{"id":"11ab6d49-e8bc-4795-ae20-2ac064f4fe1e","arxiv_id":"2412.02237","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Head Relevance Vectors, built by counting which attention head best matches each of 34 concept words, reveal concept-specific cross-attention head patterns and enable targeted concept steering in Stable Diffusion.","lead":"This paper introduces Head Relevance Vectors (HRVs), 128-dimensional scores that rank which cross-attention heads in Stable Diffusion respond to each of 34 visual concepts such as color, material, or weather. These vectors give a head-level control knob for text-to-image models, and the authors use them to reduce polysemous word errors, improve image editing, and reduce catastrophic neglect in multi-concept generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HRV construction assumes the 34 concepts cover all visual content, but the paper never measures how often the argmax assigns out-of-set concepts to one of the 34.","rationale":"The reader's designated weakest assumption is the exhaustiveness of the 34 concepts under the argmax tally. I agree this is the most load-bearing concern. The central claim, that HRVs align with human visual concepts and can be reliably constructed, depends on the argmax votes being correctly attributed to the true concept. If the true concept is absent from the 34, the vote is misattributed. The paper's ordered weakening and downstream results provide empirical support on a handful of concepts, and Appendix J shows a single added concept does not perturb the original HRVs, but neither step measures coverage over the actual prompt distribution. A direct coverage check—adding many held-out concepts and measuring ranking shifts, plus margin analysis of argmax confidence—would settle the issue. Because the concern is real but not yet demonstrated to be fatal, the existing CONDITIONAL verdict remains appropriate; a more severe verdict would require evidence that the contamination is large enough to break the observed ordered-weakening separation. The reader's other notes (manual prompt selection, missing error bars) are valid but secondary; they affect the strength of the downstream demonstrations, not the core validity of the HRV construction as directly as the coverage assumption does.","tokens_in":26565,"tokens_out":6708,"duration_ms":80013,"concrete_test":"Reconstruct HRVs using the original 34 concepts plus 15-20 held-out concepts that are common in the 2,100 prompts (e.g., sky, water, ground, action, number, text, fire, smoke, shadow, reflection). For each of the original 34 concepts, compute the Kendall rank correlation between the head-priority ordering from the 34-concept HRV and the 34-concept rows from the 50-concept HRV. If any correlation falls below about 0.9, the original HRVs are being materially contaminated by out-of-set content. Additionally, for the original 34-concept construction, record the margin between the top-1 and top-2 concept strengths at each argmax step; if a large fraction (e.g., >20%) of observations have a margin below a small threshold, the argmax is frequently assigning near-ties and the coverage assumption is unreliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 and Algorithm 1 construct each head relevance vector by, for every head/timestep, computing the cross-attention map against keys for one sampled word per concept and tallying the argmax over the 34 concepts (Eq. 4). This design makes an implicit exhaustiveness assumption: every head-timestep observation must correspond to one of the 34 visual concepts. The 2,100 prompts used for construction include ImageNet classes and PromptHero prompts that likely depict content outside the list (e.g., sky, water, actions, textures, or objects such as coral reef and lighthouse). When such content is present, the argmax still assigns the observation to the closest of the 34 concepts, contaminating the HRV for that concept. The paper does not quantify how often this occurs, nor does it measure the confidence of the argmax (e.g., the margin between top-1 and top-2 concept strengths). Appendix J adds one concept (Tableware) and shows the original 34 HRVs are largely unchanged, which is a useful but narrow check; it does not establish coverage against the prompt distribution. If misattribution is systematic, the claim that HRVs 'reliably' encode human-specified concepts is weakened, because a head that truly encodes an out-of-set concept would have its votes split among unrelated in-set concepts, distorting their vectors. This is the single most load-bearing concern because it threatens the validity of every HRV, not just the downstream applications.