Final selection — Mini Photo Colorizer v3.0
Decision
Release the round6 palette checkpoint at step9000 with the new, verified image pipeline. All 3,994,676 learned parameters, including the encoder, are in model.safetensors. No discriminator, teacher, second model or weight ensemble is loaded in production.
The final two training runs tested conditional colour critics with strengths0.04 and0.12. Both completed 2,500 fine-tuning updates, with checkpoints saved and verified every500 updates. They produced stronger colours in some images but reintroduced local colour patches. They were rejected after visual review and matched reporting. Two earlier startup attempts failed on a typed optimiser argument before training; that argument was corrected. Every successful run is retained under experiments/final-20260928.
Matched 80-image reporting set
78 images had appreciable original colour. These are repeated reporting/development images, not an independent unseen benchmark.
| Checkpoint | Blotch proxy (lower) | Missed colour (lower) | Colour coverage |
|---|---|---|---|
| Selected round6 palette9000 | 0.779 | 0.169 | 0.474 |
| Gentle colour critic | 1.070 | 0.170 | 0.487 |
| Strong colour critic | 1.413 | 0.171 | 0.497 |
The critics increased the patchiness measure by about37% and81%. On reviewed images, this corresponded to red patches on faces, small coloured areas within otherwise consistent objects, and colour bleeding. More saturation did not make them better releases.
Round7's four objective/degradation variants were also rejected: they were visually almost unchanged and did not improve blotches.
What improves over the previously deployed app
The previous Space used the old v2 U-Net. The release uses pretrained semantic features and a shared palette. In the prior pinned300-image fresh COCO comparison (272 chromatic images), selected round6 lowered the patch-excess proxy from1.714 to0.747 and missed-colour fraction from0.223 to0.159. Those are approximately56% and29% reductions under that evaluation pipeline, not universal guarantees.
The production image pipeline additionally retains aspect ratio, uses original-resolution lightness to guide chroma upsampling, preserves EXIF orientation and alpha, checks the image-size limit, compresses out-of-gamut chroma, and reports input errors. See the actual pipeline comparison appended below for its distinct measurements.
Verification
QA.json records CPU inference on tiny, portrait, landscape and1024×768 images; exact alpha retention; worst observed lightness difference below0.22 Lab L after8-bit quantisation; black/white and saturation-zero behaviour; Gradio function and HTTP calls; and ONNX parity with maximum observed difference0.000912 Lab units. These are engineering tests, not evidence of historical colour accuracy.
The Space pins the model release commit and uses the same tested source. ZeroGPU GPU operations are isolated in a decorated function; full-resolution rendering runs on CPU. Separate live Space verification is recorded in LIVE_SPACE_CHECK.json when completed.
Evidence and provenance
- Final low-critic training
- Final high-critic training
- Round7 review
- Round6 selected checkpoint and evaluation
- Actual release pipeline comparisons
Limits
The release may choose muted, warm or incorrect colours and can still show residual bleeding. Real archival evaluation is limited, and two familiar historical examples are not representative of the entire domain. The moon remains an out-of-domain failure. There is no basis to claim that all further improvement below4M parameters is impossible. This is the selected, tested final release from these experiments; it is not a guarantee of perfect colourisation.
Actual production image pipeline comparison
{ "n_total": 80, "n_color": 78, "summary": { "old_space": { "ab_error": 16.144625015747852, "patch_excess": 1.8169478370020022, "color_coverage": 0.47298392271384215, "missed_color": 0.24084079671968467 }, "release": { "ab_error": 14.243927399317423, "patch_excess": 0.6932277647444071, "color_coverage": 0.479247068747496, "missed_color": 0.1663338435453805 } }, "median_pipeline_seconds_L4": 0.06851194500001156 }
This comparison preserves aspect ratio and original output resolution before measuring the resized results. Median pipeline time excludes ZeroGPU queue/allocation and network transfer.