acceptodds
Under review as a conference paper at ICLR 2027

CompFlow: Composing Velocity Fields for Multi-Conditional Generation

Abstract

Generating samples that satisfy many conditions simultaneously can be achieved with a surprisingly simple operation: composing conditional velocity fields at inference time. We introduce CompFlow, a flow-matching framework for compositional inference without retraining, architectural changes, or specialized samplers. We show that weighted velocity residuals yield the score of a Product-of-Experts composition for positive weights, while negative weights introduce quotient factors to suppress selected concepts. Beyond this score identity, we derive the exact compositional defect, characterizing when the composed dynamics transport the intended density path. Computable from residuals already evaluated during sampling, this quantity connects transport compatibility to measurable interference between conditional contributions. On CLEVR, a single-object-conditioned model is composed at inference to simultaneously control shape, color, and position for up to five objects, achieving 99.1–86.5% per-object accuracy with 30× fewer network evaluations than the evaluated baselines. With pretrained FLUX.1[dev], CompFlow combines up to five conditions in high-resolution images and improves non-spatial compositional alignment over single-prompt conditioning on T2I-CompBench. The same operator also enables controlled subtraction of contexts, objects, attributes, and styles. Together, these results connect practical inference-time concept arithmetic with a principled characterization of conditional interactions.

Primary area:
generative models
Keywords:
Compositional Generation, Flow Matching, Visual Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yFU4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CompFlow, a training-free framework for multi-conditional generation by composing conditional velocity fields in flow matching models. The method combines conditional velocity residuals in a classifier-free guidance manner and provides a theoretical analysis connecting velocity composition with a Product-of-Experts formulation. The paper further introduces a composition defect metric to characterize the mismatch between composed velocity fields and the desired probability path. Experiments on CLEVR and FLUX demonstrate the effectiveness of the proposed approach for compositional generation and concept manipulation.

Critical strengths

The paper provides an interesting theoretical analysis of composing conditional velocity fields in flow matching models. Extending compositional guidance from diffusion score models to velocity fields is a meaningful direction.

Critical weaknesses
  1. The novelty compared with existing compositional diffusion methods is not sufficiently clear. Weighted residual composition is closely related to classifier-free guidance and score composition, and it remains unclear whether the proposed composition defect captures a unique limitation of flow matching or a more general issue of compositional guidance.
  2. The practical value of the proposed composition defect is unclear. Although it provides a theoretical characterization of conditional interactions, it is not demonstrated whether it can effectively predict failures or improve the sampling process.
Questions for authors
  1. How is the proposed composition defect fundamentally different from the interaction issues already present in existing compositional diffusion methods?
  2. Can the authors demonstrate whether the composition defect can be used to guide adaptive weighting or improve compositional generation quality?
4 comments

We thank the Reviewer for their thoughtful comments and constructive feedback. We appreciate their recognition that the paper provides an interesting theoretical analysis of composing conditional velocity fields, that extending compositional guidance from score models to velocity fields is a meaningful direction, and that the presentation is clear and easy to follow. We address their concerns below and hope these clarifications will help improve their assessment towards acceptance.


W1 / Q1. Novelty, and how the defect differs from the interaction issues of compositional diffusion.

We thank the Reviewer for the opportunity to state our contribution more sharply. The focus of our work is a single question: what can a single flow-matching ODE do? That is, how far composition goes with one deterministic trajectory, no corrector, no resampling, no SDE, and only the m+1m+1 network evaluations needed to evaluate the composed field. This question is the novelty of the paper, and everything in it (the algebra, the defect, the experiments) is an answer to it.

We are aware and agree with the Reviewer that, under Gaussian paths, weighted residual composition is related to classifier-free guidance and to score composition, and the original manuscript stated this (lines 45–47). Our contributions are the following.

1. An algebra of concepts for flow matching. We show that a single rule in velocity space composes an arbitrary number of conditions with signed weights: positive weights add concepts, negative weights subtract them, and the result is one ODE of a pretrained model, with no training and no architectural change (Prop. 2.2 and Cor. 2.3). To the best of our knowledge, the flow-matching literature has no such algebra: we are not aware of any work that composes an arbitrary number of conditions, with addition and subtraction, on flow-matching models.

