acceptodds
Under review as a conference paper at ICLR 2027

Discriminant Flow Steering: Training-Free Capability Transfer in Flow-Matching Models

Abstract

Open flow-matching image models fail specific, checkable and sometimes trivial capabilities that stronger, often closed, models handle like a three-column menu, an orthographic drawing, or a diagram with exactly three boxes. We ask whether we can close such a gap with no training, no teacher weights and no teacher call at inference, and from about twenty images from each model. To answer this, we propose Discriminant Flow Steering (DFS), which encodes the teacher images with the student’s VAE and uses the difference between the mean teacher latent and the mean student latent as the steering direction when generating new images. The unit vector along this difference is added to the latent during the first sampling steps; nothing else changes. We find that this vector is what guidance with a second model adds at each step, except that we compute it only once and offline. This holds exactly at the first step for any two models, and approximately at high noise if we model the latents as Gaussian. Prior steering methods fit their direction at the same noise level where they apply it. We derive the exact finite-sample law for such a fit and find that, in the first steps where steering matters most, 99% of the fitted direction is noise. This is why we estimate the direction from clean latents. We also measure how far a perturbation injected at each step travels to the final image. At the first step the mean difference is amplified 8.8 times more than a random direction, and by mid-trajectory the advantage is gone. To measure capability transfer rather than resemblance to the teacher, we build DFS-CapBench: 200 prompts in 20 capability families, each with a frozen checklist, admitted only if the teacher passes and the student fails. An open vision-language model judges every checklist item and never sees a teacher image. On the 126 strict prompts DFS passes 42.1% of images at one global strength, against 10.1% for the student on fresh seeds, and above prompt rewriting, guidance sweeps and negative-prompt guidance with the same access. Attention-space negative guidance does better than DFS when the failure can be named in a negative prompt, and worse when the missing capability is structural, so the two methods complement each other. When we combine them, they pass 63.9% of images, on par with a per-prompt LoRA trained on the same images. On structural capabilities DFS alone passes more images than that LoRA and keeps the student’s seed diversity. We also test the same method on a second student, a second teacher, and a small-scale video experiment.

Primary area:
generative models
Keywords:
inference time steering, flow matching, image generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wxdGOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed a new flow-matching model steering method. With only 20 images from the student and teacher, it can guide the student model to generate images it would not be able to generate using the delta between the student's images and the teacher's images in the student's latent space.

Critical strengths
  • The proposed method is effective, and the solution is interesting, especially when combined with a negative guidance method like NAG.
  • Including a diversity metric is a great choice. One concern I had for methods for strong-to-poor distillation/steering/improvement is that they tend to generate homogeneous output. The paper correctly addressed this concern.
  • The judge is correctly validated against humans.
Critical weaknesses
  • (minor) The headline "no call to the teacher during inference" is slightly misleading. It is correct that during denoising of the student, we do not need to call the teacher. However, we still need to collect the teacher's output per prompt. This means calling the teacher every time we use a new prompt. This also raised a question about the use cases of this method: if I need to generate a large batch of images X, it might be worth collecting the teacher's output. Otherwise, given that capability transfer is not universal, the use case is limited.
  • (minor) Table 1 bolds every DFS row, which reads as marking the best results, but LoRA outperforms DFS alone in every column. If bold is meant to highlight the proposed method rather than the best result, the caption should say so.
Questions for authors

Could the authors explain the use case of this method better?

3 comments
AuthorsOpenReview ↗

Response to Reviewer wxdG: teacher calls, use cases, and Table 1

W1. "No call to the teacher during inference".

Thank you. The phrase means that the teacher is never called while the student samples, and the camera-ready version will say it in these words. A direction needs one set of teacher images, collected once. The paper uses 20, and ten pass as often on a 30-prompt subset (46.7% against 45.0%, Table 11).

A new prompt does not always need a new set. New experiments for this rebuttal show that a direction can be reused on prompts that keep the same structure (Appendix G). Applied unchanged to 73 new prompts, 27 directions raise the pass rate from 17.6% to 45.0%, with no teacher image of the new prompts. A prompt that needs a different structure needs its own set.

