Concept Reachability in Diffusion Models: Beyond Dataset Constraints

Anastasia Borovykh; Marta Aparicio Rodriguez; Xenia Miscouridou

REVIEW 5 major objections 5 minor 13 references

Concept Reachability in Diffusion Models: Beyond Dataset Constraints

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper establishes a phase transition in concept reachability: below roughly 1% training frequency, concepts become unreachable, while activation steering recovers concepts that prompting cannot.

desk verdict Useful framework and a plausible steering-vs-prompting story, but the headline 1% phase transition is confounded with absolute sample counts and needs to be redone. read the letter →

arxiv 2505.19313 v1 pith:QYSAILSZ submitted 2025-05-25 cs.LG

Marta Aparicio Rodriguez , Xenia Miscouridou , Anastasia Borovykh This is my paper · ORCID

classification cs.LG

keywords conceptreachabilitydiffusionmodelsactivationsteeringphasetransitiondatasetscarcitycaptionunderspecificationdatabiastext-to-imagegeneration

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks when a text-to-image diffusion model can actually generate a concept it was trained on, and whether the answer depends on how the concept is requested — by text prompt or by steering the model's internal activations. Working with a synthetic dataset of coloured shapes where scarcity, caption quality, and bias are varied one at a time, it claims that concept reachability undergoes a phase transition: once a concept appears in fewer than roughly 1% of training images, the model abruptly loses the ability to generate it, while above that threshold extra examples barely help. It also claims that the layer at which steering is applied decides what is reachable — modifying the text embedding is stable, while modifying the U-net's bottleneck is fragile. The practical upshot is that steering recovers concepts prompting cannot reach, such as concepts seen only in the wrong position or concepts entangled by bias, so model providers could ship steering-based controls instead of retraining. The one limitation that binds all methods is that a concept never named in captions is unreachable by any of them.

What carries the argument

The carrying mechanism is the steering vector, a vector optimised on frozen model weights to minimise the denoising loss $\mathbb{E}\lVert \epsilon_t - \epsilon_\theta(x_t, t, y_s, v)\rVert^2$ over an auxiliary image set containing the target concept, then added either to the prompt's text embedding (prompt-space steering, $v_p$) or to the U-net bottleneck output (h-space steering, $v_h$). Reachability is scored by three trained classifiers that read off back shape, front shape, and the back-front colour pair from each generated image, with incomplete images counted as incorrect. The load-bearing phenomenon is the phase transition at $p_X(f) \approx 0.01$: a concept-agnostic, low threshold above which more data adds little and below which reachability drops abruptly. The contrast between the two steering spaces — $v_p$ stable and prompt-matching, $v_h$ fragile with accuracy falling as more concepts are modified and correlating with the optimised vector's final norm — locates the effect at a specific stage of the transformation.

What would settle it

Measure generation accuracy on a production-scale text-to-image model across concepts whose training frequency spans 0.001% to 5%: if reachability declines smoothly with frequency rather than stepping sharply near $p_X(f) \approx 0.01$, or if steering does not beat prompting on positionally out-of-distribution combinations, the paper's two central claims fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that concept reachability — defined as the fraction of generated images that contain the target concept combination $(c_1, s_1, c_2, s_2)$ — is governed by the structure of the training data in a sharp, quantifiable way, and that activation steering is systematically more resilient than prompting. In the synthetic coloured-shapes setting, reachability stays high as a concept's frequency is reduced and then collapses once the concept appears in fewer than about 1% of training images, $p_X(f) \approx 0.01$, a phase transition observed whichever of the four concept positions is thinned. When the target combination is positionally out-of-distribution — each individual concept appears in training but not in the requested position, for instance red shapes only ever appear in front but are requested in back — prompting falls to near zero while prompt-space steering reaches the combination with over 50% accuracy in one of the studied cases. When captions no longer name a concept, reachability drops for both prompting and steering, showing that captions organise the latent space and that a never-captioned concept is effectively unreachable. When two concepts are perfectly tied in the training data, steering, particularly on the prompt space, partially disentangles them, and adding independent examples of one tied concept lifts reachability of both.

