{"id":"134a0953-6cc4-4673-b3c6-fc386951e96e","arxiv_id":"2411.16969","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A scale-conditioned latent diffusion model with a cross-magnification conditioning space generates multi-scale pathology patches, large coherent 4096x4096 images, and super-resolved samples, with denoiser features that improve MIL classification.","lead":"ZoomLDM trains one latent diffusion model to generate histopathology image patches across eight magnification levels, using self-supervised embeddings and a shared conditioning space, then jointly samples scales to produce 4096x4096 images and perform super-resolution without retraining. A generalist reader should care because it targets a core obstacle in gigapixel imaging domains: generating images that are both globally coherent and locally detailed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4096px coherence claim hinges on an unvalidated orthogonality assumption for the VQ-f4 decoder Jacobian; without a quantitative check, the joint-sampling gradient direction may not match the true constraint gradient.","rationale":"The reader's weakest assumption is the same as mine; the joint-sampling approximation is the most load-bearing because it underpins two of the paper's headline capabilities. The paper does provide other support: it releases code, reports CDM-based novelty checks against memorization, and the patch-level FID improvements at data-scarce scales are large. However, none of those validate the 4k coherence mechanism. My proposed test is tractable because backprop through a single decoder at 1024x1024 is feasible, and the Jacobian check can be done offline with a few GPU hours. A negative result would not invalidate the multi-scale conditioning contribution but would require removing or reframing the 4096 and zero-shot SR claims; a positive result would leave the CONDITIONAL verdict's main technical risk resolved.","tokens_in":20891,"tokens_out":5388,"duration_ms":52429,"concrete_test":"Run the 1024x1024 joint sampling (4 patches + context) twice: (1) with the exact backpropagated e from Eq. 9, and (2) with the finite-difference encoder approximation from Eq. 12, using identical seeds, λ, δ, ζ, and 50 DDIM steps. Report Crop FID (as in Table 2) and the mean per-step reduction of C = ||ADec(ẑ0) - Dec(ẑ0^L)||^2 for both variants. In parallel, on 100 held-out latents from the VQ-f4 encoder, compute e_exact and e_approx for random e_img directions and report the cosine similarity. If the Crop FID differs by more than 1.0 or the constraint C does not monotonically decrease under the approximate direction, the orthogonality assumption is invalidated and the 4096 claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of globally coherent 4096x4096 synthesis and zero-shot 4x super-resolution rests on the joint multi-scale sampling algorithm (Supp. 8.2, Eqs. 9-12). To avoid backpropagating through the decoder, the authors replace the exact error direction e = (∂Dec/∂z0)^T e_img with a finite-difference estimate through the encoder: e ≈ [Enc(Dec(ẑ0)+ζe_img) - Enc(Dec(ẑ0))]/ζ. This is exact only if the decoder Jacobian is orthogonal and the encoder is its inverse (i.e., Dec is a norm-preserving orthogonal map and Enc ≈ Dec^{-1}). Neither the VQ codebook objective nor the KL term of VQ-f4 enforces this; it is a strong structural assumption about a deep convolutional autoencoder. The paper provides no quantitative check: no measurement of the angle between e_exact and e_approx, no verification that the approximate update decreases the constraint C = ||ADec(ẑ0) - Dec(ẑ0^L)||^2, and no comparison against the backprop version even at a tractable 1024x1024 scale. The supplementary itself concedes visible stitching discontinuities in 4096 outputs (Sec. 9.4), so the 'global coherency' claim is already qualified. If the approximation is inaccurate, the guidance steps do not follow a valid descent direction, and the 4096 coherence and SR results have no guaranteed mechanism; the observed quality could be a byproduct of the unconditional prior rather than the proposed self-guidance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ZoomLDM, a latent diffusion model trained on histopathology patches at eight magnifications, conditioned on a learned cross-magnification latent space built from SSL embeddings via a trainable Summarizer. A Conditioning Diffusion Model (CDM) is used to sample conditions without real images. The authors introduce joint multi-scale sampling to generate 4096x4096 images