acceptodds
Under review as a conference paper at ICLR 2027

ELROND: Exploring and decomposing intrinsic capabilities of diffusion models

Abstract

A single text prompt passed to a diffusion model yields a wide range of visual outputs determined solely by a stochastic process, leaving users with no direct control over which semantic variations appear. Exploring this range is difficult: random search offers no guarantee of covering it, while prompt editing is coarse, as even a small change in wording can substantially alter the generated image. We argue that systematic exploration instead requires recovering how the model itself organizes the conditional distributions it can produce. We formalize this structure as a _generative manifold_, and present ELROND, a method for recovering its tangent space at a given conditioning. To that end, we collect gradients obtained by backpropagating the differences between stochastic realizations of a fixed prompt, and decompose them into interpretable directions using Principal Component Analysis or a Sparse Autoencoder. We show that our method recovers this subspace accurately in a controlled setting and validate it behaviorally on large-scale models. We also demonstrate that recovered structure enables broader exploration of the model's capabilities than methods relying on external representations, and mitigates mode collapse in distilled models without retraining.

Primary area:
generative models
Keywords:
Diffusion models, mechanistic interpretability, SAE

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

This paper aims to investigate the intrinsic generative capabilities of diffusion models through semantic directions in the conditioning space. The authors analyze the generative manifold induced by a fixed prompt by generating multiple samples and computing gradients with respect to the conditioning based on pairwise differences between generated images. The collected gradients are decomposed using PCA or SAE to identify semantic directions. The authors argue that these directions can be used to control image generation. To validate their approach, they evaluate tangent space recovery on a synthetic dataset and demonstrate its applicability to SDXL and SDXL-DMD. The experimental results show that controlling the identified semantic directions allows the adjustment of specific attributes in generated images.

Critical strengths

The paper presents the proposed method and its validation in a clear manner. It also shows that diffusion models have semantic directions on the generative manifold, which can be used to control specific attributes in the generation process.

Critical weaknesses
  • Limited novelty and scope: While the paper provides a clear and systematic analysis, the proposed approach has a limited scope. It focuses on extracting information already captured by samples generated under a fixed condition. Although this analysis is meaningful, it is unclear whether the findings provide sufficient new insights beyond characterizing existing variations in the generated samples.

  • Limited generalizability and practical applicability: The analysis is conducted on a limited set of predefined concepts, and extending the proposed approach to different conditions requires separate analysis for each concept. This raises concerns about its scalability, and it is difficult to identify practical applications where such an analysis would be necessary.

Questions for authors
  • Generalizability and practical applicability: The proposed approach requires separate analysis for each concept, which limits its scalability. Could the authors further discuss its applicability to a broader range of conditions and provide practical scenarios where this approach would be beneficial? Can semantic directions discovered from one concept be transferred to another concept or conditioning setting? This could provide further evidence of the scalability and broader applicability of the proposed approach.

  • Generalization across models and modalities: Could the authors extend the analysis to other diffusion models, such as DiT or SD3.x, or to diffusion models in other modalities? This could clarify whether the discovered semantic directions reflect general properties of diffusion models or are specific to the models considered.

  • Analysis of limitations and future directions: The authors acknowledge that the proposed approach has limited applicability. However, the discussion mainly describes these limitations without explaining their underlying causes. Could the authors provide further analysis of why these limitations arise within the current framework and discuss potential directions for addressing them?

6 comments

We thank the reviewer for the careful review. We address all points below.

Q1 & W2: Generalizability of the discovered directions and subspaces

As suggested by the reviewer, we study the transferability of individual directions and of whole subspaces. Both transfer well to different conditionings.

Individual directions. A direction discovered for a concept induces similar changes when applied to the same concept in a different prompt and to different concepts (App. X, Fig. 26). For example, a direction found for "A picture of a dog" that turns the dog dark does the same for dogs in longer prompts (on a beach, in a convertible) and for cats, cars, monsters, and people.

Subspaces. Steering with external directions computed for the same concept in a different prompt, or even for an entirely different concept, increases diversity on average almost as much as native directions computed for the prompt itself, with similar prompt adherence (Sec. 4.4, App. V, X).