2. A single ODE is enough in practice. Existing compositional diffusion methods treat the inexactness of plain composition as something to repair in the sampler: Du et al. (2023) add MCMC steps, and Skreta et al. (2025) reweight and resample populations of SDE particles. We take the opposite route and show that the correction-free ODE is already competitive. We kindly invite the Reviewer to consider this empirical evidence, which we regard as the core of the paper:

  • On the CLEVR position task, CompFlow matches or exceeds Composable Diffusion and correction-based samplers (Tab. 1).
  • We extend CLEVR beyond the position-only conditioning of Du et al. (2023) to joint shape, color and position, with up to five objects (Tab. 2).
  • The same rule transfers unchanged to a billion-scale pretrained text-to-image model, where a plain ODE adds and removes concepts (Sec. 3.2). In the new App. C.3, CompFlow removes a concept that FLUX adds by default in 62% of the images, against 1% for negation inside the prompt.

3. An exact account of what the single ODE does. This is where the composition defect differs from the interaction issues known in compositional diffusion (Q1). Those works observe that summed scores are not the score of the product, and add a corrector. The defect instead characterizes the sampler that has no corrector: Prop. 2.4 gives its residual in closed form, for any number of concepts and for signed weights, using only quantities available at inference time. It states when no correction is needed, namely if and only if DtD_t is constant in space, which holds under conditional independence (Cor. 2.5). It also separates two sources of error (Eq. 12): a self term, which recovers the known inexactness of classifier-free guidance away from unit weight, and a cross term λiλj⟨δi,δj⟩\lambda_i\lambda_j\langle\delta_i,\delta_j\rangle, which is the interaction between distinct concepts and can be localized in the image (Figs. 3 and 12).

New results. To make this account complete, the revised manuscript adds two results (new Cor. 2.6 and Obs. 2.7, proofs in App. A.5). The defect integrated along a trajectory,

L=∫01σtt Dt(Xt) dt,L=\int_0^1\frac{\sigma_t}{t}\,D_t(X_t)\,dt,

is exactly the log density ratio between the PoE target p~1\tilde p_1 and the law q1q_1 of the sampler, up to a constant:

\qquad \mathrm{KL}(q_1\,\Vert\,\tilde p_1)=\log\mathbb E\big[e^{L}\big]-\mathbb E[L]\le\tfrac18\big(\sup L-\inf L\big)^2 .$$ The single ODE therefore reports its own distance from the target, with no additional network evaluation, and this distance is bounded by the interference between concepts: weakly interfering concepts give a nearly exact sampler. To the best of our knowledge, these results have not been studied for flow-matching ODEs. We validated both on closed-form mixtures (new App. D).

W2 / Q2. Practical value of the defect: predicting failures and improving sampling.

We thank the Reviewer for this question, which motivated the main additions to the revised manuscript. We answer its two parts separately.

Improving the sampling process. Yes. Cor. 2.6 identifies eLe^{L} as the importance weight of the PoE target with respect to the sampler. This gives a defect correction at no additional network evaluation per trajectory: draw KK samples and keep one with probability proportional to eLe^{L}. We tested it on mixtures where every field and density is available in closed form, and report the energy distance to exact samples from the PoE target (lower is better):

Settingλ\lambdaK=1K=1 (uncorrected)K=4K=4K=32K=32
Interacting concepts(1,1)(1,1)0.0190.0060.002
Subtraction(1,−0.5)(1,-0.5)0.0130.00030.0000
Guidance-like weights(2,2)(2,2)0.1530.1220.073

The correction reduces the error monotonically in KK, also with signed weights, and the gain persists with the practical sampler (Euler 20 steps). In the same setting, the identity of Cor. 2.6 matches the reference KL to four digits (Tab. 15), and the bound of Obs. 2.7 holds in every configuration with the predicted second-order scaling. Both validations are in the new App. D; the resampling results are reported here.

Regarding adaptive weighting, eLe^{L} is the exact correction that any adaptive scheme on the weights λi\lambda_i would have to approximate, and the Gram form of Eq. 12 indicates where such a scheme should act: on the pairs of concepts with large ∣⟨δi,δj⟩∣\lvert\langle\delta_i,\delta_j\rangle\rvert. We consider this a promising direction, and the conclusion of the manuscript points to it, but it is a separate line of research, beyond the scope of this work.

