Grail Computer · research record
Dated experiment record. Statements describe this run, not current submission readiness. Historical H. pylori category names do not establish infection; copy screens are bounded tests. Claims that a source does not exist mean none was identified in that recorded search, not proof of absence. Licensing interpretations in the notes remain unconfirmed. Local access details have been omitted. Current limitations and remaining work.

Pathology-native and pathology-SR route audit

Date: 2026-08-27

Audience: SGH synthetic-histopathology engineering and scientific review

Decision: BLOCKED — no GPU experiment authorised

Direct answer

No currently released route clears all five required conditions together:

  1. an explicit submission-compatible licence for code, checkpoint, and required components;
  2. an exact downloadable checkpoint or a bounded, reproducible training path;
  3. a task that can represent local gastric gland/foveolar topology and the four required categories;
  4. conditioning that does not require private real exemplars in the generated output path; and
  5. a feasible A100 40 GB execution path.

The general-image LoRA failures are therefore not followed by another speculative GPU run. HeteroTissueDiffuse is the closest released whole-tissue generator in task shape, and CorPiR is the cleanest newly released pathology-SR package, but neither solves the present normal-gastric topology problem. HeteroTissueDiffuse has only non-gastric release checkpoints and requires semantic maps plus visual tissue exemplars. CorPiR reconstructs detail from a low-resolution input; it cannot repair a structurally wrong synthetic field. All chargeable GPU compute remains stopped.

Frozen eligibility test

A route advances only if its primary sources establish the exact model and code revisions, checkpoint availability, component licences, gastric/whole-tissue task fit, conditioning contract, physical-scale compatibility, A100 feasibility, and a privacy-safe input path. A code licence is not treated as checkpoint or dependency permission. Generic or pathology-specific super-resolution is not accepted merely for sharper pixels: it must have a credible mechanism and reference evidence for restoring correct topology rather than amplifying a failed generator's structure.