Q1 and W1. The use case.

DFS is for people who can reach a stronger, closed image model but cannot or do not want to route every image through it. They buy one small set of images, 10 to 20, that show a structure their open model cannot draw. After that the open model draws this structure itself, as often as needed and for new content, with no further call. Some examples of who this serves:

A marketing team with one layout and many documents. A restaurant chain or a retailer has a menu or flyer layout that works, made with a closed model. Every week it needs the same layout with new dishes, prices and branches. With DFS the team keeps the layout as one stored direction and produces each new version on its own machines.

A designer with a template that is easier to show than to describe. Examples are a poster with coloured bands around any word, or one subject shown in four seasons in a 2×2 grid. The open model receives the full description and still gets it wrong. A few example images carry the template to new words and subjects (Figure 14), and trying variations costs no call. The examples can come from a closed model. In principle a designer could also make or correct them by hand, which we have not tested.

A company whose content must stay in-house. A pharmaceutical firm, a bank or a team preparing a product launch cannot send unreleased names or internal documents to an outside service. It can fit the direction on a harmless prompt with the same structure and then generate the confidential version locally.

A team that depends on a closed model it does not control. Closed models are updated, repriced, rate-limited and retired. A studio or publisher that built a workflow on one can store directions for the structures it relies on, and keep producing them on an open model when the service changes.

A group that needs images in bulk. A textbook publisher needs hundreds of diagrams in one format. A research lab builds a synthetic dataset of documents with a fixed layout. Both need volume and variety, and the open model supplies both.

These cases rest on two results of the paper: a direction keeps working on new subjects and reworded prompts (Appendix G), and the images keep the student's seed diversity (Table 12). Our experiments use open teachers as stand-ins for the closed model. A direction carries one structure that sits at the same place in every example, such as columns, bands, grid cells or panels. It does not carry textures or drawing styles. Not every image succeeds, so a user generates a few and keeps the good ones. The camera-ready version will state these use cases in the introduction.

W2. Bold rows in Table 1.

Thank you for catching this. Bold in Table 1 marks our method and not the best value. In the camera-ready version, bold marks the best value in each column, and our rows are marked by shading with a note in the caption. The numbers do not change. The per-prompt LoRA passes more images than DFS alone, as Section 5 states. LoRA trains an adapter for every prompt, and DFS needs no training. With NAG, DFS is level with that LoRA (60.7% against 58.1% on fresh seeds, Table 23).

Not a weakness and the author does not need to add anything to this but just some other thing I thought about (I posted it earlier but forget to select author as reader): Seems like the teacher images do not need to be from a bigger model, so my concern about the cost and use case is somewhat resolved: real images or hand crafted/rendered images can be used as "teacher". If this is true, the paper would be very interesting.

Yes, your reading is right. The teacher images do not have to come from another model. DFS uses only the images, so they could also be rendered, made by hand or edited, as long as they show the capability. In principle, DFS can then pass that capability to the student in the same way.

Reviewer_Y8e1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Discriminant Flow Steering (DFS), a training-free approach for improving the prompt-following capabilities of flow-matching image generation models using a small number of outputs from a stronger teacher model. For each prompt, DFS estimates a steering direction from the difference between the mean latent representations of teacher and student images, and applies this direction during early sampling steps without modifying the student's parameters or requiring teacher access at inference time. The authors provide theoretical analysis connecting the mean-difference direction to the velocity-field difference between ideal flow models, together with analyses explaining the advantages of clean-latent estimation and early-step intervention. To evaluate capability improvements rather than visual resemblance to teacher outputs, they introduce DFS-CapBench, containing 200 prompts across 20 capability families with verified teacher–student performance gaps. On the 126 strict prompts, DFS improves the image-level pass rate from 10.1% for the unsteered student on fresh seeds to 42.1%, reaching 63.9% when combined with negative attention guidance. The results are particularly encouraging for structural tasks, including counting, document layout, and diagram generation, where DFS outperforms the per-prompt LoRA baseline while avoiding training.

