BigCell: Generating Gigapixel Whole-Slide Images

1 Introduction

• Whole-slide synthesis matters because tissue structure and staining are coupled across distant regions, while pathology reports and diagnostic labels describe slides rather than isolated tiles.Existing histopathology generators typically produce tiles of 256 to 1024 pixels, whereas pathologists rely on spatial structures spanning thousands of pixels.Source: §1 · pp. 1–2
• BigCell progressively trains a tile generator from 256 to 2048 pixels and uses dense foundation-model embeddings as a blueprint for sliding-window generation up to 32,768×32,768 pixels.Training uses over 120,000 publicly available WSIs spanning multiple organ types.Source: §1 · pp. 2–3
• Over 3 times speedup with minimal quality degradation is achieved by using classifier-free guidance magnitude to prioritize important windows during sliding-window generation.The strategy is training-free and uses one interpretable hyperparameter controlling the fraction of important windows.Source: §1 · p. 3
• BigCell achieves superior macro-scale fidelity to prior tile generators, with higher foundation-model similarity and improved bottom-5th-percentile patch-level similarity.Prior tile generators retain high patch-level quality but show degraded structural coherence across broader regions when generating 8192×8192 images.Source: §1 · p. 3
• A separate conditioning generation model produces dense foundation-model embedding grids from pathology reports, enabling text-conditioned histopathology images at 8192×8192 pixels.The full grid is generated jointly to capture long-range dependencies, and alignment with the input report approaches that of real paired data.Source: §1 · p. 3
• Up to 18% relative improvement in performance over real-data-only baselines resulted from BigCell synthetic-data augmentation on TCGA-UT-8K and PANDA.The reported application uses diverse, class-consistent synthetic variations in few-shot classification settings.Source: §1 · p. 3

2 Related Work

• Whole-slide images require cellular morphology and staining to remain consistent across distant regions, challenging methods that assume local appearance dominates fidelity.Prior approaches synthesize overlapping windows or fuse resolution branches.Source: §2 · pp. 3–4
• Computational pathology is moving from patch-level embeddings aggregated by multiple-instance learning toward slide-level and context-aware representations.These approaches model dependencies beyond a single patch, which prior work associates with stronger downstream representations.Source: §2 · p. 4
• Earlier whole-slide generators use thumbnail cascades or segmentation masks, but their designs leave global properties such as staining or long-range tissue organization uncaptured.Harb et al. refine local blurred crops after thumbnail sampling; DiffInfinite couples distant patches only through a tissue-class mask and has not been demonstrated at gigapixel scale.Source: §2 · p. 4
• Representation-conditioned methods include LRDM sliding-window inference, ZoomLDM magnification-aware training, and PixCell training up to 1024 native pixels and inference-time stitching to 4096.BigCell instead denoises the full latent canvas at 20× in one stage using dense foundation-model embeddings extracted from a reference slide or jointly sampled by CGM.Source: §2 · pp. 4–5

3.1 Tile generator

• A flow-matching transformer generates 2048×2048 tiles from dense UNI-2 foundation-model embeddings extracted from 256×256 patches.The resulting 8×8×1536 conditioning grid enters the transformer through cross-attention; the model operates on VAE latents.Source: §3.1 · p. 5

3.2 Sliding-window generation

• For larger images, BigCell applies the tile generator to overlapping windows and aggregates their velocities into a global field for deterministic Euler denoising.The training-free procedure uses a global conditioning grid and VAE decoding after the final step; higher native tile resolution improves gigapixel fidelity, with diminishing returns after 1024.Source: §3.2 · p. 5

3.3 Efficient generation

• Adaptive scheduling ranks windows by the difference between conditional and unconditional velocities, recomputing top-r windows and reusing cached velocities for skipped windows.It is training-free, controlled by one hyperparameter, and reduces compute by approximately 1/r; compute allocation across tokens by Wu et al. does not share information across tiles.Source: §3.3 · pp. 5–6