Candidate audit

Route Exact source state inspected Rights/component result Task, privacy, and hardware result Decision
HeteroTissueDiffuse Code f65c155e5a079ceebbf43cd9fabb1f9cf833c08c; model revision 3f00a5b63060438d1d0caa81fd279d0313f71663; released Camelyon16, PANDA, and TCGA checkpoints Model metadata declares CreativeML Open RAIL-M. The code repository links that licence but contains no local licence file and is built on CompVis Stable Diffusion, so the complete component record would still need freezing before submission use. Release path is semantic_synthesis256; Camelyon is explicitly 256 px, and no gastric normal/IM/H. pylori checkpoint is released. Inference requires a semantic map plus raw tissue crop exemplars. Real exemplars would create a direct privacy/memorisation surface; synthetic exemplars would not supply the missing gastric prior. README documents an A100-compatible PyTorch 2.0.1/CUDA 11 environment. Blocked. Relevant architecture, wrong released domains and unsafe/unresolved exemplar strategy. Do not authorise GPU.
Graph-Conditioned Diffusion Code 7b6430d24dbd744eba73d10b2cee509191ac1b5b; no released weights found README claims MIT but the referenced LICENSE file is absent. No checkpoint download is provided. Requires object masks, graph features, a BYOL model, a trained graph transformer, and three cascaded diffusion checkpoints. The repository omits preprocessing/multiresolution dataset construction. Its demonstrated structural vocabulary is renal-like tubules/glomeruli, not gastric glands. It was tested on A100, but reproducing it would be a new training programme rather than a bounded challenger. Blocked. Missing licence artifact, checkpoints, preprocessing, and gastric conditioning.
CHIS Code 2fb6df9f5804dca8a5e9d8aa5a7b26fe42582b2e, MIT wrapper Runtime loads PixCell-256 (CC BY-NC-ND 4.0), UNI2-h (CC BY-NC-ND 4.0), gated SD3.5 VAE terms, and a PixCell custom pipeline. The permissive wrapper does not override those restrictions. Training-free generation needs a target mask and a reference image. A real SGH reference in the output path adds privacy/copying risk. Consumer-GPU inference is feasible, but rights and conditioning fail first. Blocked. Incompatible component rights and reference-image privacy risk.
His-MMDM Code/model card 26a1f9f12a8a3d1799496fb7dacebd722d92f3e3; checksum-validated AWS checkpoint bundle Repository and model card state MIT. Released task translates a source histology image among TCGA primary tumour types and optionally uses source genomics/transcriptomics. It does not expose normal gastric, IM, gastritis, or H. pylori domains. Model card states 1–2 V100 GPUs for inference and 24 V100s for from-scratch training. A real source image is also required. Blocked. Downloadable and permissive, but wrong domains, source-image dependency, and unbounded retraining requirement.
HistoGen Code 05e83123ce256e12d9293cb5c6656928d150af08 under CC0; model revision 37eba158b2b51ac3cbf2deaa4ec844dcd389500f marked MIT and manually gated Rights are permissive for the released assets. Generates nuclei images from instance/semantic masks. It does not generate whole gastric mucosa, gland topology, lamina propria, surface epithelium, or category-specific fields. Blocked. Correct rights, wrong task.
CtrPath / SuperDiff Code e7ee025b167f49ebd7740aeb6edc0745e15ad443; CtrPath checkpoint is linked from Google Drive README says MIT while the repository licence file is Apache-2.0. The checkpoint depends on SD1.5/CtrlLoRA components whose exact frozen chain is not supplied in the release record. CtrPath is a 512 px nuclei-mask-plus-text generator. The pathology-specific SuperDiff stage remains explicitly “to be released”. No released gastric checkpoint or final-resolution SR implementation exists. Blocked. Licence presentation/component chain unresolved; required SR stage absent.
PathDiff Code e883b0e7080b4e7327d011cb03d92c4e827cf2b2; Google Drive checkpoint linked No repository licence file. README states research and educational purposes only. Text/mask generator trained on PathCap/CoNIC-style sources. Authors warn that randomly paired conditions may be pathologically implausible. No gastric category checkpoint or complete submission rights exist. Blocked. Rights and category validity fail.
PixCell Code ea163d680f4ed77a6191e7f421d5adc043cf4af8; PixCell-256 f3800a5d9d726ebc8e2badce4d84722f9478202f; PixCell-1024 bc4ac7adc319d906f939f0c6c24d024e5d8c450e Code is CC BY-NC 4.0. Both released checkpoints are CC BY-NC-ND 4.0 and the model card expressly limits them to non-commercial research; UNI2-h dependency is also CC BY-NC-ND 4.0. Strong pathology-native generator and technically relevant resolution, but direct company/challenge use and derivative fine-tuning are not authorised. Blocked pending written rights. Do not use weights or outputs for submission.
CytoSyn Model revision 2a6e3c312707b9b4ad32686ac36595921a2ea124, gated, licence metadata other Existing frozen audit records non-commercial academic terms and prohibition on commercial use/monetisation of derivatives. Current access remains gated and supplies no compatible replacement licence. Pathology-native capability does not cure the rights failure. Blocked pending written rights.
CorPiR (newly released SR lead) Code d89b68e34aa6f6a91ebefc91576075c5722b4a9e; GitHub release v1.0.0 (2026-08-08) includes exact LIIF encoder plus x2/x4/x8 generator checkpoints Apache-2.0 repository; README records Apache-derived SR3/LIIF components. This is the cleanest complete release chain found in this audit. A two-stage pathology SR reconstructor evaluated on 128 px pan-cancer patches. It derives LR input by downsampling an HR target and reports 13 s per 128 px image on RTX 4090. Applied to the failed synthetic normal fields, it would reconstruct/sharpen their existing micro-rosette/oval topology, not supply missing gastric gland structure. Do not run now. Rights/checkpoints are promising, but the scientific task cannot correct the current causal failure. Revisit only after a topology-credible low-resolution gastric generator exists.
S3-Diff (newly released SR lead) Code d220beebc76230db16e147e01507c2df19c94f0b, MIT Requires SD1.5, DINOv3, SAM and CLAM under their own terms. Repository explicitly includes no pretrained S3-Diff checkpoint. x4 pathology SR trained on paired HR/LR patches and TCGA-LUAD/KIRC/LIHC. Using it would require new model training and would still condition on an already wrong LR topology. Blocked. No released task checkpoint; component chain and causal fit fail.
PathoGen (newly released inpainting lead) Code 388b1ff07fd8b730f41899eb8e9f52179524165a; HF revision 259b91ff8855bc59f6487c27aa6b0665375ab04e exposes attention.pt HF metadata says MIT, while the code repository has no actual licence file despite its README statement. Base inpainting UNet is a separate component. 512 px lesion inpainting requires an original WSI crop, mask, and source pattern crop. Published domains are kidney, skin, breast, and prostate, not gastric normal/IM/H. pylori. Real conditioning would also create direct privacy exposure. Blocked. Incomplete licence artifact, wrong domains, and real-image conditioning.
PathLDM / MuPD and other search leads PathLDM code a717d0388b4bac624f4d995ab91e7b1752b7c7c5; MuPD paper found but no released official code/checkpoint found in the bounded search PathLDM repository has no licence file; MuPD release rights cannot be established without artifacts. PathLDM is 256 px TCGA-BRCA text generation. MuPD is scientifically interesting but presently paper-only for this audit. Other discovered systems were nuclei-pair synthesis, virtual staining, or non-histology reconstruction. Blocked. No complete rights/task/checkpoint chain.