Critical strengths

DFS offers a particularly simple approach to improving a weaker generative model using only teacher-generated images, without additional model parameters or forward passes during inference. The empirical gains are substantial on the constructed benchmark, particularly for structural requirements such as counting, diagram generation, and document layout. On the structural subset, DFS achieves a 56.8% pass rate compared with 45.8% for the per-prompt LoRA baseline, suggesting that the method can be particularly effective for correcting certain types of generation failures without fine-tuning. The complementary performance with negative attention guidance further demonstrates the practical usefulness of the approach.

Beyond proposing a simple steering vector, the paper investigates why the mean-difference direction is effective, why estimating it from clean latents avoids substantial noise contamination, and why early sampling steps are particularly sensitive to the intervention. The finite-sample analysis and perturbation-transport experiments provide valuable insights into the behavior of flow-matching models and offer useful guidance for designing efficient inference-time steering methods.

Critical weaknesses
  1. Insufficient distinction between prompt-specific distribution steering and generalized capability transfer.

The main concern is whether the experiments adequately support the paper's central interpretation of capability transfer. DFS estimates a separate steering direction for each evaluation prompt using teacher and student outputs for that same prompt. Consequently, the demonstrated improvements primarily establish that the student can be guided toward desired outputs for previously observed prompts, rather than that it has acquired a reusable capability.

For example, a steering vector derived from teacher images of a three-column café menu may encode the specific layout and appearance of those examples rather than a general ability to produce multi-column documents. The paper evaluates generalization across random seeds, but this is different from generalization across prompts requiring similar capabilities. An evaluation using directions estimated from one set of prompts and applied to held-out prompts within the same capability family would help substantiate the capability-transfer claim. Alternatively, the authors could clarify the intended scope of capability transfer and position the contribution more explicitly as prompt-specific, sample-based adaptation.

  1. The theoretical motivation and novelty relative to existing reference-based steering methods require clearer justification.

The connection between teacher–student mean differences and two-model guidance is interesting, but the exact theoretical result applies to ideal flow fields defined by endpoint distributions. It does not directly establish an equivalent relationship between the learned velocity fields of practical pretrained models. Indeed, the paper reports substantially weaker alignment when evaluating real networks (Appendix D.1), suggesting that the idealized theory provides only a partial explanation for the empirical performance. A more explicit discussion of this gap would strengthen the theoretical contribution.

Furthermore, the relationship to recent reference-based flow steering and training-free distribution adaptation methods is insufficiently discussed. In particular, Follow the Mean: Reference-Guided Flow Matching (Curvo et al., 2026) investigates flow guidance through reference endpoint means, while MMD Guidance (Sani et al., ICML 2026) explores training-free adaptation using reference distributions. Although these methods differ in their precise formulations and computational requirements, they address closely related principles. A clearer comparison would help establish the distinct contributions of DFS, especially regarding its teacher–student formulation, fixed steering direction, computational efficiency, and theoretical analysis.

Questions for authors

Q1. Generalization of steering directions. Have the authors evaluated whether steering directions estimated from one prompt can improve performance on unseen prompts within the same capability family? For instance, could a direction derived from a three-column café menu improve generation of a three-column newspaper or brochure? Such an experiment would help distinguish transferable capability improvements from prompt-specific adaptation.

Q2. Connection between ideal flow theory and practical pretrained models. Appendix D.1 reports substantially weaker alignment between the estimated mean-difference direction and the target-ward displacement of actual networks than the idealized analysis suggests. Could the authors clarify which assumptions account for this discrepancy and whether directional alignment correlates with downstream steering success across prompts? This would help determine how directly the theoretical analysis explains the empirical improvements.

Q3. Relationship to reference-based guidance. Could the authors clarify the mathematical and practical differences between DFS and recent methods such as Follow the Mean and MMD Guidance? In particular, what advantages arise specifically from estimating the teacher–student mean difference and applying it as a fixed latent-space intervention, rather than using other reference-distribution guidance formulations?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Y8e1 (1/6): reuse on held-out prompts, and the scope of the claim [W1, Q1]

Q1 and W1. Do directions help held-out prompts that need the same capability?