Load-bearing premise

The synthetic coloured-shapes dataset and the small 3.7M-parameter model faithfully represent production-scale text-to-image models, so the $p_X(f) \approx 0.01$ threshold and the steering advantage transfer to real systems.

Editorial extensions

If this is right

Concepts present in at least roughly 1% of training images are reliably learnable, so data collection can stop once a concept crosses the threshold instead of chasing large counts.
Steering in the prompt space reaches concepts that prompting cannot, so model providers could ship user-facing steering controls in place of costly retraining and dataset curation.
A concept that never appears in captions is unreachable by every method tested, making caption completeness a hard lower bound on what a model can generate on demand.
Positionally out-of-distribution combinations are mostly lost to prompting but partially recovered by steering, so reachability depends on how a concept is requested, not only on whether it was learned.
The intervention layer itself determines what is reachable: text-embedding steering is stable across targets, while bottleneck steering degrades as more concepts are modified.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the phase transition is generic, comparable sharp thresholds should appear when steering language models toward learned concepts, and the critical frequency may shift with model scale and task complexity rather than staying at 1%.
The correlation between h-space reachability and the optimised vector's final norm suggests a cheap, image-free probe: estimate reachability from vector norms to audit which concepts a model can access.
Repeating the scarcity sweep across model sizes within one architecture would separate two explanations — a property of the learning task versus a property of model capacity — that the paper's single small model cannot distinguish.
A corollary for evaluation practice is that prompt-to-image alignment scores understate model capability, and steering-based evaluation would give a fairer measure of what a text-to-image model has actually learned.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a genuinely useful empirical framework for thinking about when concepts are reachable in text-to-image models, and the steering-vs-prompting comparison is worth engaging with. But the headline claim—a sharp phase transition at around 1% training-data presence—is not actually supported by the experiments, because the design confounds concept frequency with absolute sample count.

The good stuff first. The paper introduces a clean distinction between compositionally OOD and positionally OOD concept combinations, which maps well onto how diffusion models can fail. The synthetic setup is well-controlled: 54 concept combinations of colored shapes on a black background, captions generated from known factors, three separate dataset-level obstacles varied systematically, and a released implementation. The finding that prompt-space steering remains effective under scarcity while prompting collapses is interesting and, based on the qualitative Stable Diffusion examples, plausible. The h-space results are more fragile, but the observation that intervention location matters is a useful point. The underspecification results—that reachability tracks what is actually in the captions—are solid and align with prior work.

Now the soft spots. The central quantitative claim in the abstract and Section 5.2 is the phase transition at p≈0.01. Section 4.1 states that when concept presence is reduced, the total dataset size is preserved by increasing the remaining combinations. That means the proportion p and the absolute number of images containing the concept are varied together. A drop at p=0.01 corresponds to a specific absolute count (roughly 540 images for the concept and roughly 30 for the exact target combination, if the reduction is uniform). The paper reads this as "small number of samples relative to the train set," but that is not what the experiment shows. To support that claim you would need to vary the total dataset size or fix the absolute count and vary the proportion independently. The CelebA check uses a single dataset size and cannot resolve this either. This is a real confound, not a matter of transferability to production-scale models.

Other issues are minor by comparison: results are means over four seeds with no error bars; the 0.01 threshold is drawn by eye; the starting prompts for h-space steering are chosen post-hoc to favor h-space; and the real-data section is qualitative. The abstract's phrase "reliably reachable" is stronger than the evidence.

Who this is for: anyone working on controllable generation, dataset curation, or activation steering in diffusion models. The framework and the negative results about prompting are worth having, but the empirical threshold claim needs to be redone or heavily tempered.

Recommendation: send it to review. A serious referee will ask for variance estimates, a disentangling of frequency from absolute counts, and softer real-data language. Even if the 1% threshold does not survive, the rest of the paper is worth publishing.