and perform zero-shot 4x super-resolution, and they extract multi-scale denoiser features for multiple instance learning. The paper claims state-of-the-art patch FID at all scales, globally coherent 4k synthesis under a tractable budget, and improved MIL performance.","tokens_in":21240,"tokens_out":5143,"duration_ms":47638,"significance":"If the central claims hold, the work is significant for large-image generative modeling in histopathology and satellite domains: a single weight-shared model with scale conditioning, a learned cross-scale latent space, and a tractable joint sampling scheme would be a practical step beyond patch-based or cascaded approaches. The paper has concrete strengths: code release, satellite-image generalization, a CDM for data-scarce novel sampling, ablations on the SSL encoder and summarizer, and an 8-minute 4k synthesis time versus hours for baselines. However, the load-bearing approximation in the joint sampling algorithm is unvalidated, and several stated claims are stronger than the quantitative evidence (notably Table 2 at 4096x4096 and the reconstruction-like behavior noted in Supp. 9.3).","major_comments":[{"comment":"The finite-difference approximation e ≈ [Enc(Dec(z0)+ζ e_img) − Enc(Dec(z0))]/ζ is exact only if the decoder Jacobian is orthogonal and Enc ≈ Dec^{-1}. The paper asserts this is enforced by VQ or KL training, but provides no quantitative check. Since Algorithm 2 is the mechanism behind the 4096x4096 coherence and 4x super-resolution claims, please report (a) the cosine similarity between the exact backpropagated e and the approximate e at representative timesteps and scales, (b) the decrease of the constraint C = ||A Dec(z0) − Dec(z0^L)||² over the sampling trajectory, and (c) for a tractable 1024x1024 case, a direct comparison against the backprop-based version of [16]. If the approximation fails, the guidance steps need not follow a valid descent direction, and the observed coherence may be a byproduct of the prior rather than of self-guidance.","section":"Supplementary Section 8.2, Eqs. (9)-(12)"},{"comment":"Supp. 9.3 states that for magnifications below 5x the model 'can almost perfectly replicate the source image' because the SSL embeddings contain enough information to reconstruct the patch. This raises the concern that the large low-scale FID gains in Table 1 (e.g., 0.15625x: 13.42 vs. 106.14) partly reflect near-reconstruction of the conditioning real image rather than novel sample quality. The CDM row already addresses fully novel conditions, but for the real-condition samples the paper should report a novelty/diversity metric (e.g., distance to the nearest training neighbor) and, ideally, an FID computed over generated images whose conditions are sampled from the CDM only. The abstract's 'state-of-the-art generation quality across all scales' is stronger than the current evidence.","section":"Table 1 vs. Supplementary Section 9.3"},{"comment":"ZoomLDM's 4096x4096 CLIP FID (6.75) and Crop FID (18.90) are worse than both Graikos et al. (2.75 and 11.30) and ∞-Brush (2.63 and 14.76). The paper argues for qualitative superiority and practical inference time (8 min vs. hours), which is a legitimate strength, but the phrase 'we are the first to generate 4k pathology images with both detail and global coherency' is not supported by the reported metrics. Please either add a quantitative evaluation that directly supports the coherence/detail claim (e.g., layout or structure metrics, or a blinded human study) or explicitly restrict the state-of-the-art quality claim to 1024x1024.","section":"Table 2, 4096x4096 row"},{"comment":"The MIL features are extracted from ZoomLDM's U-Net while conditioning on UNI embeddings (Section 4.5; see also Supp. Table 7). Comparing these features against UNI's own features is not a fair head-to-head because UNI embeddings are an input to the proposed feature extractor; the improved AUC may partly reflect the conditioning encoder's information being re-read out rather than a genuinely new multi-scale representation. A fair baseline would fuse UNI embeddings with a simple learned projector inside the same ABMIL pipeline, or use a different conditioning encoder for ZoomLDM (e.g., HIPT) and compare against that encoder's raw features.","section":"Section 4.5 and Table 4"}],"minor_comments":[{"comment":"The sentence 'Our LDM configuration consists of a VQ-f4 autoencoder and a U-Net model pre-trained on ImageNet' is ambiguous; please state explicitly whether the U-Net is initialized from an