Yes. We thank the reviewer for this question. We built a second benchmark for it, CapBench-Reuse, which we will release together with DFS-CapBench, and we ran the reviewer's example. The results are in the new Appendix G of the revised paper, and the camera-ready version will report the main result in the main text.

Design. Everything in CapBench-Reuse was fixed before we generated any of its images: the source directions, the new prompts, the checklists, the seeds and the analysis. Nothing was chosen after seeing a result. The source directions follow one rule: every strict benchmark prompt whose own direction passes at least 6 of its 8 images, which gives 36 prompts. The new prompts were written by a language model that saw no image and no result. Each keeps the source's wording and changes only the subject. For example, the direction fitted on "a 2×2 grid whose top two cells contain checkmarks" is applied to the same grid with butterflies, guitars or rockets. As in DFS-CapBench, a new prompt enters the benchmark when the teacher passes it and the student fails it on separate seeds. This gives 73 new prompts for 27 directions. Each direction is applied unchanged, and no teacher image of a new prompt is used.

Result. On new seeds the student passes 17.6% of images and the reused direction 45.0% at the paper's default strength. An unrelated direction gives 16.1%. A reused direction was fitted on another prompt, so its best strength can differ from the default, and in practice a user tries a few strengths. With the best of three strengths per prompt, 57.4% of images pass. With the best of three per image, 66.1% pass. An unrelated direction under the same two rules gives 23.0% and 24.2% (Table 21).

The reviewer's example. We applied the café-menu direction to twelve other three-column documents with differently worded prompts, among them a newspaper front page, a travel itinerary and a school timetable. On the newspaper front page the student draws three columns in 1 of 4 images, and with the café-menu direction in 4 of 4. Across the twelve documents the rate rises from 39.6% to 87.5% at strength 5, and every document improves. At the default strength of 7 it is 83.3%, and the teacher reaches 97.9%. What transfers is the three-column layout. None of the steered pages contains menu content.

A direction is therefore not tied to its prompt. It carries one specific structure and applies it to new subjects. Figure 13 shows the café-menu direction on other documents and two other directions on new subjects, together with one direction that does not transfer.

W1. The scope of "capability transfer".

The reviewer asks us to either show reuse or narrow the claim. We show reuse, and we state the scope exactly.

Main claim. DFS-CapBench tests one claim under controls. On 126 prompts that the teacher passes and the student fails, a direction estimated from twenty teacher images raises the pass rate from 10.1% to 42.1%, at one fixed strength for all prompts. In practice a user tries a few strengths. With the best of three strengths kept for each image, 59.5% pass (Table 9). The student is not trained, and the teacher is used only through its images.

Reuse. Whether a direction is tied to its prompt is a separate question, and CapBench-Reuse answers it (Q1). A direction is not tied to its prompt. Applied unchanged to 73 new prompts, 27 directions from the paper raise the pass rate from 17.6% to 45.0% at the default strength, and an unrelated direction gives 16.1%. The gain holds when the new prompt is also reworded (12.5% to 49.1%). With the best of three strengths for each image, 66.1% pass. A new direction fitted on the new prompt itself, with twenty new teacher images, passes 65.8% at the default strength.

Scope. What a direction carries is one specific structure, such as a 2×2 grid or a three-level org chart. It is not a broad skill such as diagram drawing in general. A family of DFS-CapBench is broader than that, because it groups prompts that need different structures. Within a family, the mean direction of the other prompts passes 14.4% of images, against 9.6% for the student and 43.1% for the prompt's own direction, and a direction from a single other prompt gives no gain. We use "capability" in this specific sense. Appendix G of the revised paper states this definition, and the camera-ready version will use it in the main text.

AuthorsOpenReview ↗

Response to Reviewer Y8e1 (2/6): what a direction carries [W1]

W1. Does the vector encode the layout and appearance of its examples, or the ability?

The new experiments separate the three things the vector could carry: the layout, the content of the examples, and their appearance.

Layout. The vector carries the layout. This is the capability, and it moves to new subjects (our answer to Q1).