3.4 Conditioning generation model (CGM)

• The conditioning generation model (CGM) learns to generate dense UNI embedding grids from pathology reports encoded by TITAN.It models p(y|c) in UNI representation space and samples grids for 8192×8192 regions, enabling text-conditioned image generation through BigCell large.Source: §3.4 · p. 6

3.5 Synthetic data augmentation

• For classification augmentation, CGM perturbs and denoises a real sample’s embedding grid under a class-specific text prompt before BigCell large generates a synthetic image.The procedure preserves the original image’s global structure while adding prompt-driven class-consistent variation; augmentation uses t′ = 0.5.Source: §3.5 · p. 6

4.1 Data

• Training uses 120,703 WSIs from six public sources, while CGM training uses about 30,000 patient-level reports and 500,000 paired grid-text samples.The 2048×2048 tiles are split at slide level; text-conditioned evaluations hold out 150 reports and corresponding slides, and CGM operates at 8k because larger regions are too scarce for training.Source: §4.1 · pp. 6–7

4.2 Metrics

• Evaluation measures tile-level similarity and Fréchet distance, slide-level foundation-model similarity, report agreement, and report-image text alignment and retrieval.Tile metrics include average and bottom-5th-percentile similarity; slide metrics use TITAN and MOOZY, report agreement uses MedGemma-4B, and text alignment uses TITAN and PRISM.Source: §4.2 · p. 7

4.3 Baselines and evaluation procedure

• The study compares tile generators and large-image methods on 100 held-out 8k ROIs, evaluates adaptive scheduling against matched-NFE alternatives, and tests few-shot augmentation on PANDA and TCGA-UT-8K.Pathologist evaluation uses three board-certified pathologists, each completing 100 trials with real and generated images; Table 1 reports the 8k comparison and significance convention.Source: §4.3 · pp. 7–8, Table 1 · p. 8

5.1 Comparison against prior methods

• BigCell tile achieves the highest slide-level similarity to reference images among the compared methods, indicating more faithful global tissue arrangement.LRDM and ZoomLDM have the best Fréchet distances but the lowest agreement with reference images; those distances use unpaired crops and do not measure conditioning agreement.Source: §5.1 · p. 8
• BigCell tile leads on patch-level cosine similarity, with V25th at 0.737 versus 0.632 for PixCell and 0.502 for CytoSyn.LRDM and ZoomLDM attain the best Fréchet distances while obtaining the lowest agreement with reference images; Fréchet distance is computed over unpaired crops.Source: §5.1 · p. 8
• BigCell agrees with reference reports on organ in 69% of cases and diagnosis in 46%, versus 63% and 42% for PixCell, respectively.The methods were evaluated on identical ROIs, so these rates are meaningful relative to one another rather than in absolute terms.Source: §5.1 · p. 8, Table 2 · p. 9

5.2 Efficient gigapixel generation

• At NFE ≈11000, adaptive scheduling achieves TITAN similarity of 0.951, matching 0.957 at full compute, and speeds 32k generation 3.1×.Runtime falls from 15.6 hours to 4.97 hours on one Intel Data Center GPU Max 1550; across compute budgets the schedule improves over fewer-step sampling.Source: §5.2 · pp. 8–9, Figure 3 · p. 9

5.3 Text-conditioned generation

• On 150 held-out reports, generated images approach real-pair text-image similarity: TITAN 0.487 vs 0.551 and PRISM 0.454 vs 0.499.Randomly paired images and reports score substantially lower, and TITAN recall@5 matches the real-pair baseline at 0.655.Source: §5.3 · p. 9, Table 3 · p. 9
• Pathologists match synthetic queries to source reports as accurately as real queries (0.57); their real-versus-synthetic detection accuracy is 55.3% (95% CI 50–61%).Inter-rater agreement was at chance (Cohen’s κ = 0.020); report matching is upperbounded because each patient report covers multiple distinct 8k regions.Source: §5.3 · pp. 9–10, Figure 4 · p. 10
• CGM-based augmentation improves few-shot classification across TCGA-UT-8K and PANDA at every shot count, with relative gains up to 7% and 18%, respectively.PANDA is held out from CGM training, providing an out-of-distribution evaluation.Source: §5.3 · p. 10, Table 4 · p. 10