ImageNet latent diffusion checkpoint (which one?) and what is trained from scratch.","section":"Section 4.1.1"},{"comment":"There is a typo in the supplementary heading ('Summerizer' should be 'Summarizer'), and the figure caption uses 'magt' where 'magnification embedding' is meant.","section":"Supplementary Section 8.1 and Figure 5"},{"comment":"The term 'Conditioning Diffusion Model' is used in Section 3.1, but 'Condition Diffusion Model' appears in the Conclusion; please unify the terminology.","section":"Throughout"},{"comment":"The caption refers to 'SoTA' models, but for the lower magnifications the baselines are standalone models trained by the authors with the same architecture; please clarify which entries are published SoTA and which are internal baselines.","section":"Table 1 caption"},{"comment":"The admitted stitching discontinuities along patch borders and unresolved thin structures should be acknowledged in the main text near the 'globally coherent' 4096x4096 claim, not only in the supplementary.","section":"Supplementary Section 9.4"}],"recommendation":"major_revision","confidential_remarks":"The authors should be encouraged to run the validation experiments for the orthogonality/finite-difference approximation before publication. If the approximation is shown to be poor, the central 4k-coherence and super-resolution contributions would need substantial rethinking. The MIL comparison also deserves a more carefully matched baseline before the strong outperformance claim is made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. ZoomLDM is a real contribution: a single latent diffusion model with shared weights across eight magnification levels, conditioned on SSL embeddings via a learned summarizer. Training one model across scales and sharing data from abundant high-mag patches to data-scarce low-mag patches is the right idea, and the FID gains at low magnifications are credible. The conditional diffusion model for sampling the conditioning latent without a reference image is a nice touch, and the satellite results suggest the approach generalizes. The 1024x1024 results genuinely beat the baselines on both CLIP FID and Crop FID, and at 8 minutes per 4096x4096 image the method is the only practical option at that size. The supplementary is thorough, and the memorization check is decent.\n\nThe paper also has honest self-awareness: the supplementary admits stitching discontinuities in 4k outputs and the memorization check is decent.\n\nThat said, the strongest claims outrun the evidence. First, the 4096x4096 numbers in Table 2 are worse than both baselines on both metrics (CLIP FID 6.75 vs 2.75/2.63; Crop FID 18.90 vs 11.30/14.76). The authors fall back on qualitative superiority, which is fine, but it should be framed that way and not as SOTA. Second, the low-scale FID gains partly reflect the fact that the SSL embeddings essentially encode the patch content; at magnifications below 5x the model almost perfectly replicates the source image (Supp. 9.3). The CDM row, which measures true novel generation, is substantially worse and at 20x/10x/5x it is actually worse than the single-scale baselines. So the 'state-of-the-art across all scales' claim holds only with reference embeddings. Third, the MIL comparison is partly circular: the ZoomLDM feature extractor is conditioned on UNI embeddings, so it is expected to carry UNI's information; beating UNI is less surprising than the paper suggests.\n\nThe most load-bearing concern is the joint-sampling gradient approximation. The finite-difference update through the encoder (Eq. 12) is only valid if the VQ-f4 decoder Jacobian is approximately orthogonal, which is a strong structural assumption with no supporting evidence. There is no comparison to exact backprop even at 1024x1024, and the supplementary's own admission of stitching artifacts complicates the coherence claim. If the assumption fails, the guidance steps are not following a valid descent direction and the 4k coherence is not guaranteed.