Gap reconciliation

Required claim Best evidence found Confidence Remaining blocker
A released model can generate coherent whole gastric tissue HeteroTissueDiffuse and PixCell are the strongest released whole-tissue architectures inspected High that neither is currently usable here HeteroTissueDiffuse has no gastric checkpoint and needs exemplars; PixCell rights are incompatible
A released pathology SR model can repair the failed topology CorPiR has a complete, permissive checkpoint release; S3-Diff has code only High that current evidence does not support repair CorPiR reconstructs input topology; S3-Diff requires training and also starts from LR topology
Rights are sufficient for a company challenge submission HistoGen and CorPiR have the cleanest permissive releases High Both solve the wrong task; usable generators remain restricted or incomplete
A bounded A100 experiment is ready HeteroTissueDiffuse documents A100 inference; several inference paths fit lesser GPUs High that hardware is not the decisive blocker Scientific/task/rights gates fail before hardware
Another broad search is likely to change the decision today Bounded search covered requested systems and current high-signal releases Moderate-high Recheck only when a checkpoint, licence, gastric model, or written permission changes

Decision and next actions

  1. Keep all GPUs stopped. No normal-only preregistration is created because no route passes the exact audit.
  2. Preserve CorPiR as a conditional future SR component, not a correction for failed topology.
  3. Seek written challenge/commercial permission for PixCell and CytoSyn if the organisers and licensors can provide it. Permission must cover checkpoint use, fine-tuning if needed, generated-output submission, and finalist transfer/use requirements.
  4. Ask the HeteroTissueDiffuse authors whether a gastric checkpoint, gastric semantic vocabulary, and a privacy-safe non-real-exemplar path exist under a complete frozen licence chain. Do not infer those capabilities from the current release.
  5. Prioritise GI-pathologist calibration of the patient/slide-disjoint real reference envelope for Sydney-QA. This does not create a generator, but it converts the current uncalibrated engineering features into a safer rejection gate for any future route.
  6. Re-run this audit only on a material source change: a new exact checkpoint, compatible licence, released SuperDiff/S3-Diff weights, gastric-domain release, or written permission. Do not spend GPU on unlicensed checkpoints, generic SR, or paper-only methods.

Evidence boundary

This is an engineering and source-availability decision, not legal advice or clinical validation. Repository/model-card terms can be incomplete or change; final submission requires a frozen rights review. No DINO, sharpness, Sydney-QA, or source audit establishes clinical correctness. A generated field cannot establish a patient diagnosis or OLGA/OLGIM stage, and GI-pathologist sign-off remains mandatory.

Primary-source ledger

All sources were accessed 2026-08-27. Exact repository and model revisions are recorded in the candidate table.

Search stopped when every required route had a primary-source disposition, the highest-impact rights/task claims were independently checked against repository/model metadata, and the additional high-signal releases did not produce a route that could change the no-GPU decision.

Download public Markdown export