Penultimate architecture and curriculum decision — 28 September 2026

Objective and decision scope

Select one reproducible course for the user's next, longer training run. Prioritize coherent plausible colors, fewer within-surface patches, and retained color over recovering an unknowable original hue. Strictly fewer than 4,000,000 deployed learned parameters, including encoder and any selector. This screen cannot establish a universal architecture ceiling. No experimental model replaces root production weights during this pass.

Evidence audit

The uploaded project log describes an older classification U-Net. Current production is round6 palette step9000 at base revision 1a9eb8af2754ad2329a24cfe50d388cb559441d0, 3,994,676 parameters. The authoritative current repo revision before this pass was 26d731862cc0ea531e7198fa823cb09db25bd243.

The semantic encoder/palette change improved substantially over the old U-Net. Round7's easier-of-original-or-teacher objective selected original targets only 5.8–9.5% of the time; it did not learn explicit alternative outputs. Critic variants (including 27M, smaller, and dense critics) failed visual release review. A region auxiliary and repeating the same decoder did not produce a clear gain. These results motivate separating color ambiguity from spatial assignment, not another saturation/critic sweep.

Primary research and implications

Zhang et al., Colorful Image Colorization: https://richzhang.github.io/colorization/ — classification/rebalancing addresses uncertainty; pixel-wise decoding still needs spatial coordination. Earlier log language that classification is universally required was an overstatement: later palette regression already outperformed that baseline.
Deshpande et al., Learning Diverse Image Colorization: https://arxiv.org/abs/1612.01958 and https://arxiv.org/html/1612.01958v2 — low-dimensional color-field variables coordinate alternatives over long spatial distances. Motivates global modes instead of pixel noise.
Wang et al., PalGAN: https://arxiv.org/html/2210.11204v1 — probabilistic palette prediction and assignment separate ambiguity from boundary handling. This project's small palette model is not a PalGAN reproduction.
Lee et al., Stochastic Multiple Choice Learning: https://arxiv.org/abs/1606.07839 — winner-based learning supports distinct hypotheses, but oracle quality is not automatic-default quality. We explicitly measure the gap and train an image-only selector.
Mehta and Rastegari, MobileViTv2: https://arxiv.org/abs/2206.02680 — efficient global attention in a mobile encoder. Test the ImageNet-pretrained 0.75 variant as a distinct backbone challenger, with all deployment parameters counted.
Saharia et al., Palette: https://arxiv.org/abs/2111.05826 — diffusion supports colorization diversity. Not a justified final-run choice here without training/validating a compact denoiser; repeating the old decoder did not test diffusion. Excluded from this finite screen because establishing a from-scratch diffusion model and sampler is substantially larger work than these warm-started alternatives.

Controlled comparisons

Six arms: single palette vs four coherent palettes, crossed with teacher / mixed-target / staged curriculum. Same MobileNet v3 initialization, fixed8192 training images, seeds, image-index batches, augmentation draws, AdamW, batch24, 3000 updates. Encoder LR1e-5, decoder LR1e-4, warmup150 and cosine decay to15%, weight decay0.01, frozen encoder BatchNorm, gradient norm cap1, BF16 forwards/FP32 losses. New mode tensors and selector are included in the decoder group. The multi head adds only1,156 parameters (3,995,832 total).

Multi head: shared spatial masks and encoder; four latent palette variants; image-only score selects one complete output. Training loss is95% winner reconstruction +5% mean reconstruction +0.1 selector cross-entropy against the detached winning index. Reconstruction is L1/20 +0.25 sliced color distribution distance +0.1 boundary difference. Winner is selected per entire image, never per pixel. Report mode usage, between-mode diversity, default vs oracle losses and mode-switching under input resize. No original color is available to the runtime selector. No hue rotation, saturation bonus or unvalidated generated 'plausible' labels.