\n\nWho is this for? Researchers working on generative models for histopathology or satellite imagery. It deserves a serious referee. My recommendation is to engage, but ask for: (1) a backprop-vs-approximation validation at 1024, (2) reporting CDM-based FID alongside embedding-conditioned FID without overclaiming, and (3) a non-circular MIL baseline. After those revisions, the core method stands.","headline":"Solid multi-scale diffusion method with credible mid-scale gains, but the 4k coherence and state-of-the-art claims rest on unverified approximations and favorable evaluation choices.","tokens_in":21795,"tokens_out":4431,"would_cite":true,"duration_ms":41107,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single latent diffusion model, conditioned by scale through a shared embedding space, generates pathology patches at eight magnifications and, sampled jointly, yields coherent 4096×4096 images, 4× super-resolution, and…","keywords":["latent diffusion models","multi-scale image generation","histopathology whole slide images","self-supervised conditioning","cross-magnification latent space","joint multi-scale sampling","zero-shot super-resolution","multiple instance learning"],"falsifier":"Compute the empirical singular-value distribution of the VQ-f4 decoder Jacobian, $J = \\partial \\mathrm{Dec}(\\hat{z}_0)/\\partial \\hat{z}_0$, at real 20× pathology latents (via finite-difference products $J^\\top J$ or a few Hutchinson-style estimates); if the singular values deviate substantially from 1, the identity $\\left(\\partial \\mathrm{Dec}/\\partial \\hat{z}_0\\right)^\\top e \\approx \\left(\\partial \\hat{z}_0/\\partial \\mathrm{Dec}\\right) e$ behind Eq. 12 fails. A behavioral check that is feasible today: run 1024×1024 joint sampling with the paper's Algorithm 2 and with the exact backpropagation-based algorithm of [16], and compare CLIP and crop FID — statistical equivalence validates the approximation, while a gap shows the claimed coherence is not attributable to the stated mechanism.","tokens_in":20703,"feed_emoji":"🔬","tokens_out":16774,"duration_ms":131672,"temperature":0.7,"pith_summary":"ZoomLDM claims that one latent diffusion model, trained on 256×256 patches cut from whole-slide images at eight magnifications, can generate realistic histopathology images at any of those scales, and that sampling several scales jointly produces globally coherent 4096×4096 images in about eight minutes. The mechanism is a cross-magnification latent space: a trainable 'Summarizer' transformer compresses the SSL embedding array of a tissue region, together with a magnification token, into a fixed set of conditioning tokens, so a single weight-shared denoiser serves all eight scales and the data-scarce low magnifications inherit strength from the data-rich ones (FID 13.4 versus 106.1 at 0.15625× with real conditions). A second, small diffusion model learns to sample that conditioning space, allowing novel images to be drawn with no reference slide. If the claims hold, multi-scale generation, zero-shot 4× super-resolution, and slide-level classification features all come from one checkpoint instead of per-scale models.","feed_headline":"A single diffusion model now generates coherent 4K pathology images","feed_subtitle":"The same model shares weights across eight scales: 4K synthesis, super-resolution, and slide classification in minutes.","key_machinery":"The load-bearing object is the cross-magnification latent space produced by the Summarizer, a 12-layer transformer that consumes the array of UNI SSL embeddings describing a region (up to 8×8 tokens, padded or pooled to a fixed size) together with a learned magnification embedding, and emits 65×512 conditioning tokens that the U-Net denoiser attends to. Projecting every scale into this shared space lets one weight-shared denoiser train across magnifications, and lets a Conditioning Diffusion Model (a Diffusion Transformer trained afterward) draw novel conditions from the space. The second mechanism is joint multi-scale sampling, an adaptation of a fast constrained-sampling algorithm for diffusion models in which the current estimate of the coarser image $\\mathrm{Dec}(\\hat{z}_0^L)$ plays the role of the measurement $y$ for the finer patches, with the constraint error propagated through the linear downsampling operator $A$; the expensive backpropagation through the decoder is replaced by a finite-difference step through the VQ-f4 encoder, justified by the assumption that the VQ-f4 autoencoder (the 4×-compressing VAE of the base latent diffusion model) has an approximately orthogonal decoder Jacobian (Supplementary Section 8.2, Eqs. 10–12).","core_discovery":"The central claim is that scale can be a conditioning variable in a latent diffusion model: one weight-shared denoiser generates patches at eight histopathology magnifications, 20× down to 0.15625×, and does so at the lowest FID reported at every scale, with the biggest margins where training data is thinnest. Because all scales share one learned conditioning space, the