Image-path reopening audit

Date: 2026-08-27

Outcome

The smallest defensible next experiment is a split-corrected Z-Image normal-specialist LoRA. Z-Image is the only currently eligible challenger that already has matched project evidence, a public ungated checkpoint, an explicit publisher licence trail, and an official maintained LoRA implementation. No GPU is authorised by this audit alone; the exact training runner and immutable remote input gate must pass first.

Candidate audit

Generic natural-image super-resolution is not selected. It can sharpen hallucinated nuclei, vacuoles, or oversized organisms without restoring correct gastric topology. The pathology-specific SuperDiff implementation is not released, and no other audited released SR checkpoint supplies both a clear submission-compatible chain and evidence for gastric H&E at this scale.

Frozen challenger

• Question: Can Z-Image's already stronger tissue prior learn corrected normal gastric topology from the exact 32-field slide-disjoint normal corpus where Qwen-Image-2512 failed?
• Model: Tongyi-MAI/Z-Image at 04cc4abb7c5069926f75c9bfde9ef43d49423021.
• Trainer: official Hugging Face Diffusers Z-Image LoRA path pinned before deployment; no third-party community trainer.
• Data: qwen2512-normal-specialist-slide-disjoint-v2, 32 exact 1024×1024 RGB images from nine non-held-out slides at 0.25 µm/pixel. The name is historical; the image/caption/provenance bytes are model-independent.
• Training: BF16 LoRA rank 32, batch one, four epochs, one dataset repeat, learning rate 1e-4, cached latents, gradient checkpointing, 8-bit Adam, fixed training seed, one whole checkpoint per epoch.
• Validation: same fixed normal prompt, negative prompt, 1024×1024 geometry, 50 steps, CFG 4.0, and seed 41001 at every whole checkpoint. Compare against the immutable M1 Z-Image zero-shot normal cell, failed corrected Qwen normal checkpoints, and untouched M1 real normal reference. Never mix epochs or cells.
• Automatic gates: exact revision/config/corpus hashes, four checkpoints and validation cells, 1024×1024 RGB, no NaN/empty outputs, privacy thresholds of DINO cosine below 0.995 and dHash distance above 4 against both corrected corpora, and Sydney-QA feature extraction with status remaining CALIBRATION_REQUIRED.
• Engineering gate: at least one whole checkpoint must improve coherent normal gastric gland/foveolar architecture over the Z-Image zero-shot and Qwen controls without goblet-like/metaplastic clear-cell excess, implausible repetition, pasted structures, or loss of tissue coverage. Sharpness is descriptive, never sufficient.
• Stop rules: first run is normal-only; do not run IM, gastritis, mixed, canvas generation, or candidate pooling unless a whole normal checkpoint passes. Runner EXIT shutdown, explicit signal failure markers, guest shutdown, GCE STOP, and cumulative S$700 control remain mandatory.

Evidence boundary

This audit establishes an engineering choice, not clinical validity or final legal advice. Sydney-QA, DINO, sharpness, and visual review cannot replace GI-pathologist sign-off. The Z-Image model repository uses Apache-2.0 metadata and its publisher code repository includes an Apache-2.0 licence; the exact frozen submission still requires a final rights review across model, trainer, dataset, and output manifests.

Reproducible implementation gate

The official trainer is wrapped by scripts/run_zimage_lora_worker.sh; fixed-checkpoint inference is isolated in scripts/run_zimage_validation.py; scripts/verify_zimage_challenger.py binds the exact model, trainer, corpus, split, training, and validation configuration; and gcp-zimage provides the short bounded control surface. The worker archives the exact model card, publisher licence, Diffusers licence, upstream commits, dependency lock, config, corpus gate, GPU identity, checkpoint hashes, validation manifests, terminal markers, and complete artifact hashes before shutting the guest down.

