Standard GICO on CIFAR-10 and ImageNet-64 uses KID supervision, including RF++ and EDM VE interpreted as 1-RF. GICO over frozen BézierFlow also uses KID. Only the GICO-TF target-fidelity variant uses LPIPS.
| Method | Objective protocol | Paired terminal measurement |
|---|---|---|
| GICO on native RF++ / EDM-as-1RF | paired-image-kid-v1 |
Unbiased KID between generated and real-reference feature blocks |
| GICO + frozen BézierFlow | paired-cifar-kid-v1 |
Unbiased KID with the frozen two-grid generator binding |
| GICO-TF | paired-lpips-v1 |
VGG-LPIPS to a same-noise/class high-accuracy target |
The image objective is explicit in every measurement and artifact. deterministic and stochastic choose the policy architecture under that objective; they do not choose the supervision metric. Missing objectives and mixed or mislabelled LPIPS/KID evidence fail validation.
genode-image-gico prepare accepts a backbone manifest, rows, and ImageNet's native class table. Use metrics: {"kid": value} with measurement_protocol: "paired-image-kid-v1". The image_objective.generator is the native backbone/context binding plus backbone and clock_protocol: "density64-euler". The features object pins weights_sha256, implementation_sha256, input_protocol: "edm-uint8", and estimator: "unbiased-cubic". Compute the implementation identity with feature_implementation_identity from genode.gico.image_evaluator.
Each row includes a sample_block with complete unique seeds and corresponding noise_sha256 entries, and a reference_block with the source dataset_sha256 and unique absolute dataset indices. Both blocks contain at least two members. Set ensemble_size to generated block size, seed to its first seed and reference_id to kid_objective.reference_identity(reference_block). ImageNet blocks also carry their native class_id; generated and reference blocks must agree. A panel_id identifies the comparison panel. CIFAR context is zero; ImageNet context is the frozen native class embedding. Research collection covers all 1,000 ImageNet classes once, assigning 40 classes per density; fitting and held-out classes are disjoint. Final dataset-wide reports require all 1,000 classes and equal class weight.
Keep raw unbiased KID estimates, including negative values. Average repeated block measurements within each comparison cell, then compute uniform KID minus candidate KID. Divide by one frozen training-only standard deviation, balancing NFEs and classes. Uniform utility is exactly zero. No log ratios, numerical floors, shrinkage or per-context reward normalization apply. Each candidate uses the complete unchanged generated/reference block. Reference data, generation noise and panel identities cannot cross fitting and held-out splits. No high-accuracy generated target is used.
For the same KID supervision over a frozen transformed sampler, use the separate GICO + BézierFlow contract through the common fitting API.
GICO-TF optimizes fidelity to a frozen high-accuracy target generator. The learned GICO reward utility surrogate predicts the paired LPIPS improvement; it is distinct from that target generator. Supply metrics: {"lpips": value}, ensemble_size: 1, measurement_protocol: "paired-lpips-v1", panel_id, and the shared solver/density/grid fields. Every row carries an image_objective with:
protocol: "paired-lpips-v1";target_generator:backbone, checkpoint SHA-256, solver, positivertol/atol, nativetime_range, andprecision: "float32";lpips:network: "vgg", complete scorer weights SHA-256, implementation SHA-256, version, andinput_protocol: "decoded-float32-no-resize-no-clamp".
The target records seed, class_id (null for CIFAR), noise_sha256 and image_sha256. Compute reference_id with genode.gico.image_objective.target_identity(objective, target). The same noise/class must use the same target across clocks/NFEs. Panels, noise and targets must not cross training, calibration, selection or test splits.
Measure VGG-LPIPS on the pinned comparator's decoded float32 tensors without resizing, clamping or uint8 conversion. Average repeated measurements within each panel/class/settings group, then take uniform LPIPS minus candidate LPIPS and divide by the frozen training-only scalar scale. For a matched-supervision GICO-TF/BézierFlow comparison, use identical target pairs and train/selection splits. Retain method-specific optimizers and disclose target preparation, fitting and selection costs separately. Check target-solver convergence on training noise before reuse. This LPIPS comparison is distinct from KID-supervised GICO over frozen BézierFlow.
Both objectives use the shared utility surrogate, deterministic and stochastic policies, utility surrogate-weighted distillation and differentiable score chasing. Initial generation noise is never a policy input. At inference only the selected policy is loaded; no target generation, utility surrogate scoring or reward selection occurs.
Select the utility-surrogate checkpoint by balanced calibrated component MSE, then its temperature by measured held-out reference-mixture utility regret, with equal panel and density-family weight. Temperature is in raw KID-improvement units for GICO or raw LPIPS-improvement units for GICO-TF, before scalar reward normalization. Auxiliary-score normalization and clipping remain separate. Select deterministic and stochastic policies by raw/expected calibrated frozen-utility surrogate utility under their separate 15% density-KL and 20% full-distribution-KL gates. Both use the common context holdout and make no generator/scorer calls.
Fresh fits use 2,000 utility surrogate and policy steps, zero image dropout, refinement weight 0.05 and temperature candidates 0.05/0.1/0.5. Checkpoint MSE ties prefer the earlier step; temperature-regret ties prefer 0.05. Overrides remain explicit. Calibrations record their actual training NFEs; an explicit calibration subset must remain inside eligible fitting observations of the completed collection.
image_protocol_metadata() describes native GICO KID; image_protocol_metadata(method="GICO-TF") describes explicit LPIPS. The descriptive identities are image_euler_kid_collection and image_euler_lpips_collection. The collection protocol budgets 10,000 images per NFE for CIFAR and 125,440 for ImageNet, including the two repeats and reusing uniform-assigned outputs. Metadata reports dataset-wide costs; fitting and selection add zero generator solves. Reference/target preparation and final reporting are separate. For LPIPS use CollectionConfig(image_objective="lpips") to plan individual target pairs within the same image allowances. Historical metadata and artifacts remain untouched and require their archived runtimes.
FID50k independently evaluates the resulting distribution. Neither lower fitting KID nor lower LPIPS establishes a FID improvement. This correction performs no beta sweep, benchmark rerun or new performance comparison.