acceptodds
Under review as a conference paper at ICLR 2027

AcFlow: Controlling Text-to-Image Diffusion Transformers via Learned Conditional Activation Flow

Abstract

Text-to-image diffusion transformers (DiTs) are powerful generators, yet direct prompting provides limited control interface for style intensity and can fail to suppress unwanted concepts. To enable these controls, we introduce AcFlow, an inference-time controller that transports intermediate layer image-token activations through a learned concept-conditioned velocity field while keeping the base DiT frozen. A textual concept description specifies the desired intervention, while the integration horizon provides a continuous control parameter. The field produces token-varying, activation-dependent updates. With parameters shared across concepts within each task family, one field covers over 15,000 style descriptions or over 1,000 suppression concepts, and generalizes to concepts unseen during training without per-concept fitting. On style control, AcFlow achieves the best style–content trade-off among the evaluated baselines in the high-style-alignment regime. At a fixed operating point, AcFlow attains style–content alignment of 0.5365/0.2860, compared with 0.4397/0.2684 for the baseline with the highest style alignment. On concept suppression, AcFlow reduces the fraction of images showing the concept from 95.3%/82.1% to 41.6%/40.5% on held-in/held-out concepts, including cases where deleting them from the prompt fails to remove them. Our analyses support the learned velocity field as an adaptive control mechanism, with update directions varying across tokens and depending on their activation states. Our code is available at https://anonymous.4open.science/r/AcFlow-review.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Models, Controllable Generation, Activation Steering, Continuous Control

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ez94OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AcFlow, a trained controller to produce velocity vectors for each image token that is used to steer the diffusion at each timestep to manipulate image generation. The method is implemented using concept cross-attention with queries as image activations and keys/values and concept text embeddings followed by self-attention + feedforward operations. The controller is trained with distillation, a frozen model with the trainable controller (student) operating on the source prompt should reproduce the denoising prediction of the same frozen model operating on a target prompt (e.g., with the concept removed, added). Experiments are conducted on style and concept supression tasks.

Critical strengths
  • The paper introduces strength controlling in text-driven image editing, which is interesting and needed
  • The method is simple and elegant
  • The paper overall is well-written and it is an interesting read
  • It is nice to see the authors use the simple Euler's method for estimating the differential equation - rather than complicated solvers.
  • I also liked the idea of using activation flow rather than fixed steering, since the velocity can change with the current hidden state, using a fixed direction is not ideal.
Critical weaknesses
  • The paper misses a very relevant work and borrows ideas from it, without comparing to it neither citing it [R1]. The reference paper does exactly the same thing -- at a timestep t (defined according to temporal dynamic analysis) they switch the prompt from the source to the target, manipulating style, removing or adding concepts. [R1] does not even require training, and operates completely at inference time. The proposed method on the other hand requires training, and it actually trains on [R1] method. So it is unclear why we need the proposed method if R1 (or the training of the controller in the authors work - which is fully borrowed from [R1]) directly provides the ground-truth and in no training. The paper further borrows concepts from [R1] such as using VLMs to judge if a concept is present or not. Again without citing [R1].
  • The method degrades original content as the strength increases (Table 2 and Figure 5 and Figure 6, and 40% failures of concept supression). This is a huge downside, because if the only novelty and advantage of the method is strength control, then this comes at a huge expense on content degradation.
  • There are many works on text-driven image editing that perform editing in a completely training-free manner [R2, R3, R4, R5, R6, R7]. The paper does not compare with them and does not provide the advantage over these methods. Why would I train if I have other training-free methods?
  • The method goes somehow opposite of the current advancamaents in diffusion/timestep distillation to reduce the number of timesteps. With 28 denoising timesteps, the proposed method requires 28 × 3 = 84 velocity-field evaluations per generated image. I dont think that's a clever solution.

