurs
Target-attribute response ↑CLIP0.23130.22520.2444
SigLIP0.06530.05170.0764
BLIP0.25220.21980.3230
Continuity ↑CLIP0.17790.16700.2091
SigLIP0.17240.15070.2019
BLIP0.16050.14890.2085
Content-preservation distance ↓LPIPS-Alex0.27190.15040.0454
LPIPS-VGG0.30040.18220.0721
DINO0.12080.06020.0421
Face-ID0.36620.31990.2829
Leakage ↓CLIP0.06380.05240.0256
SigLIP0.00710.00590.0029
BLIP0.05590.03480.0193

We measure leakage using the CLIP, SigLIP, and BLIP extrapolation scores computed on the non-target person with the target-attribute description; lower values indicate less unintended attribute influence. Our method achieves higher target-attribute response and continuity scores across all three vision–language models, lower distances on all four content-preservation metrics, and the lowest leakage scores. These results provide quantitative support for instance-level control in the evaluated two-person scenes.

We also acknowledge that FLUX’s internal attention maps generally cannot reliably distinguish two people described by the same noun. Such cases require an external detection or segmentation model to obtain a reliable target-subject mask. We have clarified this limitation in the revised manuscript (Appendix K).

Response to Question 3.

We use N = 100 prompt pairs, a SAMB blending weight of η = 0.6, and a maximum slider scale of λ = 16.

To evaluate the effect of the number of prompt pairs, we conducted additional experiments on 25 attributes, using the directions obtained from 10K prompt pairs as references. With 100 prompt pairs, the average cosine similarity between the estimated directions and the reference directions is approximately 0.992, supporting the stability of direction estimation under this setting.

The ablation of the SAMB blending weight is reported in Appendix N. We select η = 0.6 to balance attribute response, continuity, and preservation of attribute-irrelevant content.

For the maximum slider scale, we evaluated each candidate value of λ and measured the corresponding CLIP, SigLIP, and BLIP attribute responses. All three metrics reach their maximum at λ = 18. However, more than 80% of the generated results exhibit hand artifacts at this scale. We therefore set the maximum slider scale to λ = 16 to balance attribute response and image quality. We selected the maximum slider scales for the other comparison methods using the same procedure to ensure a fair comparison. Detailed results are provided in Figures 17 and 18 of Appendix L.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.