Content. The vector carries little of the examples' content. On the twelve other documents, no steered page shows menu items. In CapBench-Reuse, every checklist has an item that asks whether the source's subject is absent from the image, for example "Is the image free of checkmarks?" for the grid direction. We checked all 292 steered images for it. The source's subject appears in 23 of them (8%), and 20 of the 23 come from five directions. The last row of Figure 13 shows one. The items about the new subject pass more often with the direction than without it (54.1% against 30.6%). The direction therefore does not replace the new subject with its own.

Appearance. For the café-menu direction, little of the examples' appearance comes along, and how much depends on the strength. The steered pages do not take on the cream paper or the wooden table of the café-menu images (Figure 13). At the default strength the text on many pages turns blue, and a lower strength reduces this. The direction also works when the twelve documents are requested as white text on navy. At strength 5, three columns appear on 75.0% of those pages, against 43.8% for the student on the same prompts.

Scope. How far a direction reaches depends on the direction. In CapBench-Reuse the reused direction raises the pass rate for 22 of the 27 directions, lowers it for 2 and leaves 3 unchanged. Some directions stay closer to their examples. The 2×2 grid direction works when the new prompt asks for a white page (6.2% to 45.8% on twelve other objects) and has no effect on a black page. The direction in the last row of Figure 13 draws its source's squares in place of the requested stars. For comparison, a LoRA trained on the same source images passes 32.8% of the new images, against 45.0% for the reused direction (Table 21). Appendix G reports all of these results.

AuthorsOpenReview ↗

Response to Reviewer Y8e1 (3/6): the theory and real networks, part 1 [W2, Q2]

Q2 and W2. Which assumptions account for the weaker alignment in real networks?

We thank the reviewer for this question. Proposition 1 is exact for the ideal flow fields of the two image sets. The review asks how far it carries over to trained networks, and we measured this for the rebuttal. We reran App. D.1 with the same two networks at their default guidance, on all 184 prompts with 16 seeds each. We also corrected for the fact that each mean comes from only 20 images, so the results refer to the mean difference that unlimited images would give.

The predicted displacement is present at the predicted size. The first-step displacement between the two networks contains the mean difference with coefficient 1.03 (95% interval 0.96 to 1.11). Proposition 1(i) predicts 1.

Why the cosine is 0.34. The real displacement also has parts that Proposition 1 does not describe, and a cosine counts them. Three factors give the measured value.

SourceFactor on the cosine
The first-step estimate of a guided network is offset from the mean of its own images. The two offsets add a part orthogonal to the mean difference0.52
The mean difference is estimated from 20 images per model0.88
App. D.1 computes the cosine for each seed, and 61% of the displacement is shared by all seeds of a prompt0.78
Product0.36 (measured: 0.34)

The assumption behind the first factor. Proposition 1 concerns the two image sets, which are drawn with guidance as in practice. It uses one property of the ideal field of an image set. At the first step, the field's estimate of the clean latent is the mean of the images. Classifier-free guidance adds one term to this. It extrapolates the network's output at every step, which shifts the first-step estimate away from the mean of the images. This shift contributes to the offset of the first factor. It comes on top of the predicted displacement and does not replace it. The coefficient stays at 1.03. In the paper's setup (28 sampling steps), removing guidance lowers the teacher's offset from 245 to 23 latent units and the student's from 365 to 135 (40 prompts). DFS itself uses the teacher's images and never runs the teacher network, so the teacher's offset does not enter the method.

Why a fixed vector works at this cosine. DFS does not need the whole displacement. It adds the part that is the same for every seed and that can be computed from images alone, and this part is present at full size. The sampler also carries it further than other directions. At the start of sampling, a push along the mean difference is amplified 8.8 times more than a random push of the same length on FLUX.1-dev, and 4.1 times more on SD3-medium (App. D.1). The coefficient is still close to 1 at the second step (1.07 on 40 prompts). These first two steps also carry the effect. In the schedule ablation (App. D.3), a push on the first two steps alone reaches 43.3% at its best strength, against 45.0% for the full schedule.