[R1] Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventions
[R2] Localizing Knowledge in Diffusion Transformers
[R3] Training-Free Layout Control with Cross-Attention Guidance
[R4] Stable Flow: Vital Layers for Training-Free Image Editing
[R5] Prompt-to-Prompt Image Editing with Cross Attention Control
[R6] Training-Free Multi-Concept Image Editing
[R7] SliderEdit: Continuous Image Editing with Fine-Grained Instruction Control\

Questions for authors

The paper is interesting but has major flaws. See weaknesses. For that reason, sadly enough, my decision will be negative.

3 comments

We thank the reviewer for noting that the method is simple and elegant and that activation flow is preferable to fixed steering. The review contains three factual mistakes, which we correct first; we then address the other concerns one by one. Line numbers refer to the revised PDF.

Comment 1: "it actually trains on [R1] method"; "the training of the controller in the authors work - which is fully borrowed from [R1]".

Fact: [R1] switches the prompt at a chosen denoise step along the trajectory; AcFlow is trained by velocity distillation, not by prompt switching. Each training step only matches velocities at a single noise level (Eqs. 8–11, lines 232–245), and no denoising trajectory is run. This distillation target is that of Concept Ablation (Kumari et al., 2023) and, with negative guidance, of ESD (Gandikota et al., 2023), published more than two years before [R1].

Comment 2: "The paper further borrows concepts from [R1] such as using VLMs to judge if a concept is present or not."

Fact: Judging concept presence with a VLM is standard practice in text-to-image evaluation and predates [R1], e.g. TIFA (Hu et al., 2023) and VQAScore (Lin et al., 2024). We use the judge model of the Holistic Unlearning Benchmark (HUB; Moon et al., 2025), and App. H applies HUB's official evaluator directly.

Comment 3: "There are many works on text-driven image editing that perform editing in a completely training-free manner [R2, R3, R4, R5, R6, R7]."

Fact: [R2] and [R7] are not training-free. Their abstracts state: "our localized fine-tuning approach enables efficient and targeted updates" ([R2], for personalization and unlearning) and "our method learns a single set of low-rank adaptation matrices" ([R7]).

References ([R1]–[R7] are the references of the review.)

  • Gandikota R, Materzynska J, Fiotto-Kaufman J, et al. Erasing concepts from diffusion models[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2023: 2426-2436.
  • Hu Y, Liu B, Kasai J, et al. TIFA: Accurate and interpretable text-to-image faithfulness evaluation with question answering[C]//2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2023: 20349-20360.
  • Kumari N, Zhang B, Wang S Y, et al. Ablating concepts in text-to-image diffusion models[C]//2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2023: 22634-22645.
  • Lin Z, Pathak D, Li B, et al. Evaluating text-to-visual generation with image-to-text generation[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 366-384.
  • Moon S, Lee M, Park S, et al. Holistic unlearning benchmark: A multi-faceted evaluation for text-to-image diffusion model unlearning[C]//2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025: 16356-16366.

Concern: why train a field when PCI needs no training? PCI (Görgün et al., 2026) switches from the user's prompt to a target prompt after k of the 28 denoising steps, which requires a ground-truth target prompt for every image as an extra input, whereas AcFlow takes only the user's prompt and a concept text. The comparison is therefore not like-for-like, and PCI has two limits: (1) its strength is not continuous, as the style results show; (2) it needs a clean target prompt, which often does not exist, as for suppression. We agree that PCI should have been cited; it is now in Sec. 2 (lines 113–115). We ran it on both tasks under our protocols (same cells, initial noise and scorers).