Same concept, different promptDiversity ↑CLIP score ↑
Unsteered0.2210.313
Random direction0.2380.312
Native0.3530.300
External (from "A picture of a ⟨concept⟩")0.3390.301
Different conceptDiversity ↑CLIP score ↑
Unsteered0.2940.223
Random direction0.3400.220
Native0.5380.207
External (other four concepts)0.5250.207

The results of our method are therefore highly reusable across prompts and concepts, which we find an interesting result in itself and which strengthens the practical applicability of our method. Practical scenarios are discussed in W1 & W2 below.

Q2: Other models and modalities

The submission already includes results showing generalization to the DiT-based Flux Schnell (App. G). However, as requested, we now provide additional results for SD3.5 Large Turbo (MMDiT, App. W). The model is 1.8–3.2× more sensitive to the discovered directions than to random directions of the same norm (sensitivity: DreamSim distance between the image before and after the perturbation), and applying the directions increases diversity (mean pairwise DreamSim distance) substantially on every concept.

Sensitivity ↑carcatdogmonsterperson
Random direction0.1550.0310.0540.1630.132
ELROND0.4180.0980.1400.3010.314
Diversity ↑carcatdogmonsterperson
Unsteered0.4690.1240.2200.2580.415
ELROND0.5810.1930.2950.4610.508

As for other modalities, in App. O we extend our method beyond images and apply it to 3D protein structures generation (see also W1 & W2, Transfer and practical applicability beyond images).

Q3: Causes of the limitations and possible remedies

We have updated the paper with a further analysis of the limitations and of future directions (Sec. 5, App. H). We discuss them one by one.

  • Effects beyond the concept. As stated in the paper, token embeddings are contextual, so the tangent space of a token also contains directions that influence the other tokens, and at high strength a direction can change other subjects of the prompt (App. H). Projecting out the subspaces of the other tokens, or applying negative steering to the affected token, could isolate the concept-specific part.
  • Uninterpretable directions. These are inert, entangled, or artifact-inducing. Inert directions likely come from noise in the gradient estimate, as the fraction of noise directions decreases with more gradients (App. E). We hypothesize that entangled directions arise when the decomposition does not separate the space into monosemantic directions, and artifacts when large steps leave the region where the linear approximation of the curved manifold holds. Better filtering of noise directions, sparser decompositions, or modeling the curved manifold directly could address these cases.
  • Computational cost. The cost comes from backpropagating through the denoiser for every gradient. We already provide mitigations: a small number of gradients captures the main directions (App. E), a single timestep suffices (App. F), and recomputing the subspace is often not needed, as directions transfer across prompts and concepts (Q1).

W1 & W2: Scope and practical applications

We thank the reviewer for finding our analysis clear, systematic and meaningful. We believe, however, that its scope is broader than characterizing the variation already present in the samples: ELROND turns this variation into directions that control it and reach beyond it. Directions estimated from SDXL-DMD's own collapsed samples raise its recall with respect to the teacher from 0.15 to 0.52 (mean over concepts, App. P, Tab. 20), i.e. the steered model generates parts of the teacher's distribution that its samples never contained.

Such an analysis is useful whenever one wants the variation a model can produce without knowing it in advance:

  • Restoring diversity in distilled models, without training or the teacher at inference (Sec. 4.3).
  • Creative exploration without naming attributes (Sec. 4.2).
  • Subject-specific control in a multi-subject prompt.
  • Model auditing: revealing which attributes a model varies for a concept to determine model biases.

Transfer and practical applicability beyond images. We additionally apply ELROND to RFdiffusion3, a diffusion model that generates 3D protein structures (see App. O). We select the task of motif-based generation, which keeps a small protein fragment, called a motif, fixed while generating different structures around it. We discover a shared set of noise-specific PCA subspaces from protein generations without a fixed motif and reuse them across 10 different motifs, without repeating gradient collection or PCA fitting for each condition. For each motif and method, we generate 32 candidate structures, each 100 amino acids long. Steering modifies the conditioning of the surrounding structure during high-noise denoising while preserving the specified motif constraints.

We measure diversity as mean pairwise 1−TM-score1-\mathrm{TM}\text{-score}, calculated outside the motif among samples passing a backbone-geometry screen that checks bond lengths and angles, backbone conformations, and atomic clashes. Results are averaged equally across motifs and reported as mean ± standard error, comparing native generation with ELROND and randomly scrambled steering directions at the same strength.