The revised App. D.1 reports this decomposition. The camera-ready version will state after Proposition 1 what is proved for the two image sets and what is measured on real networks.

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

DFS estimates a steering direction for each prompt from the normalized difference between mean teacher and student latents. It injects this direction during early sampling, enabling additional generations of the same prompt without training or further teacher calls. The paper analyzes the direction through ideal flow fields, noise contamination, and perturbation transport, and evaluates it using a capability-gap benchmark, ablations, additional model pairs, and exploratory video experiments.

Critical strengths
  1. A simple intervention improves the student without training. Fresh-seed results support a meaningful steering effect beyond the originally selected failures. DFS requires only teacher output images and adds no model evaluations during student sampling. Diversity measurements also support retaining variation across student seeds.

  2. The analysis provides insight into estimating and applying steering directions. The finite-sample analysis explains why noisy latents can contaminate direction estimates, while perturbation-transport measurements motivate early intervention. Estimator and schedule studies support these choices, making the work informative beyond benchmark performance.

Critical weaknesses
  1. Transfer to new prompts is not demonstrated. Each direction is computed for one prompt, and the fresh-seed experiments keep that prompt unchanged. The paper does not show that a direction works on new prompts requiring the same capability, or that multiple directions can be combined. DFS-plus-NAG does not test direction composition. Users with varied requests may therefore still need new teacher images and a new direction for each prompt. This limits the practical scope of the claimed capability transfer.

  2. The paper does not show when DFS saves cost over using the teacher directly. Preparing a direction requires teacher images, student generations, and latent encoding. If a user needs only a few images, the teacher outputs collected during preparation may already meet the request. Appendix F reports local preparation cost, but does not include teacher-image acquisition or compare the total cost per successful output with direct teacher use. It remains unclear how many additional images must be generated before the preparation cost pays off.

Questions for authors
  1. Assessment-critical validation: Can a direction estimated for one prompt improve held-out prompts requiring the same capability without additional teacher images? Is there evidence that multiple DFS directions can be combined for requests requiring several capabilities?

  2. Practical validation: For what volume of repeated generation does DFS offer an advantage over direct teacher use? Please include teacher-image acquisition, student preparation, and unsuccessful generations in the comparison, and credit usable teacher images obtained during preparation.

  3. Comparison protocol: Section 5 states that DFS's strength was selected on the evaluation seeds, whereas Table 1 reports NAG and LoRA at predefined defaults. This gives DFS a tuning advantage in the main comparison. Table 7 already reports a higher pass rate for NAG at scale 8 than for DFS at α=7\alpha=7, showing that their ordering depends on the strength reported. Could you select each method's strength on the same validation seeds and then compare all methods on the same fresh test seeds?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tWob (1/2): held-out prompts, and combining directions [W1, Q1]

Q1 and W1. Held-out prompts without new teacher images.

Yes. We built a second benchmark for this question, CapBench-Reuse (Appendix G of the revised paper). Everything in it was fixed before any of its images was generated: the source directions, the new prompts, the checklists, the seeds and the analysis.

Design. The sources are the 36 strict benchmark prompts whose own direction passes at least 6 of its 8 images. The new prompts keep the source's wording and change only the subject. A language model wrote them without seeing any image or result. For example, the direction fitted on "a 2×2 grid whose top two cells contain checkmarks" is applied to the same grid with butterflies, guitars or rockets. As in DFS-CapBench, a new prompt enters when the teacher passes it and the student fails it on separate seeds. This gives 73 new prompts for 27 directions. Each direction is applied unchanged, and no teacher image of a new prompt is used.

Result. On new seeds the student passes 17.6% of images and the reused direction 45.0%, at the paper's default strength. An unrelated direction gives 16.1%, so the gain comes from the direction and not from steering as such. The gain holds when the new prompt is also reworded (12.5% to 49.1%) and on all 288 candidate prompts without the gate (26.0% to 47.1%). With the best of three strengths kept for each image, 66.1% pass, against 24.2% for an unrelated direction under the same rule.

What new teacher images would add. A new direction fitted on the new prompt itself, with twenty teacher images, passes 65.8%. Reuse recovers 57% of that gain with no teacher image.