Referee Report

5 major / 5 minor

Summary. This paper studies concept reachability in text-to-image diffusion models under three dataset limitations: concept scarcity, caption underspecification, and co-occurrence bias. The authors introduce a synthetic dataset of colored shapes (54 concept combinations, 54,000 images) with known generative factors, train small 3.7M-parameter diffusion models, and compare prompt-only generation against steering in prompt-embedding space and U-Net bottleneck h-space. They report (i) a sharp, phase-transition-like drop in reachability when concept presence falls below roughly 1% of the training data, (ii) that the layer/space where steering is applied changes reachability, and (iii) that steering maintains reachability when prompting fails. They supplement the synthetic study with qualitative Stable Diffusion examples and a small CelebA experiment.

Significance. If the results hold, the paper provides a useful empirical map of when concepts can be accessed by prompting versus activation steering, and it argues for user-facing control mechanisms as an alternative to dataset curation. The synthetic design is a strength: the generative factors are known, the dataset modifications are controlled, and the code is released. The authors also honestly acknowledge limitations in Section 6. However, the central quantitative claims are not yet supported at the precision required: the scarcity experiments conflate relative frequency with absolute counts, the p=0 steering results lack a described source for the steering-image set, and the phase-transition threshold is drawn without uncertainty quantification. The real-data sections are qualitative or use a max-ensemble reporting choice that makes quantitative comparison difficult.

major comments (5)

[§4.1, Appendix B, §5.2] The scarcity manipulation changes pX(f) while preserving the total dataset size (54,000 images), so the relative frequency and the absolute number of images containing the target are varied together. At the claimed threshold p ≈ 0.01, the target combination (red, triangle, green, square) has roughly 540 images of the concept and only about 30 images of the exact combination (since the subset [c1=red]X contains 18 combinations). The design therefore cannot separate a frequency effect from an absolute sample-count effect, and the statement that the model needs 'few data points, relative to the train set size' is unsupported. The authors should run control experiments in which the total dataset size is varied (e.g., scaling all combinations) while holding pX(f) fixed, or report absolute counts and show that the same threshold is obtained across different total sizes.
[§4.2 and §5.2] The steering method in Section 4.2 requires a collection Z of images containing the target concept combination to optimize the vector vp or vh. In Section 5.2, results are reported at pX(f)=0, where the target combination is absent from the training set, but the manuscript does not specify where Z comes from in this condition. This makes the p=0 steering results irreproducible and weakens the claim that steering can reach concepts 'even when no training images contain the reduced concept.' The authors must state the source of Z (e.g., external images, images with the concept in another position, or model-generated images) and confirm that the same Z is used consistently across conditions.
[§5.2, Figures 5 and 14, Appendix G.3] The phase-transition claim is based on mean-over-four-seeds curves with no error bars, no per-seed plots, and a threshold marked 'by eye' with a dashed red line. Given the frequency/abundance confound raised above, the strong wording 'distinct phase transition' and 'concept-agnostic, low threshold' in the contributions is not yet justified. The authors should provide per-seed results or confidence intervals, and ideally use an automated or pre-specified procedure to estimate the threshold and its uncertainty, for both the synthetic and CelebA experiments.
[Appendix G.3, Figure 21] The CelebA scarcity result reports the steering curve as max(prompt space, h-space). Taking the maximum of two methods per point is a form of cherry-picking and prevents a meaningful comparison of steering against prompting, which is the stated purpose of the experiment. The authors should plot both methods separately with error bars, or justify a principled way of selecting one method ahead of time.
[§5.5 and Appendix F] The Stable Diffusion experiments are presented as evidence that 'concepts often remain reliably reachable through steering' in real settings, but the evaluation is qualitative: 50 images per concept are generated and a few sample panels are shown, with no quantitative accuracy or counts reported. The text acknowledges that steering sometimes fails (e.g., Figure 18), yet the summary claim is stronger than the evidence. The authors should either report the fraction of successful/improved images for the 50-image evaluation or explicitly label this part as anecdotal and move the quantitative claim to the synthetic and CelebA results.

minor comments (5)