Initial local CPU preflight passed 32 images, nine non-held-out source slides, all exact hashes, and the frozen training/validation contract. Its config SHA-256 was b758f9dc098ce430f3d5a44a325f65af8b535a582af0e6bc498deb9133abeb74; this is retained as the launch-time evidence for attempt 1.

Attempt 1 (20260826T231203Z-zimage-normal) failed before model download or training because the pinned Diffusers commit contains the official Z-Image trainer but not the requirements_z_image.txt file referenced by its README. The recovered run has exit code 1 and RUN_FAILED; no checkpoint or validation image exists. The correction follows the exact trainer's embedded PEP 723 dependency declaration, excludes unused Prodigy, preserves the CUDA image's Torch/torchvision pair, and pins the direct requirement envelope in requirements/zimage_trainer_d57cecde.txt at SHA-256 336da007b4dcb72d542b793e2888c32a97196c7e05df64ff2a496ec5c85d51d1. The corrected immutable config SHA-256 is 6a3c66826c7d8cb937a6c1d0c66edf468d195cc16a57aa08ad9e0a8eadc43cf1. Local preflight passes again. No scientific variable changed.

Attempt 2 (20260827T001927Z-zimage-normal) also failed before model download or training. The corrected trainer envelope installed, but Diffusers import reached Transformers' audio utilities and loaded the image's ABI-incompatible torchaudio 2.11.0+cu129 against torch 2.9.1+cu129; _torchaudio.abi3.so failed on undefined symbol torch_library_impl. The recovered run has exit code 1 and RUN_FAILED; no checkpoint or validation image exists. Its worker.log SHA-256 is 7ddf12c2116bf07f1448c31d637f14e5def61758ec1295f90e5713e4d8b6092b.

The second minimal correction pins the exact compatible CUDA 12.9 trio used by the successful project runners: torch 2.9.1+cu129, torchvision 0.24.1+cu129, and torchaudio 2.9.1+cu129. The lock at requirements/zimage_torch_runtime_cu129.txt has SHA-256 e5359bf567d77ad8c502696cf8e3caf6605471d77ee398e913024ec9cf2bd082; the worker installs it from the official PyTorch wheel index and asserts every version before any model download. Corrected config SHA-256 is c443d7a646315c428ed3b000bd0c3922ac243eedda6cf97c0172dea89eab272f. The model, data, training schedule, prompts, seed, validation, and gates remain frozen.

Attempt 3 (20260827T010003Z-zimage-normal) successfully passed the corrected runtime/import/input gates and downloaded the exact model revision, but failed before training because Accelerate 1.14.0 rejects the literal --report_to none as an unsupported tracker. Its recovered exit code is 1 with RUN_FAILED; there is no checkpoint or validation image. worker.log SHA-256 is 2b654ab5c18fa528c55f74419003ed5a5d31df218fcf2a8b6eb072cb3f024960. The smallest correction uses the already installed, local-only supported tensorboard tracker and freezes report_to in the immutable config. Corrected config SHA-256 is ce6319bc8fc2e21dba72e63401bd48c21d6c892c648755e19871e62ea4a581d3. No scientific variable changed.

Attempt 4 (20260827T013450Z-zimage-normal) is the first valid scientific run: exit code 0, four whole LoRAs, four fixed validation cells, and all local integrity/privacy gates pass. Full-resolution review rejects every checkpoint. Step 32 mostly preserves the zero-shot composition, while steps 64-128 increase elongated-lumen repetition, stain drift, and simplified cellular topology rather than converging on the untouched local normal gastric reference. No category or pool advances. Full evidence is in reports/ZIMAGE_NORMAL_CHALLENGER_V1_REPORT.md.

The only authorised follow-up is the inference-only exact-caption alignment in config/zimage_normal_prompt_alignment_v1.json: same model, four checkpoint hashes, negative prompt, seed, dimensions, sampler settings, and controls, with the positive prompt changed to the byte-identical caption used by every training record. If that single aligned prompt does not rescue one whole checkpoint, close the Z-Image LoRA path.