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Head Relevance Vectors (HRVs) as concept-specific importance scores for the 128 cross-attention heads of Stable Diffusion v1.4 (and SDXL). HRVs are built by generating 2,100 random images and, for each head and timestep, computing cross-attention maps against the concatenated CLIP key embeddings of one randomly sampled word per concept (34 concepts, 10 words each), argmaxing over concepts, and incrementing the winning concept's entry (Eqs. 2-5); vectors are L1-normalized. The authors validate HRVs with an ordered weakening analysis (MoRHF vs LeRHF), then apply HRV-based rescaling ('concept strengthening' and 'concept adjusting') to reduce polysemous-word misinterpretation, to improve P2P editing on five attributes, and to improve Attend-and-Excite in multi-concept generation.","tokens_in":26819,"tokens_out":9359,"duration_ms":99632,"significance":"If the central claim holds, HRVs provide a simple, training-free, head-level control mechanism for text-to-image diffusion models and a new interpretability probe. The paper's strengths are its clear methodology, released code, the random-order weakening baseline, human evaluations, three downstream task benchmarks, the SDXL extension, and the Appendix J concept-addition check. The main risks are that HRV construction assumes the 34 concepts cover all visual content encountered, the primary validation is partly self-referential, and the quantitative claims lack uncertainty estimates. These issues are fixable and do not, in my view, invalidate the approach, but they currently prevent the stronger conclusions ('effectively and reliably constructed') from being fully supported.","major_comments":[{"comment":"The argmax in Eq. (4) assigns every head-timestep observation to exactly one of the 34 concepts, so any head that actually encodes an out-of-set concept (e.g., sky, water, motion, or an ImageNet/PromptHero category outside the list) has its votes absorbed by the nearest in-set concept, contaminating that HRV. The paper never reports the distribution of argmax margins or any coverage statistic over the 2,100 construction prompts. Appendix J's addition of Tableware shows that the original 34 HRVs are largely unchanged, but this checks robustness to adding one concept, not coverage of the prompt distribution. Please report top-1 vs top-2 margins, the fraction of low-confidence assignments, and a leave-one-concept-out or hold-out-prompt analysis.","section":"Section 3, Algorithm 1, Eq. (4)"},{"comment":"The ordered weakening test uses the same cross-attention maps from which HRVs were estimated, so it is partly self-referential: a summary of which heads attended most to a concept will naturally identify heads whose weakening affects that concept. The random-order baseline in Appendix C.3 mitigates this, but it does not test whether the HRV construction generalizes across prompt distributions. A split-half validation (construct HRVs on one half of the 2,100 prompts, evaluate ordered weakening on the other half) and a random head-weighting baseline for the downstream applications (e.g., uniform or shuffled HRVs in A&E and P2P) would substantially strengthen the evidence that the vectors carry concept-specific information.","section":"Section 4, Figures 3, Appendix C.3"},{"comment":"All quantitative comparisons are reported as point estimates without error bars, confidence intervals, or significance tests, despite the use of multiple seeds and prompts (e.g., 30 seeds, 500/250 edited images, 100 generated images per model in the polysemous evaluation). Improvements of 2.3-6.3% in Table 2 and 2.32-11.79% in Section 5.2 could be within seed or prompt variance. Please report bootstrap or paired tests over seeds and prompt sets for the main comparisons.","section":"Section 5 (Tables 1-2, Figure 7) and Appendix D.2"},{"comment":"The polysemous-word evaluation uses 10 prompts hand-picked from known failure cases, and for each prompt the desired and undesired concepts are manually chosen (e.g., Color vs Plants for 'lavender'). This makes the 63.0% to 15.9% misinterpretation reduction hard to interpret as a general claim. The AMT protocol is well designed, but it evaluates only this curated set. Please broaden the prompt set, pre-specify selection criteria, or run the LLM-based automatic selection pipeline described in Appendix D.1 and report its performance.","section":"Section 5.1, Table 7, Appendix D"},{"comment":"The rescaling coefficients in the core methods are fixed without systematic justification: ordered