[§4.3] The text says 'Reachability, as defined in Section 5', but Definition 3.1 is in Section 3.4; please correct the cross-reference.
[§5.1] The paragraph on starting-prompt selection is ambiguous: it says the starting prompt that maximizes h-space performance is chosen, then says 'unless stated otherwise' the starting prompts describe the target concepts. Please clarify which starting prompt is used for each experiment and whether the h-space choice was made on held-out targets or on the same targets reported later.
[Figures 5 and 13] The p-axis ranges differ between the main figure (0 to 0.08) and Appendix E.3 (0 to 0.175), making visual comparison of the threshold behavior across figures harder. Please use consistent axis ranges or clearly annotate the difference.
[Contributions bullet list] The contribution states a 'concept-agnostic, low threshold', but only four concept values (red, green, square, triangle) are varied in the scarcity experiments. Please soften this to 'observed for the tested concepts' or provide evidence across a wider set of concepts.
[Appendix G.3] The CelebA figure caption says the dotted red line marks threshold 0.01, but no error bars or standard deviations are shown for the four seeds; please add them or state why they are omitted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claims are empirical and not forced by the paper's definitions, fits, or citations.

full rationale

The paper's central claims are experimental observations rather than derivations, and I found no step in which a result reduces by construction to its inputs. (i) Concept reachability is defined in Definition 3.1 as the measured accuracy of generated images, and the phase-transition claim in Section 5.2 is an empirical reading of Figure 5, not an algebraic consequence of the definition or of the steering losses in Section 4.2. (ii) The steering vectors are fitted to images containing the target concept combination, but the paper then evaluates whether adding the fitted vector during sampling actually produces the target concepts; the fit does not by construction determine the reported accuracy. The acknowledged reliance on an auxiliary image set Z is a practical limitation of the method (discussed in Section 6), not a circular derivation. (iii) The reference list contains no self-citations by Aparicio Rodriguez, Miscouridou, or Borovykh, and no load-bearing uniqueness theorem or ansatz is imported from the authors' prior work. (iv) The skeptic's concern that the 1% threshold is confounded with absolute sample counts is a validity or generalizability criticism of the experimental design, not a circularity: varying p while preserving total dataset size changes absolute counts, but this does not make the observed accuracy curve a tautology or a renamed input. Similarly, the real-data sections are qualitative or use small subsets, which is a strength-of-evidence concern rather than circularity. Overall, the derivation chain is self-contained: the definitions and losses are used as measurement machinery, and the conclusions are not equivalent to the inputs by definition.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claims rest on domain assumptions about the generative structure (concepts as factors, injective image and caption functions) and on the fitted steering vectors; no invented entities are introduced.

free parameters (2)

Steering vector vp (prompt space) = R^{1x10x512} (varies with caption length)
Optimized via Adam to minimize denoising loss on target concept images; the steering results depend on this fitted vector.
Steering vector vh (h-space) = R^{128x8x8}
Optimized similarly; h-space reachability is sensitive to this vector.

assumptions (3)

domain assumption Images are generated by a set of concept variables (F_i) via an injective function g, so each concept combination uniquely determines an image.
Section 3.2; this is the foundation of the synthetic dataset and of the reachability definition, and may not hold for real images.
domain assumption Captions are produced by an injective function h, so the concepts of interest are fully specified by the caption.
Section 3.2; later relaxed in underspecification experiments, but the idealised setup assumes it for balanced data.
ad hoc to paper The three selected dataset obstacles (scarcity, underspecification, bias) are the dominant causes of prompting failure in real models.
Section 1 and Section 5.5; the paper selects these three based on prior work and anecdotal Stable Diffusion examples, but does not demonstrate they are exhaustive or the most important.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Concept Reachability in Diffusion Models: Beyond Dataset Constraints." pith.science (2026). https://pith.science/paper/QYSAILSZ