ConditionGeometry pass (%) ↑Diversity among passers ↑
RFD391.25 ± 2.120.5786 ± 0.0272
Random 0.0595.00 ± 2.140.5822 ± 0.0263
ELROND 0.0591.25 ± 2.270.6377 ± 0.0117

ELROND increases diversity by 10.2% over native generation and 9.5% over matched random steering, while geometry pass matches native generation and is 3.75 percentage points below random (paired tt-tests versus random: significant diversity gain, Holm-adjusted p=0.0147p=0.0147; geometry p=0.1475p=0.1475). This illustrates a practical application: generating structurally varied scaffold candidates around a prescribed motif for downstream screening, without specifying the desired structural changes in advance. The same learned subspaces support multiple design specifications, sharing the analysis cost across conditions.

Reviewer_ftrQOpenReview ↗

Thank you

Thank you for the detailed response and additional experiments.

The transfer experiments help address my concerns. However, the evaluation is still limited to a small set of concepts and prompts. Could the authors further evaluate transferability across a broader range of unseen concepts, prompts, and categories, including generation quality and semantic consistency?

I still have concerns about the scalability and generalizability of the proposed method.

Reviewer_qS2WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, authors introduce ELROND, a method for discovering semantic directions in the conditioning space of diffusion models. Authors generate different realizations of a fixed prompt and backpropagate a latent reconstruction loss between image pairs to the concept-token embedding. The collected gradients are decomposed using PCA or an SAE, and the resulting directions are added to the token embedding during generation. Authors motivate this construction through the tangent space of a generative manifold and evaluate subspace recovery on a synthetic shapes dataset. Experiments on SDXL and distilled models demonstrate applications to exploring concept variations and mitigating mode collapse.

Critical strengths
  • The synthetic experiment is interesting. Having access to a known attribute subspace makes it possible to evaluate recovery directly, rather than relying only on visual examples of steering.
  • The idea of using differences between realizations of the same prompt to discover controllable variations is well motivated. The mode-collapse experiments provide a useful application, including transferring directions from the teacher to the distilled model without retraining.
Critical weaknesses
  • The connection to prior work on understanding noise realizations is not discussed sufficiently. ELROND uses differences between images generated from the same prompt as its discovery signal, while prior work [1] shows that initial seeds induce recurring structure in object arrangement. It is not clear how the recovered directions relate to this structure, or what additional understanding is obtained by representing these variations in conditioning space (as in: is it equivalent to find interpretable directions in conditioning space vs. in the noise space).
  • The large-model experiment demonstrates sensitivity to the discovered directions, but does not establish recovery of the full tangent space: showing that the gradients lie in the space does not show that they span it.
  • It is not clear how much of the steering effect comes from the discovered semantic directions, rather than simply changing the strength of the concept embedding. A comparison with e′=(1+λ)ee' = (1+\lambda)e would help clarify this.
  • Please see the questions below as well.
Questions for authors

Questions and Suggestions:

  • Prior work [1] shows that initial seeds induce recurring object arrangements across prompts. Could the authors discuss how the variations discovered by ELROND relate to this seed-dependent structure? In particular, does the method make existing seed-induced variations systematically controllable, or also expose kinds of variation that are difficult to reach through resampling?
  • How would simple embedding scaling, e′=(1+λ)ee' = (1+\lambda)e, perform compared to adding the discovered directions? Prompt-to-Prompt [2] provides a related control by reweighting a token's attention map, although this is not equivalent to scaling the embedding.
  • In Figure 3, which semantic concept or attribute does the discovered direction represent? Around lines 352–357, why should a direction outside the recovered subspace leave the image unchanged, rather than induce a change that is not aligned with the intended concept? The comparison shows greater sensitivity along the discovered directions, but could relevant directions still remain outside the recovered subspace?
  • Could the authors explain the notation dΦcd\Phi_c when it is introduced around line 176?
  • Section 4 mentions collecting 10000 images and 10000 gradients per concept. Since each gradient is computed from an image pair, how are these pairs selected from the 10000 images? Is each image paired with one other image, or are pairs sampled independently with images reused across pairs?
  • Beyond the equal-compute comparison in Appendix I, what are the time and GPU memory requirements for running ELROND with the configuration used in the main experiments?
  • The main evaluation covers five concepts. How well do directions discovered with the generic prompt transfer to different contexts for the same concept? Appendix N provides multi-subject examples, but a discussion of when rediscovery is needed would help clarify the practical scope.