Predicting failures. On the trained CLEVR model, the defect behaves as the theory predicts (App. B.3):

  • It is localized: the cross term concentrates on the pixels where two objects compete (Figs. 3 and 12).
  • It follows the interaction: it falls by a factor of 7.2 as the two objects separate (Tab. 8).
  • It is positively associated with scene failure at all three detection tolerances, with confidence intervals that exclude zero (Tab. 9). The effect is moderate, since the defect measures the distance from the PoE target, whereas the detector measures whether the requested objects appear.
  • The failure modes of App. B.2 (merged, stacked or color-bleeding objects) occur in the regime where Cor. 2.5 predicts the largest cross term: co-located or conflicting conditions.

We hope these clarifications and revisions address the Reviewer's concerns, and we would be grateful if they would consider reassessing their evaluation and score based on the updated text and the above comments, and support the acceptance of the work.

Thank you for the authors' response. It has partially addressed my novelty concerns. However, from my perspective, the practical significance of the Composition Defect remains limited, as its contribution is primarily theoretical and its practical benefits are not sufficiently demonstrated. Therefore, I will maintain my original score.

We thank the Reviewer for the reply, and for acknowledging that our response partially addressed the novelty concerns. We would like to clarify the role of the composition defect.

The paper does not aim to correct the defect. It shows that a single uncorrected ODE composes successfully with the defect present, and explains why. Prior compositional samplers remove this inexactness at extra cost (MCMC correctors, SMC over particles, rejection sampling); our results show this cost is not needed:

  1. Its effect is bounded. The sampler error is second order in the interference between concepts, KL(q1∥p~1)≤18(sup⁡L−inf⁡L)2\mathrm{KL}(q_1\Vert\tilde p_1)\le\tfrac18(\sup L-\inf L)^2 (Obs. 2.7), and vanishes exactly when the defect is constant in space (Cor. 2.5).

  2. Composition succeeds. With no corrector, CompFlow is competitive with correction-based samplers on CLEVR (Table 1), and on the subtraction benchmark it removes the concept while keeping the subject in 62% of the images, against 1% for in-prompt negation at matched compute.

As today is the last day of the discussion period, we would be glad to address any remaining concern on novelty, and we would be grateful if the Reviewer would reconsider the score in light of this clarification.

Reviewer_ABPoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CompFlow, a training-free combinatorial generation method. It weights and sums the velocity differences of each condition relative to the unconditional model for sampling in the flow matching model: positive weights make the image satisfy multiple conditions simultaneously, and negative weights are used to remove a certain concept. The experiment combines the shapes, colors and positions of up to five objects on CLEVR. On FLUX.1[dev], the non-spatial class prompt words of T2I-CompBench were split and then combined.

Critical strengths

The method is training-free, needs no architectural changes, and is easy to implement on top of any pretrained flow-matching model. Addition and subtraction of concepts are handled by a single rule.