authors then sample scales jointly, using the current estimate of a coarse 256×256 context image at 1.25× as a constraint that guides the sixteen-by-sixteen 20× patches through the linear downsampling relation $x_{s+1}=A x_s$; the result is 4096×4096 images that keep global tissue structure and local detail, which they state is the first practical 4K histopathology synthesis: about 8 minutes per image, where prior methods took 4–12 hours and produced either blur or contextually wrong stitching. The same joint sampler, fed a single low-resolution image through an inversion step that optimizes conditioning embeddings (adapting textual inversion), performs 4× super-resolution with no super-resolution training. Finally, the paper argues the denoiser's intermediate features are themselves strong representations: fusing 20× and 5× ZoomLDM features raises BRCA-subtyping AUC to 94.91 and HRD AUC to 88.03, above UNI, Phikon, and CTransPath, and 20× features alone already beat UNI's 20× features.","pith_inferences":["Testable extension beyond the paper: the Conditioning Diffusion Model can in principle be trained on paired (summarizer-latent, biomarker) data, so that sampled conditions carry a specified mutation or grade; the paper's closing remarks gesture at this, and a concrete check is whether CDM-sampled images of BRCA-mutant versus wild-type slides differ in the expected histological features.","The orthogonality assumption in Supplementary Section 8.2 is the part most worth probing: if the VQ-f4 decoder Jacobian is far from orthogonal, the encoder finite-difference 'gradient' is a heuristic rather than the constraint gradient, and the observed 4K coherence may come mostly from the model's prior; a controlled comparison of Algorithm 2 against the exact backprop-based algorithm at 1024×102","Because the method couples scales by a linear downsampling constraint, joint three-scale (or deeper) sampling should compound coherence; the paper demonstrates two-scale pyramids, and a natural stress test is whether a 2.5×→5×→20× pyramid at 8192×8192 stays coherent or accumulates drift, especially given the patch-border stitching the paper itself notes at 4K.","The result that single-scale 20× features beat the very SSL encoder they were conditioned on suggests generation acts as representation refinement; a testable consequence is that fine-tuning ZoomLDM on a new tissue type should improve downstream MIL accuracy more than fine-tuning the conditioning SSL encoder alone."],"forward_implications":["A single ZoomLDM checkpoint replaces eight per-magnification generative models for histopathology, and the shared-weights training transfers strength to scarce scales: FID at 0.15625× drops from 106.1 to 13.4, and even CDM-sampled conditions (26.1) beat the single-scale baseline.","Joint multi-scale sampling makes 4096×4096 histopathology synthesis tractable at roughly 8 minutes per image on the reported hardware, where patch-stitching took about 4 hours and infinite-dimensional diffusion about 12 hours; the paper reports its 4K CLIP/crop FID trails the latter, while arguing that local detail and global structure are both present for the first time.","Zero-shot 4× super-resolution: given a 256×256 image at 5×, ZoomLDM infers conditioning by inversion and generates 1024×1024 output that beats diffusion super-resolution baselines (CompVis, ControlNet, ResShift, StableSR) on LPIPS and pathology-similarity metrics, without any SR-specific training.","Multi-scale denoiser features improve slide-level classification: 20×+5× fusion reaches 94.91 AUC for BRCA subtyping and 88.03 for HRD prediction, above the SSL encoders used to condition the model, implying that generative multi-scale training refines the representations of the conditioning encoder.","The recipe transfers to a second gigapixel domain, satellite imagery: four-resolution NAIP generation lowers crop FID for 1024×1024 synthesis from 43.8/48.7 to 13.3, supporting the paper's claim that the mechanism is domain-general."],"supporting_citations":[{"why":"Supplies the fast constrained-sampling algorithm for diffusion models under linear constraints that joint multi-scale sampling modifies.","marker":"[16]"},{"why":"The learned-representation-guided patch diffusion baseline ZoomLDM extends, and the main comparison for patch FID and large-image generation.","marker":"[17]"},{"why":"Provides the latent diffusion framework, the VQ-f4 autoencoder, the U-Net architecture, and the training code and checkpoints ZoomLDM builds on.","marker":"[39]"},{"why":"The