@misc{pith2026250519313,
  author       = {Pith},
  title        = {Pith review of: Concept Reachability in Diffusion Models: Beyond Dataset Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYSAILSZ}},
  note         = {Machine review of arXiv:2505.19313}
}

read the original abstract

Despite significant advances in quality and complexity of the generations in text-to-image models, prompting does not always lead to the desired outputs. Controlling model behaviour by directly steering intermediate model activations has emerged as a viable alternative allowing to reach concepts in latent space that may otherwise remain inaccessible by prompt. In this work, we introduce a set of experiments to deepen our understanding of concept reachability. We design a training data setup with three key obstacles: scarcity of concepts, underspecification of concepts in the captions, and data biases with tied concepts. Our results show: (i) concept reachability in latent space exhibits a distinct phase transition, with only a small number of samples being sufficient to enable reachability, (ii) where in the latent space the intervention is performed critically impacts reachability, showing that certain concepts are reachable only at certain stages of transformation, and (iii) while prompting ability rapidly diminishes with a decrease in quality of the dataset, concepts often remain reliably reachable through steering. Model providers can leverage this to bypass costly retraining and dataset curation and instead innovate with user-facing control mechanisms.

Figures

Figures reproduced from arXiv: 2505.19313 by the authors.

**Figure 1.** Images generated by Stable Diffusion (Rombach et al., 2022) that fail to produce the desired outcome due to hypothesised dataset limits: (L) a scarce concept, (C) underspecification in a caption, (R) biases. See Appendix F.2 for details. Existing work has shown that the dataset structure plays a key role in reachability. For GANs and VAEs, Zhao et al. (2018) conclude that biases in the train set will influence the g… view at source ↗

**Figure 2.** Visualisation of the structure of the dataset ac [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Concepts of interest in the dataset. The concepts of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (19 more)

**Figure 4.** Figure 4: Reachability to different concept combinations [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Accuracy of prompting and steering on the prompt [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: a) Average reachability across 10 randomly chosen concept combinations [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Reachability to concept combinations containing [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: Images sampled from the same random seeds used [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 9.** Figure 9: Diagram of spaces in the architecture where steering is implemented. A: prompt space, the concept vector is added [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: Examples of image-caption pairs from our synthetic dataset. Each image shows two shapes of different colours, [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

**Figure 11.** Figure 11: Relation between accuracy of steering on the prompt space and [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

**Figure 12.** Figure 12: Steering accuracies for different starting prompts [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

**Figure 13.** Figure 13: Accuracy of prompting and steering on the prompt space and [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

**Figure 14.** Figure 14: Accuracy of prompting and steering on the prompt space and [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]

**Figure 15.** Figure 15: Average reachability across 10 randomly chosen concept combinations [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

**Figure 16.** Figure 16: Reachability to concept combinations containing only non-red triangles in the front (vertical axis) and only red [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]

**Figure 17.** Figure 17: Comparison between sampling on Stable Diffusion with no steering (top row) and with additional steering on the [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]

**Figure 18.** Figure 18: Comparison between sampling on Stable Diffusion with no steering (top row) and with additional steering on [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]

**Figure 19.** Figure 19: Comparison between sampling on Stable Diffusion with no steering (top row) and with additional steering on the [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]

**Figure 20.** Figure 20: a) Reachability to different concept combinations when prompting and when steering from the starting prompt [PITH_FULL_IMAGE:figures/full_fig_p023_20.png]

**Figure 21.** Figure 21: Accuracy of prompting and steering from the starting prompt [PITH_FULL_IMAGE:figures/full_fig_p023_21.png]

**Figure 22.** Figure 22: a) Average accuracy of prompting and steering to the different concept combinations for different levels [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages

[1]

a{s 1}behind a{c 2} {s2}

Removec 1: replace the caption with “a{s 1}behind a{c 2} {s2}”

work page
[2]

a{c 1}shape behind a{c 2} {s2}

Removes 1: replace the caption with “a{c 1}shape behind a{c 2} {s2}”

work page
[3]

a{c 1} {s2}behind a{s 2}

Removec 2: replace the caption with “a{c 1} {s2}behind a{s 2}” 15 Concept Reachability in Diffusion Models: Beyond Dataset Constraints

work page
[4]

a{c 1} {s2}behind a{c 2}shape

Removes 2: replace the caption with “a{c 1} {s2}behind a{c 2}shape”

work page
[5]

a{c 2} {s2}

Removec 1 ands 1: replace the caption with “a{c 2} {s2}”

work page
[6]