Critical weaknesses
  1. Limited novelty. Under Gaussian paths, residual composition in velocity space is mathematically equivalent to Composable Diffusion (Liu et al., 2022), including negation. The defect analysis largely overlaps with Du et al. (2023) and FKC (Skreta et al., 2025), and it is used only as a diagnostic, never for correction. The authors need to rewrite the contribution, they are overclaimed, especially contribution 1.
  • Liu, N., Li, S., Du, Y., Torralba, A., & Tenenbaum, J. B. (2022). Compositional Visual Generation with Composable Diffusion Models. ECCV 2022.
  • Du, Y. et al. (2023). Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC. ICML 2023.
  • Skreta, M. et al. (2025). Feynman-Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts. ICML 2025.
  1. Weak diagnostic. The defect correlates only weakly with failure (Spearman ρ = 0.12). Its “late accumulation” may simply be an artifact of the t/σ_t weighting. The closest distance bins likely request physically unrealizable configurations, which confounds interference with infeasibility.
  2. Flawed CLEVR evaluation. Training conditions on one object in multi-object scenes, so extra objects are expected, yet neither metric penalizes them. Figs. 10 and 12 show duplicated or extra objects and inconsistent placements, for example the same coordinate (2,−1) rendered at different locations. Hyperparameters are tuned on the test set, and Table 1 compares against baselines copied from another paper rather than the same backbone.
  3. Claims inconsistent and unreliable with figures and data.
  • B.3 reports 75–87% scene accuracy per bin, but Table 9 reports 69% overall.
  • Fig. 9 disagrees with Table 2.
  • B.2’s text does not match Table 7.
  • The captions of Figs. 6 and 16 claim subject preservation, but identity and breed visibly change.
  1. Narrow T2I evaluation. Only the non-spatial category is reported, while Appendix C.4 admits failures on binding and spatial prompts. FLUX.1[dev] is guidance-distilled, so the theory’s assumptions do not hold. The LLM-decomposition confound is not ablated, and the gain is only about 2σ.
  2. Weak and unconvincing qualitative results. The figures in the paper shows a lot of failure cases. The only additional CLEVR samples in figure 10 can see a lot of extra objects and incorrect position and size. For concept composition addition and reduction, almost all of the illustrated images are bad at identity preservation. It also generate images far away from the prompt object. For example in Fig. 5 the horse’s color and breed change across steps. Concept subtraction has no quantitative metrics or baselines, and samples collapse at λ ≥ 1.5.
  3. Overclaimed performance and Presentation issues. The authors overclaimed their performance and contribution, The authors need to make sentence logical and smooth. These paper include broken caption (Fig. 27), inconsistent notation, and incorrect citations.
Questions for authors
  1. Beyond Composable Diffusion (equivalent in ε-space) and the PoE correction weights in FKC, what is the specific theoretical contribution of this paper?
  2. Can you report metrics that account for extra objects (e.g., precision, object-count accuracy, exact-scene accuracy)? Can you also report results with hyperparameters selected on a separate validation set?
  3. Can you explain the following numerical inconsistencies?
    • B.3 reports 75–87% scene accuracy per bin, but Table 9 reports 69.0% overall.
    • Fig. 9 disagrees with Table 2.
    • Table 6 gives 92.1% per-object accuracy at m=2, w=1, yet the two-object scene accuracy in B.3 is only 69%. What accounts for this gap?
  4. For the FLUX experiments:
    • What guidance scale does the single-prompt baseline use?
    • How does CompFlow compare under matched NFE?
    • How does it compare when the decomposed sub-prompts are concatenated back into a single prompt?
    • Which LLM is used for prompt decomposition?
  5. Can you report results on the attribute binding, spatial, and complex categories of T2I-CompBench?
  6. Does the claim that "the defect accumulates late in sampling" still hold without the t/σ_t weighting?
  7. Can you provide quantitative metrics for concept subtraction, and a comparison with native negative prompting?
8 comments

We thank the Reviewer for their comments. We appreciate their recognition that the method is training-free, requires no architectural changes, and is easy to implement on top of any pretrained flow-matching model, and that addition and subtraction of concepts are handled by a single rule. The review also identified reporting errors in our appendix, which we have corrected. We now address each of the Reviewer's points, and answer the questions by number in the third comment.


W1. Novelty

We are aware that, under Gaussian paths, residual composition is equivalent to composing diffusion, score-based or energy-based models, and the manuscript already states this (lines 45–47).

The main contribution of our manuscript is to assess how far the composed signal vtcompv_t^{\mathrm{comp}} produced by a deterministic flow-matching ODE can go without corrections, resampling, or network evaluations beyond the m+1m+1 required to implement the product of experts. The works the Reviewer cites as overlapping respond to the inexactness of plain composition by taking the opposite route: Du et al. (2023) add MCMC steps, and Skreta et al. (2025) reweight and resample populations of SDE particles.

We go in the opposite direction, and show two main things.

1. Correction-free composition is competitive. Under flow matching and the PoE framework, our correction-free, zero-shot composition rule is competitive with samplers that rely on correction. We show this in the controlled CLEVR setting, where CompFlow matches or exceeds naive Composable Diffusion and correction-based frameworks. These results establish, both qualitatively and quantitatively (Tab. 1), that correction-free sampling and composition is robust on its own. Moreover, we push the CLEVR setting beyond the position-only conditioning of Du et al. (2023) and test whether our composition rule in flow-trajectory space can compose additional attributes; again, Tab. 2 reports robustness quantitatively. Finally, we show that our formulation transfers seamlessly to a modern, billion-scale, pretrained text-to-image backbone, where a plain ODE sampler is still able to add and remove concepts from an arbitrarily large number of requested concepts. To the best of our knowledge, neither the cited works nor the broader literature report results of this kind: complex conditions, in arbitrary number, on flow matching.

2. Exact characterization of the uncorrected sampler. Since we push the correction-free rule to its limit, we characterize exactly how a composed FM ODE transports probability mass. In particular, Theorem A.2 gives the exact residual of our uncorrected sampler under the flow-matching dynamics, in closed form, using only information available at inference time. It identifies when no correction is needed: the mass transported by the dynamics is exact if and only if the defect DtD_t is constant. Corollary 2.5 extends this result and shows that this is the case under conditional independence.