(1) Style: PCI's control is not continuous. Here the target is the styled prompt, which the paper reports as a reference (Table 2), and PCI's strength is discrete: only the first few switch steps matter (k ≥ 6 moves style only from 0.154 to at most 0.198). Its usable levels and AcFlow's horizons on the Fig. 3 grid:

MethodSettingStyle
FLUX.1-dev–0.154
PCIk = 4 / 3 / 2 / 1 / 00.245 / 0.293 / 0.363 / 0.476 / 0.548
AcFlowT = 0.8 / 1.0 / 1.2 / 1.4 / 1.6 / 1.8 / 2.0 / 2.50.353 / 0.402 / 0.450 / 0.480 / 0.500 / 0.519 / 0.536 / 0.562

PCI offers only these five levels, with jumps of up to 0.11 in style between neighbours; finer settings would need more denoising steps, and nothing goes beyond the styled prompt. AcFlow's horizon is a real number: any T is valid, and T = 2.5 goes beyond the styled prompt (0.562 vs 0.548). Even where a target prompt exists, PCI therefore offers a few fixed settings rather than a continuous strength.

(2) Suppression: a clean target prompt often does not exist. The target must be a clean prompt that no longer implies the concept, and such a prompt is hard to write: deleting the concept's words often leaves context that still implies it. This is the difficulty reported in Sec. 4.2 (lines 363–366; examples in App. C): on our IP2P set, even the concept-free caption still yields the concept for 28.9%/15.6% of prompts. On those prompts, PCI is bounded by its target: it removes the concept in 25.6/30.4%, against 60.5/73.9% for AcFlow at T = 3 (95% CIs [14.0, 39.5] / [13.0, 47.8] vs [46.5, 74.4] / [56.5, 91.3]). Negative prompting, the inference-only method with the same input as AcFlow, collapses before it suppresses (31.0/36.6% at MUSIQ −29).

References

  • Görgün A, Sammani F, Deligiannis N, et al. Temporal concept dynamics in diffusion models via prompt-conditioned interventions[C]//International Conference on Learning Representations. 2026, 2026: 66640-66669.

Concern: training-free editing methods [R2–R7]. The review groups six methods that address different tasks; only one of them applies to our setting, text-to-image generation with a DiT from text input alone. [R2] and [R7] trains (factual mistake 3). [R6] and [R7] edit a given input image. [R3] and [R5] manipulate the cross-attention layers of U-Net models, which DiTs replace by joint updates of text and image states (lines 51–53); [R3] also needs bounding boxes. The remaining one, Stable Flow [R4], is the FLUX-native training-free editor, and we ran it on our benchmarks: it runs a parallel source branch (+84% latency), reaches style 0.274 on the style grid (FLUX 0.154; AcFlow T = 0.8: 0.353), and removes the concept in 21.8/26.8% of IP2P images (AcFlow T = 2: 57.0/54.2%). The paper also compares against the training-free negative prompting and SLD (Schramowski et al., 2023) in App. G/H.

Concern: 84 field evaluations versus step distillation. Reducing the number of steps targets latency, and AcFlow adds little: +5.0% FLOPs and +6.1% latency in total (App. A, Table 4, lines 723–734), because each evaluation is a single transformer block on the image tokens (0.30T FLOPs, 1.7% of one DiT forward pass; lines 706–708). The field is evaluated N times per denoising step, so its share of the cost stays about the same when the number of steps is distilled.

Concern: content decreases as strength increases. CLIP penalizes correct stylization: the ground-truth MegaStyle images, i.e. the intended outputs for the same content prompts and style descriptions, score 0.290 CLIP versus 0.302 for unstylized FLUX. At T = 2, AcFlow is 0.022/0.018 CLIP below unstylized FLUX (held-in/held-out), and the ground truth itself is already 0.012/0.014 below. VQAScore (Lin et al., 2024) even rates the ground truth above unstylized FLUX (0.929 vs 0.856). What matters is how much content a method keeps at a given style, and in the high-style regime AcFlow keeps the most among the methods compared in the paper (Table 2 and Fig. 3, lines 270–276). The "40% failures" cited in the review is not a content measure: it is the share of images in which a strict judge still finds the concept at T = 2, down from 95.3%/82.1%, and the content cost of this suppression is about 0.015 CLIP and 2 MUSIQ points (lines 358–361). The remaining change of composition is stated as a limitation (lines 483–485).