Curricula:

Teacher: deterministic cached DDColor artistic targets.
Mixed: independently sample original photograph or teacher target with probability0.5 per image each update. Never average two hues and never pick the easier target. Both represent complete color proposals; teacher remains fallible.
Staged: first15% teacher, next35% ramp original probability to0.5; final third adds35% mild spectral/film/noise augmentation plus same-mode resize consistency (mean coefficient0.1, computed every fourth update). This tests a curriculum bundle, not isolated causal effects of each component. Anchored reconstruction remains active to guard against consistency-induced gray collapse.

Seventh arm: MobileViTv2-0.75 + single palette, copied compatible decoder weights and new lateral projections. 3,481,205 parameters.1000 extra teacher alignment updates, then3000 staged updates; encoder LR2e-5, decoder2e-4. Pretrained encoder pinned to the revision recorded in encoder_provenance.json. This is a practical challenger with unequal prior colorization training and an extra alignment stage, not a pure equal-initialization ablation. A negative short pilot does not prove the encoder incapable.

Data and evaluation

Training subset sampled from existing filtered16,230-image Imagenette/COCO training pool; no new teacher expense. Existing target cache hashes and row IDs checked. Pin dataset revisions. New128 development and160 test images from COCO val/0001 exclude earlier round7, region, recurrent evaluation IDs and exact-byte training duplicates. Teacher/encoder upstream exposure and near duplicates remain possible. Fix splits before training. Choose each arm's candidate from evaluations every1000 updates on development only; test opened after choices frozen. Keep all evaluated checkpoints.

Screen on ordinary and film-style grayscale. Record raw and guided patch metrics, chroma retention, missed color, spill, auxiliary original-color error, resize and flip sensitivity, latency, and image-bootstrap paired intervals. Gray output can win a smoothness metric; do not rank on patch score alone. Screening guardrails allow missed color/spill at most+1.5points, coverage at least95% baseline, and error at most110% baseline. These are diagnostic eligibility conditions, not proof of plausibility. All checkpoint scans are exploratory and intervals do not correct multiple comparisons.

Visual review: 32 fixed fresh examples per group, archival photos never used as neutral targets, seven familiar probes, and aspect-preserving production-pipeline outputs. Label comparisons anonymously with saved keys; assistant review is qualitative, not an independent human preference study. Archive source URLs and checksums retained. A modest historical panel is not a representative archival benchmark.

Targeted fine-tuning and RL

Late-stage robustness and consistency are tested rather than assumed helpful. No RL launched in this screen: no independently validated preference reward currently exists, and previous learned critics were exploitable. Prefer direct differentiable supervision for measurable consistency; if a later reward is proposed it must first rank coherent alternate hues above gray collapse and patchwork on held-out images. No claim that RL is intrinsically unsuitable.

Persistence, budget, and final selection

One L4 job with60min hard limit and50min internal training deadline; all outputs checkpointed to this main-branch experiment folder with SHA256 download verification. Source/access/preflight checks before training. The first attempt was canceled before training after about12min because of repeated NPZ decompression. Corrected version2 materializes each cache once and logs data preparation. A separate confirmation, if warranted, is capped18min; total maximum hardware charges at the verified$0.80/h rate approximately$1.20, excluding no separate service purchases. Current account balance is not exposed by the connector. No automatic paid reruns.

Choose architecture/curriculum from test plus visual review, with production baseline retained as an eligible fallback. If a candidate clearly wins, confirm against the matching control on a second seed or targeted ablation within the remaining cap. Save final_config.json and NEXT_PASS.md with exact checkpoints, source revisions, dataset/curriculum, learning rates, retention gates and validation instructions. The next pass starts the longer run and performs Space/model-card/repository work. This pass does not claim any longer run already completed.

Version2 corrections before training

NPZ cache members are decompressed once per shard, not once per selected image; a synthetic equivalence and timing check is saved. MobileViT now uses a separate post-warmup curriculum step. Six additional National Archives inputs supplement the original historical panel. Version1 was canceled before any training checkpoint; its source/preflight remains for audit.