New results. To strengthen this characterization, we added two new results to the revised manuscript, both direct consequences of Theorem A.2 (new Corollary 2.6 and Observation 2.7). The complete statements and derivations are in the updated manuscript. We hope they clarify the scope of our work, and we summarize them here for quick reference.

Corollary 2.6. The defect integrated along a trajectory,

L=∫01σtt Dt(Xt) dt,L=\int_0^1\frac{\sigma_t}{t}\,D_t(X_t)\,dt,

is exactly the log density ratio

log⁡p~1q1(X1)=L−log⁡Z1Z0.\log\frac{\tilde p_1}{q_1}(X_1)=L-\log\frac{Z_1}{Z_0}.

As a consequence,

KL(q1 ∥ p~1)=log⁡E[eL]−E[L],\mathrm{KL}(q_1\,\Vert\,\tilde p_1)=\log\mathbb E\big[e^{L}\big]-\mathbb E[L],

so the uncorrected ODE reports its own distance from the target, using only quantities it already computes and no additional network evaluation.

Observation 2.7. This distance is bounded as

KL(q1 ∥ p~1)≤18 (sup⁡L−inf⁡L)2,\mathrm{KL}(q_1\,\Vert\,\tilde p_1)\le\tfrac18\,\big(\sup L-\inf L\big)^2,

which shows that the error of the correction-free sampler is bounded by the interference between the concepts.

Finally, Cor. 2.6 shows that eLe^{L} can be interpreted as a weight toward the PoE target; we updated the manuscript with this observation. We stress, however, that the claim of the paper concerns what the flow-matching ODE achieves with no correction.

To the best of our knowledge, these results have not been studied in the literature for FM ODEs. We validated both results empirically on closed-form mixtures, and added a discussion of the results in App. D.

W2. Diagnostic

We address the two main points of this weakness separately.

The σt/t\sigma_t/t weighting and the "late accumulation" of the defect. We thank the Reviewer for this sharp observation. The weighting term is not arbitrary but required: it is the coefficient of the continuity residual (Eqs. 3 and 22), and therefore the rate of the log-density ratio (Cor. 2.6). In the newly added App. D, we show that the exact field computed with this weight matches the numerical derivative of the true KL along the whole path, and that its time integral returns the final KL. An unweighted overlap has no such meaning.

Nonetheless, the Reviewer's request is fair, and we measured the effect of the weight. If the late concentration were an artifact of the weight, removing the weight should make it disappear: the unweighted overlap would then be spread evenly over the trajectory, placing about 50%50\% of its integral after t=0.5t=0.5, or it would be concentrated early. We therefore recomputed the profile of App. B.3 without the weight, i.e., the residual overlap 2⟨δ1,δ2⟩2\langle \delta_1, \delta_2 \rangle alone, on the same 768 condition sets (6,144 trajectories). The late concentration persists: the unweighted overlap peaks at t≈0.65t\approx 0.65, and 64%64\% of its time integral lies at t≥0.5t\geq 0.5, against 88%88\% with the weight. The weight therefore sharpens the late concentration, but does not create it: the interaction between concepts itself builds up in the second half of the trajectory. We report this analysis in App. B.3 of the revised manuscript.

Interference and infeasibility. We respectfully disagree with the Reviewer's framing. Interference has a formal definition in the paper: it is the cross term ⟨δi,δj⟩\langle \delta_i, \delta_j\rangle of the defect DtD_t (Thm. A.2), which vanishes when the residuals are orthogonal (Cor. 2.5). "Infeasibility" is defined neither in the review nor in the paper, and we are not aware of any formal definition of it in this setting. The only precise reading we can give it is a probabilistic one: a request is "unrealizable" when its conditions are far from conditionally independent, so that the product of experts is not exact. That is the definition of interference, not a confound of it. It is the regime where Thm. A.2 predicts the largest cross term, and it is what Table 8 and the heatmaps (Figs. 3 and 12) show. This reading is also contradicted empirically: scene accuracy is highest in the closest bin (75.1% under 8 px at τ=0.10\tau = 0.10) and lowest in the farthest (62.4%). Restricted to pairs farther apart than one object footprint (20 px), where no overlap is possible, the association is unchanged.