UNI self-supervised pathology encoder whose patch embeddings serve as the conditioning signal, paired with patches at all eight magnifications.","marker":"[8]"},{"why":"The infinite-dimensional diffusion baseline (∞-Brush) compared on 1024×1024 and 4096×4096 large-image generation.","marker":"[26]"},{"why":"The prior multi-scale pathology diffusion model that lacked a conditioning mechanism, the gap ZoomLDM's conditioning design targets.","marker":"[18]"},{"why":"The text-conditioned histopathology latent diffusion baseline used for SoTA comparison at 10× and 5× patch FID.","marker":"[49]"}],"fun_headline_variants":["Scale as a conditioning key: one model for all histology zooms","ZoomLDM draws coherent 4K pathology images in one pass","Multi-scale diffusion: state-of-the-art FID at eight zoom levels","Super-resolution without extra training from a multi-scale diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 4096×4096 joint-sampling algorithm rests on the assumption that the VQ-f4 autoencoder's decoder Jacobian is approximately orthogonal, so the guidance direction can be computed by a cheap finite-difference pass through the encoder instead of backpropagation through the decoder; the paper states this justification in Supplementary Section 8.2 but gives no quantitative check, and if the Jacobian is far from orthogonal the guidance steps are not provably following the constraint gradient.","fun_headline_variants_meta":{"raw":{"variants":["Scale as a conditioning key: one model for all histology zooms","ZoomLDM draws coherent 4K pathology images in one pass","Multi-scale diffusion: state-of-the-art FID at eight zoom levels","Super-resolution without extra training from a multi-scale diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001295,"raw_usage":{"total_tokens":5375,"prompt_tokens":1125,"completion_tokens":4250,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":741,"completion_tokens_details":{"reasoning_tokens":4176}},"tokens_in":741,"tokens_out":4250,"duration_ms":27153,"temperature":1.0,"reasoning_tokens":4176,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:40:57.121399+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the empirical singular-value distribution of the VQ-f4 decoder Jacobian, $J = \\partial \\mathrm{Dec}(\\hat{z}_0)/\\partial \\hat{z}_0$, at real 20× pathology latents (via finite-difference products $J^\\top J$ or a few Hutchinson-style estimates); if the singular values deviate substantially from 1, the identity $\\left(\\partial \\mathrm{Dec}/\\partial \\hat{z}_0\\right)^\\top e \\approx \\left(\\partial \\hat{z}_0/\\partial \\mathrm{Dec}\\right) e$ behind Eq. 12 fails. A behavioral check that is feasible today: run 1024×1024 joint sampling with the paper's Algorithm 2 and with the exact backpropagation-based algorithm of [16], and compare CLIP and crop FID — statistical equivalence validates the approximation, while a gap shows the claimed coherence is not attributable to the stated mechanism.","supporting_citations":[{"cited_title":"Fast constrained sampling in pre-trained diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the fast constrained-sampling algorithm for diffusion models under linear constraints that joint multi-scale sampling modifies."},{"cited_title":"Learned representation-guided diffusion models for large-image generation","cited_arxiv_id":null,"evidence_quote":"The learned-representation-guided patch diffusion baseline ZoomLDM extends, and the main comparison for patch FID and large-image generation."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the latent diffusion framework, the VQ-f4 autoencoder, the U-Net architecture, and the training code and checkpoints ZoomLDM builds on."},{"cited_title":"∞-brush: Controllable large image synthesis with diffusion models in infinite dimensions, 2024","cited_arxiv_id":null,"evidence_quote":"The infinite-dimensional diffusion baseline (∞-Brush) compared on 1024×1024 and 4096×4096 large-image generation."},{"cited_title":"Diffusion- based generation of histopathological whole slide images at a gigapixel scale","cited_arxiv_id":null,"evidence_quote":"The prior multi-scale pathology diffusion model that lacked a conditioning mechanism, the gap ZoomLDM's conditioning design targets."},{"cited_title":"Pathldm: Text conditioned latent diffusion model for histopathology","cited_arxiv_id":null,"evidence_quote":"The text-conditioned histopathology latent diffusion baseline used for SoTA comparison at 10× and 5× patch FID."}],"review_version":1}