That exact-caption run completed and failed. All four cells generated pseudo-label text containing the scale/caption language, violated the no-text mechanical gate, and showed stacked or mirrored papillary tissue with extreme repetition rather than local normal gastric topology. The Z-Image LoRA path is closed under its preregistered stop rule. Full evidence is in reports/ZIMAGE_NORMAL_PROMPT_ALIGNMENT_REPORT.md.

The remaining measured eligible challenger is FLUX.2 Klein Base 4B at the immutable M1 model revision a3b4f4849157f664bdbc776fd7453c2783562f4d. Its zero-shot cells were weaker than Z-Image, so it is not presumed superior. It advances only to an implementation/licence audit because the 4B base is Apache-2.0, Black Forest Labs explicitly positions it for fine-tuning, and official Diffusers now includes train_dreambooth_lora_flux2_klein.py. A separate frozen normal-only preregistration is required before GPU start.

The audit is now complete at the CPU/documentation level. The exact model card and license at the M1 revision are Apache-2.0. Publisher repository commit 50fe5162777813d869182b139e83b10743caef15 identifies the 4B Base as the limited-hardware fine-tuning model and carries an Apache-2.0 repository license. Official Diffusers commit ec94eec6cabd536c44c77647038303e96ab9355b contains the dedicated Klein DreamBooth LoRA trainer, documents local Qwen text encoding, CPU offload, latent caching, gradient checkpointing, 8-bit Adam, and NF4 quantization, and is itself Apache-2.0. Exact upstream artifact hashes are frozen in manifests/flux2_klein_challenger_upstream.csv.

The official FP8 example is not eligible on the A100 40 GB because the documentation requires compute capability 8.9 or later; A100 is 8.0. The same official trainer explicitly supports bitsandbytes NF4 as the older-card alternative. config/flux2_klein_normal_challenger_v1.json therefore freezes one normal-only 1024px NF4 rank-32 run using the corrected 32-field corpus, the exact training caption, four whole epoch checkpoints, and the same seed and full-resolution/held-out/privacy gates. This is only a preregistration. GPU start remains prohibited until the exact runner, validation code, dependency import test, local gate, remote immutable gate, and three independent shutdown layers pass.

The bounded FLUX.2 Klein run then completed successfully and produced four intact whole LoRAs and four fixed validation cells. Integrity, rights-evidence capture, split, dimensions, runtime, and both privacy screens pass. Full-resolution review rejects all four checkpoints: each preserves or rearranges the zero-shot model's repetitive micro-rosette/oval shortcut and lacks credible normal gastric nuclei, epithelium, lamina propria, and gland diversity. The preregistered stop rule closes FLUX.2 Klein category, canvas, and pool expansion. Full evidence is in reports/FLUX2_KLEIN_NORMAL_CHALLENGER_V1_REPORT.md.

The corrected Qwen, Z-Image, and FLUX.2 general-image LoRA routes have therefore all failed the same frozen normal-only gate. Further GPU work is prohibited until a new CPU/source audit identifies a released pathology-native generator or pathology-specific super-resolution system with an explicit submission-compatible model, code, and component licence chain. Generic super-resolution is not an eligible fallback because it can sharpen incorrect topology without restoring it.

That follow-on audit is now complete. No released route passes rights, exact-checkpoint, gastric whole-tissue, privacy-safe conditioning, and bounded A100 gates together, so no new GPU experiment is authorised. The exact candidate matrix and source ledger are in reports/PATHOLOGY_NATIVE_ROUTE_AUDIT.md.

Primary sources:

• https://huggingface.co/Tongyi-MAI/Z-Image
• https://github.com/Tongyi-MAI/Z-Image
• https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_z_image.md
• https://huggingface.co/didsr/HistoGen
• https://github.com/DIDSR/HistoGen
• https://github.com/xuanxu92/ctrPath
• https://github.com/bhosalems/PathDiff

Download public Markdown export