H04: modest local improvement, no convincing layout independence
7 September 2026. One-pair, one-seed engineering feasibility test. Completed 300 updates and three output arms; no clinical, generalization or submission claim.
Result
The human-mask adapter output has somewhat more consistently bounded pale pockets and coherent local walls than the disabled-adapter baseline. But the baseline already contains recognizable gland-like detail. The rotated-mask output changes local pixels/pockets while substantially retaining the original reference's longitudinal layout and recognizable field arrangement. This is not a convincing mask-driven generation success. Different hashes are insufficient: moved-layout adherence was the scientific test.
Independent blinded Astra review ranked B > A > C with a modest preference, calling the three roughly similar overall and not a step change. Unblinding: B = human, A = disabled, C = rotated. B had stronger continuity and contained pockets; A had some enlarged or merged pale areas; C had more crowding and dark clumps. Root separately inspected named outputs, reference and controls; root was not blinded. Neither review establishes disease truth.
Review images
- Human mask
- Disabled adapter: unchanged frozen baseline
- Rotated mask
- Real appearance/training reference
- Prior H01 scaffold result
Do not attribute the difference between H01 and H04 wholly to learning: H01 started from a flat scaffold, whereas H04 starts from noise, uses a different seed and compares against its own disabled baseline. The within-H04 comparison is the relevant matched test.
Verified evidence
- 72 portable file hashes passed; archive SHA256:
4e1221bdf7cd605b0265007fdfbb830698d42da4639be0a96ef6fac170f58c40. - 300 recorded updates, 912,256 adapter parameters; exact frozen-base hashes before and after match. Losses, gradients and output tensors are finite. Variable-timestep training loss is not a matched quality metric.
- The 20-step pre-fit untouched/disabled/zero-head latent and PNG parity checks passed. The post-fit disabled output also exactly matches the pre-fit baseline latent and PNG.
- All three actual initial-noise tensors match; cached training latent, UNI and mask hashes match the pre-fit manifest. The output latents differ, proving a numerical effect only.
- No real-RGB latent initialization, LoRA, checkpoint selection or seed selection. Original annotations remain unchanged.
Independent local audit: H04_A1_AUDIT.json. The limited seven-reference resemblance screen is complete: full screening report.
| Output | Aligned grayscale correlation with source | Median nearest-patch similarity |
|---|---|---|
| Human mask | 0.638 | 0.693 |
| Disabled adapter | 0.304 | 0.388 |
| Rotated mask | 0.199 | 0.356 |
Every arm's strongest match was reference 01, the training and appearance source. The exact and mildly recolored source controls scored near 1; a 32-pixel shifted source had median patch similarity 0.344. Thus the preferred human output also retains the strongest measured source resemblance. These are screening correlations, not fractions of copied pixels or a privacy threshold. Even a shifted copy can score low. The search covered all seven references at two fixed patch scales while preserving aspect ratios; its reference pool is incomplete and correlated. It establishes neither novelty nor absence of memorization.
Resource closure
Our VM 8097050405901662678 was verified stopped at 15:41:48.945 UTC, after booting at 15:29:27.143 UTC: 741.802 powered seconds. Estimated compute cost is $0.756923 using the historical planning rate of $3.673385/hour—not an invoice, and excluding storage, tax and agents. Original FLUX resources were untouched. Training, generation, review and the bounded resemblance screen have finished.
Final cleanup update: after verifying both experiments' archives and the trained adapter locally, the temporary VM and its separate 350 GB copied disk were deleted successfully. That copy is not directly recoverable, but the original assets remain untouched and the experiment outputs/checkpoint are preserved locally. No GPU or copied-disk storage resource remains in this branch. H01 plus H04 estimated compute totals $3.20, excluding storage, tax and agent usage.
Decision for review
Do not sweep this one-pair recipe or promote the visual preference as a generation breakthrough. One source with one fixed mask/reference combination cannot establish that the model learned independently controllable geometry. A useful next design would deliberately vary geometry while controlling appearance, then test on different verified source fields. That is a new experiment, not a reason to silently tune this result. H02 source preservation remains a different augmentation question; H03 is not admitted. This annotation branch is paused for user review, not declared a completed SGH solution. The other FLUX thread remains independent.