Scope. What transfers is one specific structure, such as a 2×2 grid or a three-column page, applied to new subjects. It is not a broad skill such as diagram drawing in general.

Q1 and W1. Requests that need several capabilities.

One direction can serve a request with several requirements. We fitted a direction on the full request, from one set of teacher images. On six prompts that ask for a 2×2 grid with wireframe objects, it passes 62.5% of images, against 16.7% for the student (difference 45.8 points, 95% interval 25.0 to 62.5). On ten prompts with two requirements, it meets both on 25.0% of images, against 7.5% for the student (difference 17.5 points, interval 7.5 to 28.8). The teacher itself meets both on 57.5%.

Summing two existing directions is a separate question, and it remains open. In two small tests the sum did not beat the better single direction, and it stayed below the direction fitted on the full request (29.2% and 11.2% on the prompts above). We make no claim about summing existing directions.

A direction also combines with a different training-free mechanism. Under the common protocol of your third question, DFS with NAG passes 60.7% of images on fresh seeds, against 43.3% for DFS and 44.6% for NAG alone.

AuthorsOpenReview ↗

Response to Reviewer tWob (2/2): cost against calling the teacher, and one tuning protocol [W1, W2, Q2, Q3]

Q2, W2 and the end of W1. Teacher images, new requests, and when DFS pays.

The setting. DFS is meant for a teacher that is closed and priced per image. With such a teacher every image is a paid call, and so is every retry, every variation and every further image of a batch. In our experiments open models stand in for such a teacher, and DFS uses only their output images.

What DFS needs. One set of teacher images per direction. The paper uses 20, and ten pass as often on a 30-prompt subset (46.7% against 45.0%, Table 11). The preparation also takes the same number of student generations, about five minutes on one GPU, and no training (App. F). The direction is then one stored vector. It can be applied indefinitely with no further teacher call.

The volume at which it pays. Take a user who wants M usable images of one prompt. Calling the teacher costs at least M calls. DFS costs N calls once. These N teacher images are usable themselves, so we credit them. DFS therefore needs fewer teacher calls as soon as the user wants more than N images, which is 20 in the paper's setting, or 10 with the smaller set, which saves even more. Every image after that needs no API call. It is generated on the user's own hardware, where one usable image takes 2.3 generations of FLUX.1-dev on average. Failed generations are counted in this number, because DFS passes 43.3% of images on fresh seeds. For 1,000 usable images, calling the teacher means 1,000 paid images. DFS means 20 paid images and about 10 GPU-hours on the user's own hardware. A direction is also reusable, which saves more when requests vary. Requests that keep the structure and change the subject or the wording need no new teacher image. The same direction serves them, with 45.0% against 17.6% for the student on 73 new prompts (our answer to Q1). Only a request with a different structure needs a new set. Our reply to Reviewer wxdG describes practical cases, including ones where calling the teacher is not an option at all, such as content that must stay in-house or a closed model that is updated or retired.

Q3. One tuning protocol for all methods.

Thank you for specifying the protocol. We ran it as described. Each method's strength was chosen on the four evaluation seeds, and all methods were then scored on four fresh seeds that played no part in any choice.

MethodSelected strengthFresh-seed passBest of three settings per image
Studentn/a10.1%n/a
DFSα = 743.3%56.0%
NAGscale 844.6%53.2%
LoRA per promptscale 1.558.1%73.4%
DFS + NAGα = 7, scale 560.7%74.4%

At the selected strength. NAG and DFS are level (44.6% and 43.3%, a difference of 1.4 points with a 95% interval of −10.3 to +7.5). For DFS, α = 7 and α = 10 tie on the evaluation seeds. We keep α = 7, the paper's setting. α = 10 gives 37.7% on fresh seeds.

Combination. DFS with NAG passes 60.7%. This is 16 points above NAG alone and 17 above DFS alone, and level with the per-prompt LoRA without training. The two methods gain from each other because they close different gaps (Table 2).