Typos:

  • Line 274: "mosty" → "mostly."
  • Figure 13: "Drection 2" → "Direction 2."
  • Template text is used for the Author Contributions and Acknowledgments sections.

References:

[1] Li, Shuangqi, Hieu Le, Jingyi Xu, and Mathieu Salzmann. "All Seeds Are Not Equal: Enhancing Compositional Text-to-Image Generation with Reliable Random Seeds." arXiv:2411.18810, 2025. https://arxiv.org/abs/2411.18810

[2] Hertz, Amir, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. "Prompt-to-Prompt Image Editing with Cross Attention Control." arXiv:2208.01626, 2022. https://arxiv.org/abs/2208.01626

5 comments

We thank the reviewer for the careful review. We address all points below.

W1 & Q1: Relation to seed-induced structure

We now discuss this line of work in Sec. 2. The noise and the conditioning control different things. The noise selects which sample of the distribution pθ(x∣c)p_\theta(x \mid c) is generated, so seed-induced structure, such as recurring object layouts, is variation within the support of that distribution. The conditioning selects the distribution itself, and ELROND's directions move between distributions. This is why we work in the conditioning space, and why the two kinds of directions are not equivalent: they overlap and complement, rather than one being a subset of the other:

  • ELROND reaches beyond resampling: its directions raise SDXL-DMD's recall with respect to the teacher from 0.15 to 0.52 (mean over concepts, App. P, Tab. 20), i.e. they produce images that resampling would not.
  • Noise reaches beyond ELROND: properties such as object placement usually stem from the random noise, and are rarely controllable by the conditioning - only if the model learned to do so. Our discovered directions mostly change the concept's appearance, not its composition (Tab. 3).

To answer the question (Q1) directly, ELROND does both: it makes the part of a prompt's variation that the conditioning can express systematically controllable, and it exposes variation that resampling cannot reach. Variation that only the noise controls, such as object placement, is left to noise-based methods.

W2: Recovery of the full tangent space

We agree that the sensitivity test shows that the gradients lie in the tangent space but not that they span it, and now state this explicitly (Sec. 4.1). Full recovery cannot be verified on text-to-image models, as there is no ground-truth tangent space to compare against: it would require the training data, which is unavailable, and even then the model need not have learned it exactly. This is why we verify full recovery in the synthetic setting, where the ground truth is known (recall above 90%, Sec. 4.1). On large models, we show that the recovered subspace saturates: more gradients reveal no new directions (App. E).

W3 & Q2: Embedding scaling

We ran the suggested comparison, rescaling the concept token as e′=e+λ e/∥e∥e' = e + \lambda\, e/\|e\| with λ=±20\lambda = \pm 20 and a random sign per image (App. R). Since ∥e∥\|e\| ranges from 26.5 to 35.8, larger values would remove or invert the embedding for some concepts.

Diversity ↑carcatdogmonsterpersonmean
Unsteered0.3450.1680.2360.3200.4080.295
Embedding scaling0.3440.1750.2460.3410.4170.305
ELROND PCA0.4700.3900.5460.5320.6260.513
ELROND SAE0.4530.3070.4300.4960.6380.465

Scaling barely changes the images, as it offers a single direction, while ELROND combines dozens.

Q3: Directions outside the recovered subspace

Our objective is to empirically recover the subspace Te\mathcal{T}_e. By construction (Sec. 3.2) it contains all directions that change the generated images, when added to one token's embedding. Since a token's influence is largely localized to its concept, these changes concern mostly the concept itself. However, Te\mathcal{T}_e is not strictly restricted to concept-aligned changes: a direction can also change the style or, at high strength, other subjects of the prompt (hence failure cases in App. H). A direction that induces such an unaligned change therefore still lies in Te\mathcal{T}_e, and a direction outside Te\mathcal{T}_e does not change the images at all.

A random direction changes the images only through its component inside Te\mathcal{T}_e, which is small in expectation, as Te\mathcal{T}_e has a much lower dimension than the embedding space. Relevant directions cannot lie outside the theoretical Te\mathcal{T}_e. However, they can lie outside the empirical approximation of it that ELROND computes (see W2).

As for the direction in Fig. 3, it is a randomly selected PCA direction for cat that changes both the cats' color and the style, shifting the distribution toward black, illustrated cats, and so happens to be an entangled direction.