W3. CLEVR evaluation

  • Extra objects. We followed the experimental setting of Du et al. (2023) and Liu et al. (2022): each conditional is trained as "this object is in the scene", on scenes with one to five objects. The position-accuracy metric we report is existential, as in those works, and measures whether the requested objects are in the right position. This existential protocol is widely used in well-established work, and we believe that exact object counting is outside the scope of our work. For clarity and completeness, we nonetheless report object precision below.
mmPrecisionObjects generatedObjects in target
146.32.753.66
272.22.844.25
381.73.514.60
485.64.114.83
585.64.735.00

Objects in target: average number of objects in the training scenes that contain the requested ones.

  • Figs. 10 and 12. Conditions are 3D world coordinates on the ground plane, and the camera is fixed. Object size is not conditioned, so the same coordinate can be rendered a few pixels apart across seeds. We clarified this in the caption of Fig. 10 of the revised manuscript.

  • Hyperparameters tuned on the test set. We respectfully but firmly disagree with this statement, which is a serious one and is not substantiated in the review. No trained quantity is selected on test data: the trained CLEVR UNet has a single tuned hyperparameter, the number of training epochs, which is set by early stopping on a held-out validation split of images. The architecture is inherited unchanged from Du et al. (2023). We ask the Reviewer to indicate which specific hyperparameter of the CLEVR model they consider to be tuned on the test set, so that we can address it precisely.

  • Table 1. The rows marked with †\dagger are reproduced from Yu & Gao (2025), as the caption states. We agree that they are not a same-backbone comparison: they share the type of verifier, not the model or the condition sets. No claim of the paper rests on outperforming these rows.

W4. Consistency of claims, figures and data

We thank the Reviewer for these observations, and address them point by point.

  • B.3 and Table 9. We traced the inconsistency to a labeling error in the script that generates the tables: the B.3 text reported results for τ=0.15\tau = 0.15, while Table 9 is at τ=0.10\tau = 0.10. We report the correct numbers in the revised manuscript.

  • Fig. 9 and Table 2. We thank the Reviewer for spotting this. Fig. 9 is correct: for each mm, all values are taken at the ww with the best accuracy. The error is in Table 2, whose pixel-error and DINOv2-similarity columns reported the best value of each metric over ww, instead of the value at that same ww. Table 2 now reports a single ww per mm in every column. In the same check, we found that two pairs of pixel-error entries in Table 6 were swapped (m=1m=1 at w=1,2w=1,2, and m=4m=4 at w=4,5w=4,5). They are corrected, and Table 6 now agrees with Figs. 8 and 9.

  • B.2 and Table 7. We thank the Reviewer for spotting this. Table 7 and Fig. 11 are correct; the B.2 text attributed to the Conflicting attributes row (red and green cube) the purple merge that occurs in the Co-located row (red cube and blue sphere). The Conflicting attributes samples are two-toned cubes. The text is corrected, and the conclusions of B.2 are unchanged.

  • Identity preservation. The prompts specify a category, not an instance of the subject. In Fig. 5, for example, no breed or color of the horse is requested, and a style condition moves the flow toward a different distribution over images of that category. The method preserves the category and does not claim to preserve the instance. The same applies to Fig. 6, where the positive prompt asks for a portrait, and to Fig. 16 (Fig. 17 in the revised manuscript), where no particular breed or color of the dog is requested. This is the intended behavior of the method.

Beyond the reporting errors fixed in the revised manuscript, we are not aware of any inconsistent claim or unreliable experiment, and the review does not point to a specific inconsistent claim. We would be grateful for specific instances, and will correct them.

Reviewer_LUPWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a flow-based generative model that can satisfy multiple conditions given by the user to generate an image. For example, making sure that all the concepts given in text description are present in the generated image. The overall method operates at inference time without needing to retrain the original flow-based model (e.g., Flux), by simply weighing the velocity residuals – positively for concepts to be kept, negative for concepts to be removed/suppressed. Results are shown on two tasks; (i) multi-object generation on CELVR dataset and (ii) text to image generation with a focus on composing visual concepts. The proposed method shows effectiveness on both tasks qualitatively and quantitatively, by adhering to the conditions given by the user.

Critical strengths
  • While the overall method is not easy or straightforward, the modification happens only during inference of an already pretrained model, i.e., no new model needs to be trained in principle.

  • The idea of having the generative model being aligned to the input condition (e.g., making sure to preserve the contents of text prompt) has been studied for diffusion based generative models, but this work now enables that idea for flow based models as well.

  • On the compositional generation task for CLEVR dataset, the proposed method achieves a good performance in ensuring that the conditions for the generated image are being met and the object that gets generated is placed at the accurate (target) location.