weakening uses -2, and concept adjusting uses 2*HRV(desired) - 1*HRV(undesired). Appendix C.4 explores the weakening factor, but no sensitivity analysis is given for the concept-adjusting coefficients or for the L1 normalization in SDXL, which Appendix G.3 itself identifies as problematic. A small ablation over these coefficients would clarify how much of the reported gains depend on them.","section":"Section 5 (Figure 4), Appendix C.4"}],"minor_comments":[{"comment":"The method name 'MassaCtrl' should be 'MasaCtrl' to match the reference and the surrounding text.","section":"Section 5.2, Table 1"},{"comment":"The sentence 'The two sets of HRVs for the original 34 visual concepts (Figure 48a and Figure 48b) are highly similar' is misworded; Figure 48b contains the 35-concept set, so the comparison is between the 34-concept and 35-concept constructions.","section":"Appendix J"},{"comment":"The 'Filtering process' column contains a stray 'O' rather than a description; the table should be cleaned up.","section":"Table 8"},{"comment":"The OpenAI reference entry uses the informal URL title 'Hello gpt-4o'; use a formal citation or consistent reference format.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a camera-ready ICLR version, but I evaluated it as a journal submission. The method is promising and the experiments are extensive, yet the main claims currently rest on an unquantified coverage assumption and point estimates without uncertainty. If the requested diagnostics, ablations, and significance tests are added, I would expect the claims to be supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this paper gives you a simple, training-free, per-head control signal for Stable Diffusion, and the evidence that it tracks human-specified visual concepts is decent. The HRV construction—tallying argmax concept attendance across heads and timesteps—is genuinely new relative to the layer-level or token-level manipulation in P2P, Attend-and-Excite, and Structured Diffusion. The downstream gains are modest but consistent across three tasks, and the human evaluation on polysemous words is a real plus.\n\nWhat it does well: the ordered weakening analysis is the right sanity check. MoRHF removes the target concept noticeably faster than LeRHF, and the random-order baseline in Appendix C.3 clears away the worry that any ordering would do. The SDXL extension and the t-SNE timestep analysis are useful robustness checks. The paper is honest about failure cases—the Limitation section names concepts where HRVs do not separate MoRHF/LeRHF, which is more than most interpretability papers admit.\n\nThe soft spots, in proportion. The biggest is the coverage assumption: every head/timestep is forced to vote for one of the 34 concepts. The 2,100 construction prompts include ImageNet classes and PromptHero images that almost certainly contain content outside the list—coral reefs, glaciers, lighthouses, actions, textures. Those votes go to the nearest of the 34 by argmax, which contaminates the HRVs. The paper never measures how often this happens or how confident the argmax is. This is real, and it is load-bearing for HRV purity. That said, I would not call it fatal: the ordered weakening and the downstream tasks still work, which suggests the signal survives the noise. Appendix J adds a concept and shows the original HRVs barely move, but it does not test coverage against the actual prompt distribution.\n\nMinor issues: no error bars or significance tests anywhere; the 10 polysemous prompts are hand-picked from known failures; the rescaling coefficients (2 and -1) are manual. These are worth a round of revision, not a rejection.\n\nFor whom: anyone working on T2I interpretability or training-free steering will find this useful to read and cite. It is a solid subfield advance, not a paradigm shift.\n\nMy recommendation: send it to peer review. The core claim holds up in its own empirical terms; the coverage question is a missing analysis, not a proof of error. Ask the authors to quantify coverage and add error bars, and I would accept.","headline":"A useful, training-free, per-head control signal for Stable Diffusion with decent empirical support; the load-bearing weakness is an unexamined coverage assumption in the HRV construction, not the headline claims.","tokens_in":27415,"tokens_out":2659,"would_cite":true,"duration_ms":28115,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cross-attention heads in text-to-image diffusion models encode human-specified visual concepts, and per-head relevance vectors built without fine-tuning can steer generation.","keywords":["head