Q4: Notation around line 176

Φ\Phi maps a conditioning cc to the distribution pθ(⋅∣c)p_\theta(\cdot \mid c), and dΦcd\Phi_c is its differential at cc. Intuitively, it describes how the distribution of generated images responds to an infinitesimal change of the conditioning: dΦc(v)d\Phi_c(v) is how the images a prompt produces start to change when cc is moved slightly along a direction vv, e.g. generated elves shifting toward Christmas helpers. Formally, dΦc:RL×d→TΦ(c)Md\Phi_c : \mathbb{R}^{L \times d} \to T_{\Phi(c)}\mathcal{M}, dΦc(v)=ddεΦ(c+εv)∣ε=0d\Phi_c(v) = \frac{d}{d\varepsilon} \Phi(c + \varepsilon v) \big|_{\varepsilon = 0}, is the tangent vector of the curve of distributions traced by moving the conditioning along vv. We now define this notation in Sec. 3.1.

Q5: Pair selection

Each gradient uses a pair of two distinct images sampled at random from the 10,000, independently for every pair, so images are reused across pairs. We now state this in Sec. 3.1.

Q6: Time and memory

We report per concept time and memory usage, with the configuration of the main experiments (SDXL-DMD, 10,000 images and gradients, App. Y):

StageTime (one H100)Peak GPU memory
Sampling 10,000 images≈ 1.9 h20.7 GiB (batch 4)
Collecting 10,000 gradients≈ 46 min16.1 GiB (batch 1)
PCA< 1 s (CPU)–
SAE training≈ 20 s< 0.5 GiB

Sampling and gradient collection parallelize trivially across GPUs, and steering adds no cost at inference, as it only modifies the token embedding.

Q7: Transfer to other contexts

We study the transfer of individual directions and of whole subspaces to: other contexts for the same concept and, going one step further, to other concepts. Both transfer well.

Individual directions. A direction discovered with the generic prompt induces similar changes when applied to the same concept in other prompts, and even to other concepts (App. X, Fig. 26). For example, a direction found for "A picture of a dog" that turns the dog dark does the same for dogs in longer prompts (on a beach, in a convertible) and for cats, cars, monsters, and people.

Subspaces. On 15 longer prompts that place each concept in a specific scene and style (Sec. 4.4, App. V), steering with external directions discovered with the generic prompt increases diversity almost as much as native directions discovered for each prompt, with similar prompt adherence. The same holds for directions of entirely different concepts (App. X).

Same concept, different promptDiversity ↑CLIP score ↑
Unsteered0.2210.313
Random direction0.2380.312
Native0.3530.300
External (from "A picture of a ⟨concept⟩")0.3390.301
Different conceptDiversity ↑CLIP score ↑
Unsteered0.2940.223
Random direction0.3400.220
Native0.5380.207
External (other four concepts)0.5250.207

I would like the thank the authors for their detailed responses to the reviewers' questions, and running the embedding scaling experiment. My questions are mostly answered, though I agree with Reviewer ftrQ that the evaluations are performed on a limited set of concepts/prompts. Also, I found the questions raised by Reviewer CqnT quite interesting. In particular, the recommended random-direction control beating ELROND PCA on certain settings suggest that the size of the perturbation indeed partially explain the gain (which is what I was aiming to test with the suggested embedding scaling experiments).

All considered, I would like to maintain my positive score of 3.

Reviewer_CqnTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ELROND finds the directions of variation a diffusion model can generate for a given prompt. These set of distributions are termed as "generative manifolds". The manifolds are recovered by:

  1. Generating a large number of images from the same prompt, 10k images.

  2. And then sampling an image, adding noise, and predicting the clean image with the model.

  3. Later, computing the MSE between this prediction and another image from the 10k images and backpropagating to the token embedding of the concept word

  4. Repeat this steps 2-3 on many image pairs: collect many gradients and decompose them with PCA or top-k SAE.

  5. Finally, steering by adding the discovered direction to the concept token embedding.

The authors validate this on a toy diffusion model trained on shapes so that the ground truth manifolds are known. They also show this on SDXL, SDXL-DMD and Flux Schnell for five concepts. The main finding is that the discovered directions change the image much more than random directions. And they give more diversity than SliderSpace while maintaining same text alignment. They also show that they fix mode collapse in DMD.

Critical strengths
  1. The question of what a model can produce for a given prompt and uncover its variations without relying on external semantic embedders like CLIP to define these axes is very interesting. Especially using the model's own gradients to find these directions.