References ([R1]–[R7] are the references of the review.)

  • Lin Z, Pathak D, Li B, et al. Evaluating text-to-visual generation with image-to-text generation[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 366-384.
  • Schramowski P, Brack M, Deiseroth B, et al. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023: 22522-22531.
Reviewer_NGGtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AcFlow, a text-conditioned activation controller for frozen text-to-image diffusion transformers. A learned velocity field modifies intermediate image-token activations, while its integration horizon controls intervention strength. The controller is trained by matching target-prompt velocity predictions from the frozen backbone. Separate task-specific controllers are evaluated for style control and concept suppression. Experiments examine generalization to unseen concepts, applicability to another backbone, computational overhead, and the directional structure of the learned interventions.

Critical strengths
  1. The shared conditional interface has practical value. AcFlow supports multiple concepts without fitting a separate controller for each one. The held-out evaluations provide evidence of generalization, and the default configuration increases batch-one latency from 3.58 to 3.80 seconds, suggesting a reasonable computational trade-off.

  2. The intervention analysis provides useful mechanistic evidence. The norm-matched rank-one controls retain substantially less suppression than the full updates, while the off-axis component approximately recovers their effect. This supports the importance of token-dependent directions beyond simply increasing intervention magnitude.

Critical weaknesses
  1. The comparative advantage over existing text-conditioned control remains unclear. FluxSpace is cited but not evaluated, and the direct styled-prompt reference outperforms the default AcFlow setting in Table 2. Although this reference does not provide an equivalent continuous interface, it makes comparison with a continuous conditioning-based method important for establishing practical value.

  2. The evaluation protocols are not mapped clearly across the manuscript. The main text describes one fixed description per held-in family, whereas Appendix I.3 specifies eight families crossed with 32 captions. The relationship between these settings and the fine-grained evaluation needs clarification to establish the coverage and reproducibility of the generalization claims.

Questions for authors
  1. Please provide a consistent mapping from each main result to its evaluation split, number of families, descriptions, prompts, and seeds.

  2. How does AcFlow compare with FluxSpace over matched control-strength settings, including generation quality and latency?

3 comments

We thank the reviewer for the constructive review.

W1/Q2. A continuous conditioning-based comparison: FluxSpace over matched strengths, with quality and latency. We ran FluxSpace (Dalva et al., 2025) with its released implementation (linear editing of the attention outputs of the 19 double-stream blocks) on the Fig. 3 grid. We used the same style description as the edit prompt and its released stylization recipe (edits at steps 0–3, pooled scale 0.8, no mask), and swept its strength λ. AcFlow is shown at the horizons that match FluxSpace's style range (T = 0.2–0.6 are new; same checkpoint and cells; VQAScore (Lin et al., 2024) as defined in the general comment):

MethodStyleCLIPVQAScoreMUSIQ
FLUX.1-dev0.1540.3070.92673.9
FluxSpace, λ = 1 / 4 / 8 / 160.177 / 0.176 / 0.206 / 0.2160.307 / 0.308 / 0.306 / 0.3040.917 / 0.921 / 0.909 / 0.88673.5 / 72.4 / 71.8 / 72.3
AcFlow, T = 0.2 / 0.4 / 0.6 / 0.80.168 / 0.219 / 0.296 / 0.3530.306 / 0.306 / 0.305 / 0.3020.911 / 0.910 / 0.926 / 0.91674.4 / 74.6 / 74.8 / 74.5

FluxSpace raises style from 0.154 to at most 0.216 (λ = 16), i.e., 16% of the way to the styled prompt (0.548). In a pilot on training captions, longer edit windows (steps 0–7, 0–13, 0–27) and short "⟨family⟩ style" edit prompts did not raise its style above 0.25 (styled prompt: 0.56 on the same cells). At matched style AcFlow is on par or better (T = 0.4: 0.219 / 0.306 / 0.910 / 74.6, against λ = 16: 0.216 / 0.304 / 0.886 / 72.3) and continues to 0.56 at T = 2.5, while MUSIQ stays within 1.5 points of unsteered FLUX at every horizon (72.4–74.8).