  • For text to image generation, the proposed method does a satisfactory job in improving the combination of multiple visual concepts (compared to baseline flow model). Furthermore, it also does a good job at removing undesirable concepts (Fig. 6). Overall, this is measured quantitatively through various metrics for text-image alignment.

Critical weaknesses
  • It seems that one of the key results is centered around text-to-image models, and how the natural language interface can be used to enhance or suppress the visual concepts. Considering that, the authors have not compared against many diffusion model based approaches which have the same goal [1, 2, 3], some of which also do image editing by suppressing concepts. In what way is the proposed conditional image generation problem different than the ones like those listed above? I am not asking the authors to compare against every single such baseline, but at least a few other prominent baselines' comparison will be helpful.

  • Currently, the problem formulation for image generation for CLEVR dataset is not very clear. Specifically, what is the exact input (which, I believe, serves as the ground truth)? The figure 10 shown in the appendix makes things slightly more clear, and so it certainly needs to be a part of the main paper. Otherwise, it is not clear to the reader what the target is that is being compared to, which forms the result in Tables 1 and 2.

  • While there are many qualitative examples shown in the appendix (Fig. 15-24), they all show the capabilities of the proposed method. But we do not have a lot of qualitative comparison to how different that capability is compared to the base flow model (Flux), for which we only have the two examples given in Fig. 4. This ties in to the first point, where more qualitative comparisons are needed compared to other approaches tackling similar problem. Finally, this comparison is also needed because the quantitative results (Table 3) do not show a very convincing improvement over the baselines.

References

[1] Aligning Text to Image in Diffusion Models is Easier Than You Think. Lee et al. NeurIPS 2025 [2] Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models. Chefer et al. SIGGRAPH 2023 [3] InstructPix2Pix: Learning to Follow Image Editing Instructions. Brooks et al. arXiv 2023

Questions for authors

Lines 429-431: Why should the effects not be independent?

4 comments

We thank the Reviewer for their thoughtful comments and constructive feedback. We appreciate their recognition that the method operates only at inference on an already pretrained model, that it achieves a good performance on CLEVR in ensuring that the conditions are met and that objects are placed at the target location, and that it does a good job at removing undesirable concepts. We address their concerns below and hope these clarifications will help improve their assessment towards acceptance.


W1. Baselines. We thank the Reviewer for the thoughtful comments and for the pointers to related work. The cited works are relevant and influential contributions to text-to-image alignment and image editing. While their final output is similar to ours, we think that the direction they take is orthogonal to CompFlow: they all aim at a better agreement between the text and the generated image, but under different assumptions. We detail the differences below.

  • Attend-and-Excite [2] is the closest to our work and shares the goal of concept addition. It optimizes the latent at each step against the cross-attention maps of user-selected tokens, and it does not support removal. The original manuscript already compares CompFlow with Attend-and-Excite (Table 3).
  • SoftREPA [1] improves single-prompt alignment by fine-tuning text tokens with a contrastive objective. It is complementary to our work and requires fine-tuning. CompFlow can be applied to any flow-matching backbone without fine-tuning, since the alignment is performed by the velocity arithmetic we introduce.
  • InstructPix2Pix [3] is an image-editing framework, trained on synthetic edit pairs. CompFlow is orthogonal to it as well: it requires no training or alignment, and steers the velocity field at inference time.

In summary, the setting of CompFlow differs significantly from that of the existing text-to-image baselines: a frozen, possibly off-the-shelf pretrained model, independently encoded conditions, and a single rule that steers the generation at inference time to add and remove concepts. We nonetheless consider the mentioned works relevant, and we updated the related work section of the revised manuscript accordingly, to better position our work. We thank the Reviewer again for this observation.


W2. CLEVR problem formulation. We thank the Reviewer for pointing this out. We summarize the formulation here.

Input. The input is a set of mm object specifications {(shapei,colori,(xi,yi))}i=1m\{(\text{shape}_i,\text{color}_i,(x_i,y_i))\}_{i=1}^m, drawn from the annotations of rendered scenes. No ground-truth image is given to the model or used in the evaluation.

Training. The model is a standard UNet trained on scenes containing one to five objects. Each training image is paired with the specification of a single object present in the scene, so the network only ever sees one condition at a time. This follows the setting of Liu et al. (2022) and Du et al. (2023), for a fair comparison. We added the training algorithm to the revised manuscript (Algorithm 2, App. B).