• Text-conditioned generation is limited to 8192×8192 pixels; gigapixel text-to-image synthesis would require substantially larger paired datasets than currently public.Source: §6 · p. 10
• Despite adaptive scheduling, 32k generation takes 4.97 hours on a single GPU, which the authors consider impractical for high-throughput applications.Source: §6 · p. 10
• Generated histopathology images carry misuse risk and are intended for research only, not as substitutes for real tissue in clinical or diagnostic settings.Source: §6 · p. 10

7 Conclusion

• BigCell generates coherent whole-slide images at native gigapixel resolution, reaching 32,768×32,768 pixels with adaptive scheduling, and supports 8192×8192 text-to-image generation.The conclusion reports that generated samples outperform prior methods on tile- and slide-level fidelity and improve few-shot classification on two downstream datasets.Source: §7 · p. 10

A Motivation for whole-slide generation

• Whole-slide generation addresses the need to preserve biological dependencies, global diagnostic structure, and slide-level features that independently generated patches cannot represent.Pathologists relate local findings to surrounding tissue; reports describe features such as grade, margin status, and invasion that span large slide regions.Source: A Motivation for whole-slide generation · pp. 15–16

B Clinical Applications

• The authors describe text-generated 8k ROIs as potential controllable examples for algorithm development, refinement, benchmarking, and pathology education.Examples include varying pancreatic cancer invasion, grade, and rare morphologies for trainees to examine alongside common findings.Source: B Clinical Applications · p. 16

C Extending sliding window generation to flow-matching

• For flow matching, the sliding-window method estimates a global canvas velocity from local window velocities by least squares, then uses that velocity in the Euler sampler.With uniform window weights, the global velocity is approximated by averaging overlapping local velocities at each canvas location.Source: C Extending sliding window generation… · p. 16

D.1 Tile generator

• BigCell tile is an 830M-parameter SiT-XL/2 flow-matching transformer conditioned on UNI-2 features and operating in Flux-2 VAE latent space.The 2048×2048 image maps to a 256×256×32 latent; at 32k, sliding windows use corresponding 8×8×1536 conditioning slices, aggregate velocities, and decode once.Source: §D.1 · p. 17, Table 5 · p. 17
• Tile training progressively increases resolution from 256 to 2048 pixels, initializing each stage from the previous stage’s weights.The schedule trains for one epoch at each resolution.Source: §D.1 · p. 17, Table 6 · p. 17

D.2 CGM

• The 362M-parameter CGM is a flow-matching transformer in UNI-2 CLS-token space, conditioned on unpooled TITAN text embeddings.Source: §D.2 · p. 18, Table 7 · p. 18

D.3 Compute details

• Training used 8,000 Intel GPU-cluster node-hours, including exploration and failed runs.Source: §D.3 · p. 18

D.4 Data pipeline

• The data pipeline used TRIDENT for tissue segmentation and 20× patch extraction; BigCell tile training applied neither stain normalization nor data augmentation.UNI-2 grids, VAE features, and TITAN features were precomputed and stored.Source: §D.4 · p. 18

D.5 Sampling details

• Tile and CGM sampling use Euler ODE samplers with classifier-free guidance scale 2.0, at 20 and 50 steps, respectively.Tile-generator quality metrics improve with native resolution, with diminishing returns beyond 1024; 4096 training was not performed because its cost scales quadratically.Source: §D.5 · p. 18, Table 8 · p. 18, Figure 6 · p. 19

E.1 Effect of tile generator resolution

• Slide- and tile-level metrics improve monotonically as native training resolution increases, with most gains below 1024 pixels.The 1024-to-2048 step yields smaller gains.Source: §E.1 · p. 19
• Training at 2048 pixels costs 8× more per sample than at 1024, while 4096-pixel training is estimated at ∼96× the 1024 budget.A 4096-pixel iteration benchmark measures a further 12× increase over 2048.Source: §E.1 · p. 19
• Diminishing fidelity gains beyond 1024 pixels and the cost of 4096-pixel training lead the authors to use 2048 pixels as the final operating point.The paper describes full 4096-pixel training as impractical at its dataset scale.Source: §E.1 · p. 19