Latency, re-measured in one session on one A100 (App. A protocol, batch 1): FLUX 3.61 s; AcFlow (N = 3) 3.86 s (+7.0%, 1.05× FLOPs; App. A: +6.1%); FluxSpace 3.88 s (+7.5%, 1.05×) with its released schedule, and 5.42 s (+50%, 1.36×) when editing every step.

In the final version we will add the FluxSpace results to the appendix and, since the styled prompt changes the generator's prompt, state our claim as "the best style–content trade-off in the high-style regime among the evaluated methods that leave the generator's prompt unchanged".

W2/Q1. Protocol mapping. The held-in columns of Table 2 and Fig. 3 use the aligned baseline grid. The held-out columns of Table 2 come from the fine-grained held-out set, because the per-style baselines have no held-out entries. We will state this in the caption and add the map in the next comment to App. I.

References

  • Dalva Y, Venkatesh K, Yanardag P. Fluxspace: Disentangled semantic editing in rectified flow models[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 13083-13092.
  • Lin Z, Pathak D, Li B, et al. Evaluating text-to-visual generation with image-to-text generation[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 366-384.
ResultFamilies / conceptsCondition textContent promptsImages per pointInitial noise
Table 2 held-in, Fig. 3, Fig. 158 held-in families (cubism, baroque, art nouveau, ink wash, glitch art, art deco, paper art, psychedelic)one frozen description and reference per family, shared by all methods32 COCO-2017 val captions256one seed per caption (42–73), shared by all methods
Table 2 held-out6 held-out families704 distinct MegaStyle descriptions768 paired MegaStyle prompts (128 per family)768seed-42 stream
Fig. 5, Table 3, Fig. 16held-in 39 / held-out 6 familiesper-row MegaStyle description (1,980 / 704 distinct)paired MegaStyle promptsT = 2: 2,048 / 768 (full); other T: fixed family-balanced subsets 512 / 192seed-42 stream, same for all T
Table 5, Table 7, Fig. 691 training concepts (149 prompts, none shared with training) / 150 unseen concepts (173 prompts)erase <concept>IP2P captionsone per prompt and Tseed-42 stream, same for all T
Training (style)39 families, 512 rows each15,669 distinct descriptions over 19,968 rowspaired MegaStyle prompts––
App. B (Z-Image)as Fig. 5as Fig. 5as Fig. 5as Fig. 5as Fig. 5
App. Gnudityerase nudity931 I2P prompts (sexual); Ring-A-Bell: 109-prompt subset931official per-row I2P seeds
App. H10 HUB characters (9 unseen)erase <character>64 HUB prompts each640canonical HUB per-row seeds

Thank you for the rebuttal. My concerns have been addressed, and I will maintain my original rating.

Reviewer_7ufBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AcFlow introduces a learned controller for continuous style control and concept suppression in text-to-image diffusion transformers. It learns a text-conditioned velocity field that updates intermediate image-token activations while keeping the generator frozen, with the integration horizon controlling intervention strength. Experiments show favorable style–content trade-offs among the evaluated control methods and generalization to concepts unseen during controller training. The paper also evaluates a second backbone and studies how update directions vary across tokens and denoising steps.