  2. Starting with a toy experiment with a known ground-truth subspace is very principled. Having a ground truth for these kinds of discovery papers with the precision-recall analysis is a great idea.

  3. Overall, the writing is clear and easy to follow.

Critical weaknesses
  1. Some very important works' citations are missing. As I understand it, each gradient is the transpose of the Jacobian of the clean image prediction with respect to the token embedding, times the residual. So technically, the PCA over gradients mostly recovers the top singular directions of the Jacobian weighted by the residuals. SVD of Jacobians has been explored for concept discovery in prior work. Park et al. (NeurIPS 2023), "Understanding the Latent Space of Diffusion Models through the Lens of Riemannian Geometry" and (Chen et al., NeurIPS 2024), "Exploring Low-Dimensional Subspaces in Diffusion Models for Controllable Image Editing". In fact, the LOCO edit paper uses the same Jacobian form as ELROND, but with respect to the noisy image. The author should cite them and position their work against these prior related works. One experiment I suggest is to replace the residual with a random Gaussian vector and compare the recovered subspace and steering results. If the baseline matches the current result, then the image pairs are not adding much beyond the Jacobian.

  2. When showing real model, large model experiments, the choice of noise time step was always using the highest T. At that noise level, I imagine that the gradient is just pulling the mean towards the second image. Then the whole point of the pair of images' story does not hold in this case. The first image is essentially unnecessary because you can just start with pure Gaussian noise. A good experiment would be to replace the first image with a pure Gaussian sample and compare how different the subspace would turn out. If it's not too far off, then the gradients are essentially similar to text-inversion. If that is true, I still think it's interesting, but it would need an honest framing.

  3. In the mode collapse experiments currently the current baselines are SDXL-DMD, and SliderSpace. One very important control baseline that is currently missing, is adding random noise to the token embeddings at the same strength as ELROND (lambda=40), or CADS, which adds annealed noise to the text embedding. This baseline would tell us if the directions that are found through PCA are in any way helpful or if it's just from the large perturbations that are being done to the embedding.

  4. Another very relevant training-free baseline for SDXL DMD diversity improvement is "Distilling Diversity and Control in Diffusion Models" by Gandikota and Bau who show diversity gains on SDXL-DMD without training.

  5. There is a noticeable decline in the prompt fidelity when showing the diversity-related experiments. First, in Table 17, there is a drop in precision across most concepts when compared against the base model. Second, the qualitative examples clearly show that some images do not follow the prompt at all. For example, Figure 21 shows monster images, but there are samples that do not contain any monster. Reporting VQA score or some related metric that shows how many samples do not follow the prompt would show a complete picture of the method

  6. The current evaluations cover only five concepts with a very strict template. I would encourage the authors to show results on longer prompts and more diverse prompts

Questions for authors

Please refer to weaknesses.

4 comments

We thank the reviewer for the concrete suggestions. We ran every requested experiment and address all points below.

W1: Missing citations

Thank you for pointing us those intersting works that we've ommited; both are now discussed in Sec. 2. Both compute directions from the Jacobian w.r.t. the noisy image of a single sample at a single timestep, and edit samples with a prompt, and thus the model's distribution, fixed. ELROND takes the Jacobian w.r.t. the conditioning and aggregates gradients over many samples of the prompt's distribution, so its directions describe how that distribution changes with the conditioning.

We agree that any g=J⊤rg = J^\top r lies in the tangent space, so the Jacobian fixes the space and the residual weights directions within it. To determine whether the images add anything beyond the Jacobian, we ran the suggested experiment, replacing rr by norm-matched Gaussian noise (App. S); we call these gradients Jacobian-only.

In the synthetic setting, where the ground truth is known, the difference is clear:

Synthetic data: est. dim. / precision / recallSquare (dim. 3)Triangle (dim. 4)Circle (dim. 6)
ELROND3 / 0.956 / 0.9564 / 0.939 / 0.9398 / 0.726 / 0.968
Jacobian-only11 / 0.264 / 0.96911 / 0.353 / 0.97012 / 0.486 / 0.971

Jacobian-only gradients recover the ground-truth directions (recall ≈ 0.97) but also many more: their estimated dimension is 11–12 instead of 3–6, and precision drops to 0.26–0.49. They capture every direction the model responds to, including attributes that do not vary for the given shape, while the image residual keeps only the variation actually present in the samples. On SDXL-DMD:

ELROND vs. Jacobian-only subspace (SDXL-DMD)carcatdogmonsterperson
Subspace overlap0.510.530.490.560.43
Overlap of the top-10 directions0.420.350.420.450.41
ELROND gradient variance outside the Jacobian-only subspace42%39%41%33%45%

The two subspaces overlap, but they are distinct: the leading directions differ, and a third or more of the variance of the ELROND gradients lies outside the Jacobian-only subspace.

steering, mean of 5 conceptsFID ↓Precision ↑Recall ↑DreamSim ↑concept kept (VLM) ↑
Jacobian-only49.060.0870.5770.53590.9%
ELROND48.130.0860.6100.53993.1%

Used for steering, the two perform similarly, with ELROND marginally better in FID, recall and concept fidelity. The Jacobian accounts for much of the subspace, and the images change which directions dominate. With Gaussian residuals, PCA finds the directions the output is most sensitive to. With the image residual, which at t=Tt=T is the difference between a generated image and the model's prediction from pure noise for the same prompt, it finds the directions along which the generated images actually vary.

W2: Pair of images' story does not hold for t=Tt=T

We fully agree: at t=Tt=T the first image is noise, and each gradient is close to a textual-inversion step toward a sample of the prompt's distribution. We keep the pair formulation for consistency with the synthetic setting, which uses all timesteps, where the first image does carry information. We now make this clear in App. T.

W3 & W4: Mode collapse baselines

We thank the reviewer for pointing us to the additional baselines. We add the random-direction control, CADS and Distilling Diversity to Sec. 4.3 and App. P. ELROND significantly outperforms both CADS variants. Random-direction at matched strength improve the mean FID less than ELROND (51.85 vs. 47.59 for PCA and 41.99 for SAE), so the size of the perturbation alone does not explain the gain. Distilling Diversity, which runs the teacher at inference, reaches a FID comparable to ELROND with student directions (42.87 vs. 41.99), and ELROND with teacher directions outperforms it on every concept without running the teacher.

FID ↓carcatdogmonsterpersonmean
Random direction (λ=40\lambda=40)33.0928.3030.2596.3771.2151.85
CADS (concept token)31.9026.2929.91112.8388.5757.90
CADS (whole conditioning)35.5545.1743.45162.1682.3273.73
Distilling Diversity$^\dagger$25.9220.5828.07100.2339.5642.87
ELROND PCA (SDXL-DMD)26.5723.6032.95107.4647.3547.59
ELROND SAE (SDXL-DMD)26.3519.9425.1399.6338.9041.99
ELROND SAE (SDXL)25.4819.7224.0695.0533.6139.58

†^\daggerRuns the teacher (SDXL) for the first sampling step.

W5: Prompt following

We measure the share of samples that keep the prompted concept with a VLM judge (Qwen2-VL-7B-Instruct) and report the share on 500 random images per concept and method (Sec. 4.2, App. R).

concept kept (%) ↑carcatdogmonsterpersonmean
SDXL-DMD100.0100.0100.0100.0100.0100.0
SliderSpace94.4100.089.894.2100.095.7
ELROND PCA92.096.491.089.097.093.1
ELROND SAE96.499.699.696.091.296.6

The steering costs some prompt fidelity: the unsteered model always shows the concept, while ELROND SAE loses it in 3.4% of samples, on par with SliderSpace (4.3%), and ELROND PCA in 6.9%. The precision table the reviewer refers to (Tab. 17 in the submission) is Tab. 18 in the revised paper.

W6: Extending the evaluation

We repeat our sensitivity (distance between the image before and after the perturbation) and diversity enhancement (mean pairwise distance between steered images) evaluations on 15 longer prompts (3 per concept, placing the concept in a specific scene and style) and on 3 new concepts (App. U, with example directions). Both results carry over: on every prompt, discovered directions change the image about 2–4× more than random directions of the same norm and increase diversity well beyond them. Directions found on the template "A picture of a ⟨concept⟩" (generic) work on the longer prompts nearly as well as those found on the prompt itself (native).

Long prompts (15)Sensitivity ↑Diversity ↑
Random0.1090.238
Native0.2460.353
Generic0.2270.339
New concepts (bird, house, robot)Sensitivity ↑Diversity ↑
Random0.1570.297
Native0.3760.509

Additionally, we extend our evaluation to SD3.5 Large Turbo (App. V) and RFdiffusion3, a diffusion model that generates 3D protein structures (see App. N).