In practice. A fixed strength is the right basis for a comparison. A user is not bound to it, and can generate an image at a few strengths and keep the best. The last column applies this rule to every method, with three settings each (Table 24).

The new Appendix I of the revised paper reports every method of Table 1 under this protocol, with every strength on both seed sets. The camera-ready version will report Table 1 under this protocol.

AuthorsOpenReview ↗

Follow-up to Reviewer tWob

Dear Reviewer tWob,

Thank you again for your review and for the precise questions. The discussion ending soon, we would be glad to know whether our response answers them. Here is a short summary, with the places to look in the revised PDF.

Held-out prompts (Q1). On CapBench-Reuse (Appendix G, pages 34 to 38), 27 directions applied unchanged to 73 new prompts raise the pass rate from 17.6% to 45.0%, with no new teacher image. An unrelated direction gives 16.1% (Table 21, page 35).

Figures for the held-out prompts. Figure 13 (page 37) shows the student and the steered student side by side on new prompts. Figure 14 (page 38) shows two more directions, each on 12 new prompts.

Cost against the teacher (Q2). DFS needs fewer teacher calls as soon as a user wants more than 20 images of a prompt. Our reply counts failed student generations and credits the teacher images, as you asked.

One tuning protocol (Q3). We ran the protocol you specified. On fresh seeds DFS passes 43.3% and NAG 44.6%, and DFS with NAG passes 60.7% (Appendix I, with Table 23 on page 40 and every strength in Table 24 on page 41).

If any point is still open, we will answer it before the period closes.

Thank you for the additional experiments and clarifications. They address several of my concerns and better define the method’s practical scope.

  1. Reuse and composition (W1/Q1). The new experiments provide useful evidence that successful directions can transfer a specific structure to new subjects and reworded prompts. However, the source directions are selected for strong original performance, and the new requests largely preserve their structure. Together with the unsuccessful direction-composition experiments, this leaves reuse across diverse or combined requests unestablished. The demonstrated scope is therefore narrower than general capability transfer, and this boundary should be clear in the main text.

  2. Practical benefit and preparation cost (W2/Q2). The concrete use cases and preparation and generation-cost estimates clarify the practical motivation, particularly for repeated generation with a shared structure. The benefit remains workload-dependent: fewer teacher calls do not alone establish lower total cost, and requests requiring new directions incur preparation again. Nevertheless, the response makes the intended deployment setting substantially clearer.

  3. Comparison protocol (Q3). Selecting each method’s strength under a common protocol and evaluating on fresh seeds addresses my concern about unequal tuning. The resulting comparison supports similar observed performance between DFS and NAG, with DFS retaining its advantage of requiring no additional model evaluations during sampling.

These additions strengthen the empirical case, while the limitations in the breadth and composability of the transferred capability remain material. I therefore maintain my recommendation of weak acceptance (3).

Comments on the paper

6 comments
AuthorsOpenReview ↗

Summary of the revision

We thank the three reviewers for their careful reviews. We answer each reviewer in a separate thread. The revised PDF adds three appendices, and each opens with a note that it is new.

  • Appendix G, Reusing a direction on new prompts. A second benchmark, CapBench-Reuse. 27 directions from the paper are applied unchanged to 73 new prompts, which were fixed before any evaluation (Table 21, Figures 13 and 14). For Reviewers Y8e1, tWob and wxdG.
  • Appendix H, Comparison with reference-guided sampling. Follow the Mean and MMD Guidance on both benchmarks (Table 22). For Reviewer Y8e1.
  • Appendix I, Baselines under a common tuning protocol. Every method's strength is chosen on the evaluation seeds and scored on fresh seeds (Tables 23 and 24). For Reviewer tWob.

The main text is unchanged in this revision. The camera-ready version will bring the new results into the main text, mainly CapBench-Reuse and the comparison with reference-guided sampling. We will release DFS-CapBench and CapBench-Reuse.

Hi authors, thank you for your rebuttal. But I think you placed it under the wrong reviewer. I am wxdG but the rebuttal seems to be for tWob

Thank you for flagging the misplaced comment. It is corrected, and our reply to your review is now in this thread.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.