Critical strengths
  • A single controller supports multiple concepts without per-concept fitting. Evaluation on held-out style families and suppression concepts supports this capability, with further evidence from the unseen-character results.
  • The style comparisons include strength sweeps and uncertainty intervals, supporting the reported advantage in the high-style-alignment regime. Ablations also show that effective control is possible across several intervention blocks.
  • The token-swap and norm-matched intervention experiments provide useful evidence for the image-token interface and the contribution of directions beyond the dominant shared axis. The default controller also adds relatively modest inference overhead, approximately 6% in the reported batch-one setting.
Critical weaknesses
  • The benefit of the flow formulation over a simpler learned conditional update is not fully established. The rank-one experiments show that projecting the trained field onto its dominant axis loses effectiveness, but do not establish how a simpler controller would perform if trained directly with the same objective. The one-step ablation partly addresses this question, although it also reduces the number of module evaluations. These results support the proposed controller, but leave some uncertainty about which aspects of its design account for the gains.
  • Comparative evidence for open-vocabulary suppression is limited. Appendix C shows that AcFlow can suppress concepts when prompt editing fails, although concept-free prompting performs better overall. The comparisons in Appendices G and H are useful but involve separately trained fields and different tasks. An inference-only baseline on the main suppression dataset would make the practical benefit clearer. Also, the strongest results on the prompt-editing failure subset use T=3T=3, where the aggregate results show greater content and quality degradation; the corresponding trade-off on that subset is not reported.
Questions for authors
  • How does a simpler text-conditioned update trained with the same data and distillation objective compare with AcFlow? This would help clarify the contribution of iterative transport beyond learning a shared conditional controller.
  • A negative-prompting comparison on the InstructPix2Pix evaluation would be informative, since it can use the same source prompt and target concept without per-concept fitting. Does AcFlow offer a better suppression–preservation trade-off in this setting?
  • For the subset where prompt editing fails, could the authors provide content-alignment and image-quality results alongside the suppression rates at T=2T=2 and T=3T=3? This would clarify the cost of the additional successful removals.
2 comments

We thank the reviewer for the careful and constructive review.

W1/Q1. A simpler conditional update trained with the same data and objective. We trained the simplest text-conditioned update of Sec. 3.1 with the default recipe (same data, distillation objective, optimizer, schedule and seed), changing only the parameterization and keeping the same block: the additive update v=d(c;σ)v=d(c;\sigma), a learned text-conditioned direction per noise level that depends on neither the state nor tt. It is a shared conditional controller without iterative transport: the flow reduces to hi+T d(c;σ)h_i+T\,d(c;\sigma), and TT only scales the update. Both fields were trained for 10k updates (the field in the paper: 19k) and evaluated on the Fig. 3 grid (256 cells, same seeds) at the horizons used in the paper:

Field (10k updates)TStyleCLIP
AcFlow0.8 / 1.6 / 2.50.270 / 0.389 / 0.4570.305 / 0.296 / 0.283
Additive d(c;σ)d(c;\sigma)0.8 / 1.6 / 2.50.170 / 0.222 / 0.2730.307 / 0.306 / 0.307

Over the same horizons the additive update moves style much less: at T = 2.5 it reaches 0.273, where AcFlow reaches 0.457. Where their style ranges overlap, the two keep the same content (CLIP within 0.002 at matched style; paired bootstrap intervals include 0). Learning a shared conditional controller with the same data and objective therefore covers only the low-style range at these horizons; the higher-style range of Fig. 3 is reached with iterative transport by a state-dependent field. Table 3 isolates the number of integration steps: a field trained with N = 1 reaches 0.432/0.375 style versus 0.510/0.442 for N = 3. We will add this comparison, with samples, to the appendix of the final version.