Thank you for the thorough rebuttal response with new experiments. I have a few follow up suggestions/questions:

  1. The Jacobian-only results suggest that ELROND recovers the variation present in the samples, rather than "ALL" directions the model encodes or responds to (which is how the tangent space is defined in Sec. 3.1). Do the authors agree with my interpretation here? If yes, revising the paper to match this view would be very helpful.

  2. I do not see the t=T clarification in main paper. Since all large-model results use it, and this is an important detail (and short to mention), adding a clause in the main paper would be beneficial to the reader.

  3. The random-direction control baseline already improves the mean FID from 63.50 to 51.85, and beats ELROND PCA on dog and monster. I would soften the claim "the size of the perturbation alone does not explain the gain", add the unsteered row to the table, and report percentage cases where "the queried concept was actually present in the image" for the random and CADS baselines as well.

  4. For the longer prompts, reporting prompt alignment would show whether the rest of the prompt is still followed.

I am raising my score to weak accept given the thorough rebuttal response. I would appreciate a response for my remaining questions. Thanks!

We thank the reviewer for the swift reply, and we are happy to hear that we managed to convince you with our response. Below we would like to address the follow-up suggestions, (page and line numbers refer to the revised PDF).

Q1: What ELROND recovers

We agree with the reviewer's interpretation that ELROND recovers the variation present in the samples, rather than "all" directions the model responds to, and have revised the paper accordingly (Sec. 3.1, p. 5, l. 227–229):

  • Before: "Repeating this process for many pairs gives a dataset of gradients G={g1,…,gM}\mathcal{G} = \{g_1, \dots, g_M\}, representing the semantic directions that span the tangent space of the concept in the embedding space."
  • Now: "Repeating this process for many pairs gives a dataset of gradients G={g1,…,gM}\mathcal{G} = \{g_1, \dots, g_M\} spanning the directions of Tc\mathcal{T}_c along which the prompt's samples vary."

Q2: The t=Tt=T clarification

We added the clarification to the main text, where the image pair is introduced (Sec. 3.1, p. 5, l. 216–217): "At the highest noise level, where zT(i)z_T^{(i)} is close to pure noise, the gradient is essentially a textual-inversion step pulling the prediction toward x(j)x^{(j)} (App. U)."

Q3: Mode collapse baselines

Random-direction control. We agree and soften the claim: a random direction at the same strength already accounts for a substantial part of the gain. The unsteered row is present in Tab. 4 (SDXL-DMD); we only omitted it from our response (sorry for confusion).

FID ↓carcatdogmonsterpersonmean
SDXL-DMD (unsteered)37.5538.7950.25103.9586.9463.50
Random direction33.0928.3030.2596.3771.2151.85
ELROND PCA26.5723.6032.95107.4647.3547.59
ELROND SAE26.3519.9425.1399.6338.9041.99

VLM evaluation of the other baselines. Below we report the share of samples in which the VLM judge finds the queried concept. We have now added those results to the App. S (Tab. 22).

concept kept (%) ↑carcatdogmonsterpersonmean
SDXL-DMD (unsteered)100.0100.0100.0100.0100.0100.0
Random direction100.0100.0100.0100.0100.0100.0
CADS (concept token)100.0100.0100.099.8100.0100.0
CADS (whole conditioning)98.2100.099.899.696.898.9
Distilling Diversity100.0100.0100.0100.099.499.9
ELROND PCA92.096.491.089.097.093.1
ELROND SAE96.499.699.696.091.296.6

Q4: Longer prompts

The CLIP score with the full prompt is already reported in Tab. 27. We now additionally update it, by using the VLM judge to test whether each element of the prompt is present: the concept, the subject's action or attribute, the scene, and the style (App. V, Tab. 28; mean of 15 prompts, 250 seeds each). Our steering lowers each by a few points, similarly to the results with the generic prompt (Q3), so the rest of the prompt is largely followed.

mean of 15 promptsUnsteeredRandomNativeGeneric
CLIP score ↑0.3130.3120.3000.301
VLM: concept present (%) ↑99.599.194.194.0
VLM: subject action / attribute (%) ↑89.690.085.484.9
VLM: scene (%) ↑97.596.893.293.2
VLM: style (%) ↑95.395.093.993.8

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.