Removec 1,s 1 andc 2: replace the caption with “a{s 2}”

work page
[7]

a red triangle behind a green square

Removec 1,s 1,c 2 ands 2: replace the caption with the empty string, “” E. Additional Experiments E.1. Additional Analysis on a Balanced Data Final norm of optimised concept vector is indicative of reachability when steering on the h-spaceThe differences in the accuracies achieved through steering on the h-space suggest that, from a given starting prompt ...

work page 2024
[11]

a person{ ˆh}

Removeg: replace the caption with “a person{ ˆh}”

work page

Show all 13 references

[12]

Remove ˆh: replace the caption with “a{g}”
[13]

” We steer from the starting prompt describing only seen concepts. For example, when removing ˆh from the captions, to steer to (woman,wearing a hat) we use ys = “a woman

Removegand ˆh: replace the caption with the empty string, “” We steer from the starting prompt describing only seen concepts. For example, when removing ˆh from the captions, to steer to (woman,wearing a hat) we use ys = “a woman”. Additionally, we compare the accuracy of prom...
[2017]

cc/paper_files/paper/2017/file/ 3f5ee243547dee91fbd053c1c4a845aa-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2017/file/ 3f5ee243547dee91fbd053c1c4a845aa-Paper. pdf. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lam- bert, N., Rasul, K., Davaadorj, M., Nair, D., Paul, S., Berman, W., Xu, Y ., Liu, S., and Wolf, T. Diffusers: ...

2017 arXiv
[2018]

a {c1} {s1} behind a {c2} {s2}

URL https://proceedings.neurips. cc/paper_files/paper/2018/file/ 12 Concept Reachability in Diffusion Models: Beyond Dataset Constraints 5317b6799188715d5e00a638a4278901-Paper. pdf. Zheng, X., Pang, T., Du, C., Jiang, J., and Lin, M. Intriguing properties of data attribution o...

2018 arXiv
[2022]

ISBN 9781713871088

Curran Associates Inc. ISBN 9781713871088. M´esz´aros, A., Ujv ´ary, S., Brendel, W., Reizinger, P., and Husz´ar, F. Rule extrapolation in language modeling: A study of compositional generalization on ood prompts. InAdvances in Neural Information Processing Systems, volume 37,...

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

a{s 1}behind a{c 2} {s2}

Removec 1: replace the caption with “a{s 1}behind a{c 2} {s2}”

work page

[2] [2]

a{c 1}shape behind a{c 2} {s2}

Removes 1: replace the caption with “a{c 1}shape behind a{c 2} {s2}”

work page

[3] [3]

a{c 1} {s2}behind a{s 2}

Removec 2: replace the caption with “a{c 1} {s2}behind a{s 2}” 15 Concept Reachability in Diffusion Models: Beyond Dataset Constraints

work page

[4] [4]

a{c 1} {s2}behind a{c 2}shape

Removes 2: replace the caption with “a{c 1} {s2}behind a{c 2}shape”

work page

[5] [5]

a{c 2} {s2}

Removec 1 ands 1: replace the caption with “a{c 2} {s2}”

work page

[6] [6]

Removec 1,s 1 andc 2: replace the caption with “a{s 2}”

work page

[7] [7]

a red triangle behind a green square

Removec 1,s 1,c 2 ands 2: replace the caption with the empty string, “” E. Additional Experiments E.1. Additional Analysis on a Balanced Data Final norm of optimised concept vector is indicative of reachability when steering on the h-spaceThe differences in the accuracies achieved through steering on the h-space suggest that, from a given starting prompt ...

work page 2024

[8] [11]

a person{ ˆh}

Removeg: replace the caption with “a person{ ˆh}”

work page