W2/Q2. Inference-only baseline on the IP2P set. We ran negative prompting (NP) on the Table 5 prompts, with the same initial noise and judge. NP uses the concept as the negative prompt in FLUX's true-CFG branch, at scales 1.5–5; this doubles the transformer evaluations per step. Columns give held-in / held-out values; "editing fails" is the Table 7 subset where the concept-free prompt still shows the concept (43 / 23 prompts):

Method (transformer cost)Suppr. (%), allSuppr. (%), editing failsCLIPMUSIQ (Δ)
AcFlow T = 1 (1.05×)29.6 / 31.018.6 / 17.40.287 / 0.28370.9 / 72.1 (−1.3 / −0.4)
AcFlow T = 257.0 / 54.241.9 / 39.10.273 / 0.26769.7 / 70.4 (−2.3 / −2.0)
AcFlow T = 364.1 / 58.560.5 / 73.90.254 / 0.25164.8 / 64.6 (−7.0 / −7.6)
NP, true-CFG 3.5 (2×)16.2 / 27.59.3 / 34.80.295 / 0.29058.1 / 57.4 (−13.6 / −15.7)
NP, true-CFG 531.0 / 36.625.6 / 56.50.280 / 0.27743.0 / 44.4 (−29.3 / −29.2)

Yes: with the same input (source prompt and concept), AcFlow gives a much better suppression–preservation trade-off than NP. NP needs scale 5 to remove the concept in 31.0/36.6% of images, at which point images collapse (MUSIQ −29); scales 1.5–2.5 remove at most 15%. AcFlow removes 29.6/31.0% at T = 1 with −1.3/−0.4 MUSIQ, and 57.0/54.2% at T = 2 with −2.3/−2.0 MUSIQ. We will add this table to App. C.

Q3. Cost of the additional removals where prompt editing fails. On the Table 7 subset where the concept-free prompt still shows the concept (43 held-in / 23 held-out prompts), CLIP is to the concept-free prompt and Δ is paired against the unsteered image:

TSuppr. (%)CLIP (Δ)MUSIQ (Δ)
241.9 / 39.10.285 (−0.010) / 0.277 (−0.016)69.6 (−1.7) / 71.5 (−0.9)
360.5 / 73.90.256 (−0.038) / 0.263 (−0.029)66.9 (−4.4) / 68.0 (−4.4)

Going from T = 2 to T = 3 removes the concept in 18.6/34.8 more percentage points of these prompts, at an additional −0.028/−0.013 CLIP and −2.7/−3.5 MUSIQ. On this subset the quality cost at T = 3 is smaller than on all prompts that show the concept (−7.0/−7.6 MUSIQ). We will add this table to App. C.

Comments on the paper

2 comments

We really appreciate your review and your reply. Please do not mistake our response for disrespect; we had absolutely no intention of, or cause for, being disrespectful. We are not sure which part of our response came across as disrespectful, but we will be as respectful and modest as we can in our future communication.

We listed the factual mistakes (perhaps better called misunderstandings) simply to establish common ground for our discussion. We understand that reviewers have limited time and may not have noticed every detail. We only wanted to clarify the misunderstandings and mistaken details, and we did so only where they are absolutely critical for this paper.

Below we point to where each of your remaining questions is answered. All numbers come from the paper or from the runs reported in our earlier responses; no new experiments were run.

Training-free methods [R3]–[R6]. We agree that they are training-free; our point concerned their input. [R6] edits a given input image, as you note, and that image fixes the content the edit must keep. In text-to-image generation there is no input image: the content is defined only by the prompt, and the control acts while the image is being generated. [R3] and [R5] act on the cross-attention maps of U-Net models, which FLUX replaces by joint attention over text and image tokens (lines 51–53); [R3] also requires bounding boxes. [R4] is FLUX-compatible, and we ran it on both tasks, generating the source image and editing it in the same pass (our response 3/3): style 0.274 (AcFlow at T = 0.8: 0.353) and suppression 21.8/26.8% (AcFlow at T = 2: 57.0/54.2%), at +84% latency.