relevance vectors","cross-attention heads","text-to-image diffusion","mechanistic interpretability","concept steering","polysemous words","image editing","multi-concept generation"],"falsifier":"Construct an HRV for a concept that is not among the 34, such as 'sky' or 'running', using the same pipeline, and perform ordered weakening on it. The claim would be falsified if the most-relevant-first and least-relevant-first weakening curves are indistinguishable, or if adding this distant concept substantially changes the original 34 HRVs; the paper's extension test uses Tableware, which is close to the existing set and therefore not a maximally informative check.","tokens_in":26284,"feed_emoji":"🎨","tokens_out":12250,"duration_ms":113358,"temperature":0.7,"pith_summary":"The paper's central claim is that the cross-attention heads—the connections that link each word in a prompt to image regions—carry stable, human-readable visual concepts, and that those concepts can be read out without changing the model. The authors build a Head Relevance Vector (HRV) for each of 34 visual concepts, with one entry per attention head, by counting which concept each head attends to most strongly while 2,100 random images are generated. Ordered weakening shows the counts are meaningful: removing the highest-scoring heads erases the target concept from generated images much faster than removing the lowest-scoring heads first. HRVs also work as steering controls, and rescaling head attention maps with them reduces misinterpretation of polysemous words, improves difficult image edits, and lessens catastrophic neglect in multi-concept generation. If the claim holds, attention heads give users a direct, fine-grained interface to what a text-to-image model draws.","feed_headline":"Human visual concepts sit in specific diffusion attention heads","feed_subtitle":"Per-head relevance scores fix word mix-ups, hard image edits, and multi-concept dropouts.","key_machinery":"The central object is the Head Relevance Vector (HRV), an $H$-dimensional vector for each visual concept whose $h$-th component records how often the $h$-th cross-attention head was most responsible for that concept during construction. The construction feeds randomly sampled concept-words through the CLIP text encoder and each head's key-projection layer, concatenates the semantic-token embeddings into one key matrix, computes the softmax cross-attention map with the image query, averages over spatial and word-token dimensions, and applies argmax over the $N=34$ concepts so that every head-timestep contributes exactly one scale-free vote. The steering machinery is a rescaling vector: concept strengthening uses the desired concept's HRV as per-head multipliers on the semantic-token attention maps, and concept adjusting uses $2\\cdot r_{\\text{desired}} - r_{\\text{undesired}}$ as the per-head multipliers; in both cases only the attention maps of semantic tokens are rescaled, leaving special tokens and the rest of the denoising process untouched.","core_discovery":"On the paper's own terms, the discovery is that cross-attention head positions in Stable Diffusion have concept-specific relevance patterns that align with human-specified visual concepts, and that these patterns are extractable from forward passes alone. For each concept, the procedure embeds one sampled concept-word per concept, concatenates the key projections, computes the cross-attention map against the image query, spatially averages it, and uses argmax to assign that head-timestep observation to the single most activated concept, incrementing that concept's HRV. The argmax is not a detail: cross-attention heads differ in scale by up to 8.1 times, so argmax converts each head into one scale-free vote. Validation by ordered weakening shows that weakening in most-relevant-first order removes the target concept from generated images markedly sooner than least-relevant-first order, in both Stable Diffusion v1.4 and SDXL. Applying the HRVs as rescaling vectors yields the paper's three application results: human-evaluated misinterpretation of polysemous words falls from 63.0% to 15.9%; image editing improves on five attributes relative to several baselines; and multi-concept generation improves over Attend-and-Excite by 2.3% to 6.3% across the reported metrics.","pith_inferences":["A stronger test than anything in the paper would build an HRV for a concept that lies far outside the 34, such as 'sky' or 'running'; the paper only demonstrates extension with Tableware, a concept close to the original set.","Because construction is driven by CLIP