Inference. The trained model is evaluated once per condition, for each of the mm conditions. The resulting velocities are combined with the CompFlow composition rule, and the image is sampled by integrating the composed ODE.

Evaluation. We sample a batch of condition sets and generate one image per set. An image is valid if it contains all requested objects, with the requested attributes, at the requested positions. For Table 2, a detector locates the objects, detections are assigned to the requested objects by Hungarian matching, and an object counts as correct if it lies within τ⋅128\tau\cdot 128 pixels of its target (Table 2 uses τ=0.15\tau=0.15). Table 1 uses the position-only check of the cited baselines: every requested position must be occupied by an object.

W3. Qualitative comparison with the base model. We agree with the Reviewer that the two examples of Fig. 4 were not enough to show how CompFlow differs from the base model. We added new comparisons with single-prompt FLUX.1[dev] to the revised manuscript.

Concept addition. The new App. C.6 (Fig. 29) shows single-prompt FLUX.1[dev] and CompFlow side by side on six examples: the three additive examples of Fig. 1, the green apple ++ bear example, and two prompts in the T2I-CompBench format. In all six cases the single prompt drops or weakens at least one of the requested concepts: it loses the pillow-book, weakens the amber atmosphere, replaces the wool cloak with a modern jacket, ignores the bear, loses the writer and the book, and reduces the insects to a single specimen. CompFlow retains them. The mechanism is the same in all cases: in a single prompt all concepts are encoded in one text embedding, and the dominant ones can absorb the others, whereas in CompFlow each concept enters the dynamics through its own velocity term at every integration step.

Concept removal. The new App. C.3 compares CompFlow with the base model on removal, qualitatively (Fig. 15) and quantitatively (Table 13). On 42 concepts that FLUX.1[dev] adds without being asked, CompFlow removes the concept and keeps the subject in 62% of the images, against 1% for negation inside the prompt.

These comparisons complement the aggregate scores of Table 3: they show which of the requested concepts a single prompt loses and CompFlow keeps.


Q1. Lines 429–431: why should the effects not be independent? The conditions of Fig. 5 are composed with one rule, but their effects on the sample interact, and the theory says why. By Eq. (12), concepts interact through the cross terms λiλj⟨δi,δj⟩\lambda_i\lambda_j\langle\delta_i,\delta_j\rangle of the composition defect, which vanish only in the cases of Cor. 2.5 (orthogonal residuals, or conditionally independent concepts). Fig. 3 makes this interaction visible on CLEVR. The sentence at lines 429–431 states this: composing conditions of different types does not make their effects independent.


We hope these clarifications and revisions address the Reviewer's concerns, and we would be grateful if they would consider reassessing their evaluation to support the acceptance of the paper.


  • Liu, N., Li, S., Du, Y., Torralba, A., & Tenenbaum, J. B. (2022). Compositional Visual Generation with Composable Diffusion Models. ECCV 2022.
  • Du, Y. et al. (2023). Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMC. ICML 2023.

Dear Reviewer,

As the rebuttal discussion period is approaching its end, we would like to kindly follow up on our responses and the additional results and clarifications provided.

We would appreciate hearing whether our responses have addressed your concerns or if there are any remaining points you would like us to clarify. We would be happy to discuss them further and hope you might reconsider supporting the acceptance of our work in light of the new evidence.

Best regards,

The Authors

Reviewer_LUPWOpenReview ↗

Questions about the new qualitative results

I am responding to W3, where the authors included more comparisons with Flux 1.1 model in Fig. 29. I tried using the same prompts that authors have indicated but my results showed something different. There were total of 6 cases presented (a-f), and I will briefly summarize the results that I obtained. (a) roughly the same results as author (could not generate a pillow book) (b) I suppose the amber style is not being captured by base model (?) If so, then yes, even I cannot recreate that aspect. (c) if the issue with base model was the lack of wool cloak, then I was able to generate that on my side. (d) I am able to generate an almost replica of the CompFlow result using the base Flux model. (e) I can again generate an almost exact replica like CompFlow (e.g., only the gender is different) (f) I can generate a much better example than the one given, where there are actual insects.

In other words, the only cases where the CombFlow model is slightly better than the base model are cases (a) and (b). Other 4 cases they appear roughly the same. Hence, qualitatively, it is not clear to me what the benefit of the proposed method is. Also worth noting is that I did not even try different random seeds to get better image quality for the Flux model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.