PCI's target prompt. We agree that PCI [R1] and AcFlow are competing ways to control concepts in generation; this is why we added PCI to the related work (lines 113–115) and compared with it on both tasks. AcFlow is trained on source–target pairs (Sec. 3.4), but it runs no target trajectory: each training step matches velocities at a single noise level (Eqs. 8–11, lines 232–245), and at inference it receives only the user's prompt and the concept text, also for the 6 held-out style families and 150 held-out concepts. For style, we built its target from the concept text, as you describe: the content prompt followed by the style description (the styled prompt of Table 2). Our point there was the strength: switching at k = 4, 3, 2, 1, 0 gives style 0.245, 0.293, 0.363, 0.476, 0.548, and no setting lies in between or beyond. For suppression, the corresponding target is the concept-free prompt, in most cases the prompt with the concept deleted (App. C), and this is what we gave PCI; generating from it still shows the concept for 28.9%/15.6% of prompts (Sec. 4.2, lines 363–366), and on those PCI removes it in at most 25.6/30.4% of cases.

We also note that constructing clean concept-free target prompts can be difficult for concept suppression, as in noted in Section 4.2, line 363-364 and Figure 12. Figure 12 shows cases where simply removing the concept text from the prompt and supplying it to the model fail to remove the concept in the generated image (equivalent of k=0 for PCI). In these cases, a concept text paired with an original prompt does not directly provide the target prompt consumed by PCI [R1].

CIS-guided selection. We swept the switch step instead of choosing one by CIS. CIS guidance selects which step to use and does not change what a step produces, so any step it selects is covered by the sweep: k = 0–8, 10, 12, 14, 16, 20 and 28 on the style grid, and k = 1–4, 6, 8 and 12 on IP2P, where k = 0 is the reference of Table 5.

Continuous control. Since T is real-valued, the size of each increase is set by the user rather than by the method. Mean style rises at every horizon we evaluated, in steps of 0.2 from T = 0 to 2 (0.154 to 0.536), and reaches 0.562 at T = 2.5 (tables in our response 2/3 and to Reviewer NGGt). Fig. 5 shows the response on held-in and held-out families with 95% bands, together with fixed-noise sequences; Figs. 17–18 and 21 show 22 style sequences at 13 horizons and 10 suppression sequences at 11 horizons. On content at matched style: in the high-style range, AcFlow has the best style–content trade-off among the evaluated methods that leave the generator's prompt unchanged (Table 2 and Fig. 3, lines 270–276). Among the training-free methods, FluxSpace (Dalva et al., 2025) saturates at style 0.216 and AcFlow is on par or better at equal style (response to Reviewer NGGt), and Stable Flow [R4] gives a single setting (0.274). PCI changes the generator's prompt, like the styled-prompt reference of Table 2, and is therefore outside this claim, the scope we stated in our response to Reviewer NGGt.

Suppression: content at T = 3 and for PCI. Table 5 reports T = 3: CLIP 0.254/0.251 and MUSIQ 64.8/64.6, against 0.286/0.283 and 72.2/72.6 without steering; this cost is why the paper uses T = 2 as its operating point (lines 899–903). On the prompts where the concept-free prompt still shows the concept, where we compared with PCI (Δ paired against the unsteered image):

MethodSuppr. (%)CLIP ΔMUSIQ Δ
PCI, k = 1 (its best step)25.6 / 30.4+0.009 / +0.014+0.1 / +0.4
AcFlow, T = 241.9 / 39.1−0.010 / −0.016−1.7 / −0.9
AcFlow, T = 360.5 / 73.9−0.038 / −0.029−4.4 / −4.4

PCI keeps the content, but no switch step removes more; AcFlow removes more at a measured cost, and T sets how much. Against negative prompting, the inference-only method with the same input as AcFlow, the trade-off is clearly better (response to Reviewer 7ufB): NP removes 31.0/36.6% only once images collapse (MUSIQ −29), while AcFlow removes 57.0/54.2% at T = 2 for about 2 MUSIQ points (Sec. 4.2).

Thank you again for your time and your review.

References ([R1]–[R7] are the references of the review.)

  • Dalva Y, Venkatesh K, Yanardag P. Fluxspace: Disentangled semantic editing in rectified flow models[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 13083-13092.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.