text embeddings and argmax competition, concepts with poor lexical separation—such as facial expressions, which the paper itself reports as a failure case—are the likely weak points of the method.","The finding that timesteps do not separate in the t-SNE of HRVs suggests head-level concept routing is settled globally rather than by a particular denoising phase; if so, HRV steering should transfer to shorter or distilled sampling schedules.","The authors note an LLM can choose the desired and undesired concepts automatically but do not build that pipeline; a natural next step is an end-to-end system that takes a prompt, detects ambiguous tokens, and applies concept adjusting without human selection."],"forward_implications":["A user can obtain concept control over a frozen text-to-image model by running the HRV construction once, then rescaling head attention maps at inference time.","Adding a new visual concept to the set does not materially disturb the existing HRVs, since the paper shows the 34-concept and 35-concept vectors are highly similar.","For ambiguous tokens, concept adjusting with a chosen desired and undesired concept lowers the human-perceived misinterpretation rate from 63.0% to 15.9% on the 10 prompts tested.","Coupling HRV rescaling with Prompt-to-Prompt improves image-text alignment and human preference for edits of color, material, geometric patterns, image style, and weather conditions.","Coupling HRV rescaling with Attend-and-Excite reduces catastrophic neglect in two-prompt-type benchmarks, improving over the baseline by 2.3% to 6.3% on full-prompt similarity, minimum object similarity, and BLIP score."],"supporting_citations":[{"why":"Supplies the Stable Diffusion v1.4 architecture whose cross-attention heads, U-Net, and latent denoising process are the subject of all analyses.","marker":"Rombach et al., 2022"},{"why":"Provides the CLIP text encoder that embeds concept-words and prompts into key matrices, and the CLIP similarity used as the main alignment metric.","marker":"Radford et al., 2021"},{"why":"Prompt-to-Prompt gives the cross-attention replacement and rescaling framework that underlies both ordered weakening and the P2P-HRV editing method.","marker":"Hertz et al., 2022"},{"why":"Attend-and-Excite is the multi-concept generation baseline and attention-guidance procedure that A&E-HRV augments with HRV re-weighting.","marker":"Chefer et al., 2023"},{"why":"ImageNet classes make up 1,000 of the 2,100 prompts used to accumulate HRV counts.","marker":"Deng et al., 2009"},{"why":"Provides the remaining 1,100 diverse prompts used in the same HRV accumulation.","marker":"PromptHero"}],"fun_headline_variants":["Visual concepts trace to specific diffusion attention heads","Attention head positions map to human visual concepts","Decoding visual concepts from diffusion cross-attention heads","Head relevance vectors pinpoint visual concepts in diffusion","Attention heads encode human visual concepts in text-to-image"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the 34 visual concepts, each expressed by 10 concept-words, are an exhaustive partition of what cross-attention heads can encode, because every head-timestep observation is assigned by argmax to exactly one concept and any head that truly encodes something outside the set would have its votes misattributed.","fun_headline_variants_meta":{"raw":{"variants":["Visual concepts trace to specific diffusion attention heads","Attention head positions map to human visual concepts","Decoding visual concepts from diffusion cross-attention heads","Head relevance vectors pinpoint visual concepts in diffusion","Attention heads encode human visual concepts in text-to-image"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000536,"raw_usage":{"total_tokens":2599,"prompt_tokens":994,"completion_tokens":1605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":1535}},"tokens_in":610,"tokens_out":1605,"duration_ms":11816,"temperature":1.0,"reasoning_tokens":1535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:41:59.412512+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct an HRV for a concept that is not among the 34, such as 'sky' or 'running', using the same pipeline, and perform ordered weakening on it. The claim would be falsified if the most-relevant-first and least-relevant-first weakening curves are indistinguishable, or if adding this distant concept substantially changes the original 34 HRVs; the paper's extension test uses Tableware, which is close to the existing set and therefore not a maximally informative check.","supporting_citations":[],"review_version":1}