E.2 Resolution-matched comparisons

• The resolution-matched comparisons use the same 100 held-out ROIs, while the window-selection experiment varies which windows receive fresh versus cached velocities.The selection experiment sweeps the selected fraction to vary NFE.Source: §E.2 · p. 19, §E.3 · pp. 19–20, Table 9 · p. 20
• At matched 1024 native resolution, BigCell leads PixCell-1024 on every slide- and patch-level metric.Table 9 compares both methods on 100 held-out ROIs at 8192×8192 resolution.Source: §E.2 · p. 19, Table 9 · p. 20
• With backbone, data, sampler, and training budget matched, the summarizer-conditioned model underperforms both BigCell and PixCell-1024 on every metric.The authors attribute the result to fixed-size conditioning-grid pooling and say the proposed bottleneck warrants further investigation.Source: §E.2 · p. 19, Table 9 · p. 20

E.3 Importance-based window selection

• CFG importance-based selection beats random selection across the full NFE range, with larger tail-metric gaps at low compute and a closing gap near full compute.The comparison holds denoising steps and velocity caching constant; Figure 7 shows the quality comparison.Source: §E.3 · p. 20, Figure 7 · p. 20

F Report-level agreement with MedGemma

• MedGemma-4B decodes tissue of origin, malignancy, and diagnosis from each generated and reference image, and the evaluation measures agreement on those fields.The paper also reports agreement on all three fields together.Source: F Report-level agreement with MedGemma · p. 20
• At 32,768×32,768 resolution, BigCell generations agree with reference-slide reports far more often than chance on all three criteria.The evaluation uses 256 matched real/synthetic pairs at full compute; Table 10 reports the comparison.Source: F Report-level agreement with MedGemma · p. 20, Table 10 · p. 21

G.1 Report summarization for CGM training

• For CGM training, unstructured reports are rewritten as concise diagnostic summaries, with five paraphrases generated per report.The prompt requests sentences under 50 words, including tissue type and important positive test findings when present.Source: §G.1 · p. 21

G.2 Synthetic report generation for data augmentation

• For each of 32 TCGA-UT-8K cancer types and 6 PANDA ISUP grades, the authors generate four plausible pathology reports of no more than 25 words.Claude Opus 4.7 generates the class-conditional reports.Source: §G.2 · p. 21

H Broader impacts

• Because BigCell is trained on publicly available data, it reflects those datasets’ biases; its generated images are intended for research and education, not clinical or diagnostic substitution.The paper states that generated images should not substitute for real tissue in clinical or diagnostic settings.Source: H Broader impacts · p. 21

I Additional qualitative results

• BigCell tile preserves global tissue structure and fine cellular details; PixCell, CytoSyn, and MuPaD respectively show inconsistent structures, blurriness, and collapsed layout at 8192×8192.MuPaD also poorly preserves tissue morphology.Source: I Additional qualitative results · p. 23
• Adaptive scheduling gives detailed windows more denoising steps and background windows fewer model passes, while maintaining overall gigapixel image quality.Figure 10 visualizes the importance assigned to windows.Source: I Additional qualitative results · p. 23
• The appendix shows 32,768×32,768 BigCell generations of breast, lung, and mesothelial tissue, with progressive zoom-in crops.Additional synthetic samples are available as pyramidal TIFFs.Source: I Additional qualitative results · p. 23, Figure 11 · p. 24, Figure 12 · p. 25

2. Limitations

• The checklist states that BigCell uses public datasets without proprietary data and that code and pretrained models are released.It also says the information needed to reproduce the main experimental results is provided in Appendix Section D.Source: §2 · pp. 28–29

NeurIPS Paper Checklist

• The checklist reports statistical significance tests for data-augmentation experiments and comparisons with prior tile generators.Source: §2 · p. 29

10. Broader impacts

• The authors state that users of released models and datasets must follow usage guidelines and access restrictions.Source: §10 · p. 31