schema_version: vision_question_clusters_v1 description: Conditional synthetic image generation ranking with true diffusion training and sampling. Analyst hypotheses are excluded from prompts. Energy score is comparable across parameterizations; denoising losses are not ranked. dataset_family: conditional_image_generation selection_metric: test_energy_score base_profile: image-generation-ranking-base-v1.yaml default_num_choices: 3 allowed_num_choices: - 3 - 4 - 5 instance_seeds: - 851901 - 851902 - 851903 discovery_training_seeds: - 20 - 21 - 22 - 23 confirmation_training_seeds: - 200 - 201 - 202 - 203 - 204 - 205 - 206 - 207 - 208 - 209 - 210 - 211 - 212 - 213 - 214 - 215 statistics: discovery: n_seeds: 4 split: validation method: direction_estimation ordering: mean_validation_energy_score admission_gate: false purpose: Freeze adjacent ranking directions before independent confirmation; no significance claim from four discovery seeds. confirmation: n_seeds: 16 split: test test: one_sided_paired_sign_test alpha: 0.05 correction: holm minimum_mean_gap: 0.003 estimand: Training-seed paired direction stability conditional on one fixed dataset and fixed evaluation-noise/reference stream, with observed mean energy-score gap; not a test of expected-score mean difference or a natural-image population claim. selection_contract: discovery_split: validation confirmation_split: test minimum_mean_gap: 0.003 require_nonoverlap: false multiplicity: Holm across every frozen adjacent hypothesis including rejected templates claim_scope: Fixed synthetic image distribution and fixed independent evaluation sample streams; training-seed robustness only. difficulty_policy: Evaluate every admitted triple and report rejected templates without replacement. defaults: dataset: size: 16 noise_steps: 32 condition_dim: 6 brightness_min: 0.65 brightness_max: 1.0 texture_strength: 0.12 radius_min: 0.25 radius_max: 0.42 position_extent: 0.45 reference_samples: 4 train_size: 1024 validation_size: 64 test_size: 64 ood_size: 32 evaluation_shift: {} ood_shift: position_extent: 0.55 optimizer: type: AdamW lr: 0.001 weight_decay: 0.01 betas: - 0.9 - 0.999 eps: 1.0e-08 foreach: false loss: loss_id: diffusion_mse training: batch_size: 16 training_steps: 3000 total_samples_seen: 48000 gradient_accumulation: 1 clip_norm: 1.0 schedule: constant precision: fp32 tf32: false deterministic_algorithms: true warmup_updates: 0 batch_rng_device: cpu paired_cpu_batch_stream: true persist_initial_state: true batch_seed_offset: 500000 noise_seed_offset: 600000 evaluation_split: validation eval_steps: - 750 - 1500 - 3000 prediction_type: x0 noise_steps: 32 noise_schedule: cosine loss_weighting: uniform min_snr_gamma: 5.0 sampling_steps: 16 sampling_eta: 0.0 evaluation_samples_per_condition: 4 evaluation_seed: 314159 evaluation_batch_conditions: 16 models: conv: type: image_generation_denoiser mode: conv width: 24 depth: 3 kernel: 3 norm: group activation: silu residual: true block: dense conditioning: concat coordinates: absolute time_embedding: sinusoidal max_parameters: 2000000 unet: type: image_generation_denoiser mode: unet width: 24 depth: 3 kernel: 3 norm: group activation: silu residual: true block: dense skip: true conditioning: concat coordinates: absolute time_embedding: sinusoidal max_parameters: 2000000 mlp: type: image_generation_denoiser mode: mlp width: 24 depth: 3 activation: silu conditioning: concat coordinates: absolute time_embedding: sinusoidal max_parameters: 2000000 transformer: type: image_generation_denoiser mode: patch_transformer width: 24 depth: 3 activation: gelu conditioning: concat coordinates: absolute time_embedding: sinusoidal patch_size: 2 max_parameters: 2000000 num_heads: 4 ff_multiplier: 2 contexts: textured: {} clean: texture_strength: 0.0 small_data: train_size: 96 shifted: evaluation_shift: position_extent: 0.55 surface_heuristics: - id: more_parameters field: parameter_count higher_is_better: true - id: more_training_steps field: training.training_steps higher_is_better: true - id: more_sampling_steps field: training.sampling_steps higher_is_better: true templates: - id: operator_local_global category: operator context_ref: textured budget_semantics: matched_steps hypothesis: Local convolution, down/up processing and global patch attention differ in spatial bias and finite-budget optimization. varying_axes: - model.mode choices: - id: conv model_ref: conv - id: unet model_ref: unet - id: transformer model_ref: transformer diagnostics: [] controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: operator_coordinates category: operator context_ref: textured budget_semantics: matched_steps hypothesis: Coordinate priors and spatial weight sharing interact; operator labels alone do not isolate the change. varying_axes: - model.mode - model.coordinates choices: - id: conv_absolute model_ref: conv - id: conv_relative model_ref: conv model_updates: coordinates: relative - id: flat_mlp model_ref: mlp diagnostics: - id: mlp_relative model_ref: mlp model_updates: coordinates: relative controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: kernel_depth category: block context_ref: textured budget_semantics: matched_steps hypothesis: Larger spatial support competes with depth, since coordinates already reveal location. varying_axes: - model.kernel - model.depth choices: - id: k1_d5 model_ref: conv model_updates: kernel: 1 depth: 5 - id: k3_d3 model_ref: conv - id: k5_d1 model_ref: conv model_updates: kernel: 5 depth: 1 diagnostics: - id: k1_d3 model_ref: conv model_updates: kernel: 1 - id: k5_d3 model_ref: conv model_updates: kernel: 5 controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: depthwise_width category: block context_ref: textured budget_semantics: matched_steps hypothesis: Channel mixing, spatial mixing and width need not trade off monotonically. varying_axes: - model.block - model.width choices: - id: dense24 model_ref: conv - id: dw24 model_ref: conv model_updates: block: depthwise - id: dw48 model_ref: conv model_updates: block: depthwise width: 48 diagnostics: - id: dense48 model_ref: conv model_updates: width: 48 controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: residual_depth category: residual context_ref: textured budget_semantics: matched_steps hypothesis: Depth and residuals interact under a fixed training budget. varying_axes: - model.depth - model.residual choices: - id: res_d2 model_ref: conv model_updates: depth: 2 - id: res_d5 model_ref: conv model_updates: depth: 5 - id: plain_d5 model_ref: conv model_updates: depth: 5 residual: false diagnostics: - id: plain_d2 model_ref: conv model_updates: depth: 2 residual: false controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: unet_skip_width category: residual context_ref: textured budget_semantics: matched_steps hypothesis: Skip paths preserve resolution while extra width changes optimization and capacity. varying_axes: - model.skip - model.width choices: - id: skip24 model_ref: unet - id: noskip24 model_ref: unet model_updates: skip: false - id: noskip48 model_ref: unet model_updates: skip: false width: 48 diagnostics: - id: skip48 model_ref: unet model_updates: width: 48 controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: conditioning_injection category: conditioning context_ref: textured budget_semantics: matched_steps hypothesis: Repeated modulation, input bias and explicit condition channels have different trainability. varying_axes: - model.conditioning choices: - id: concat model_ref: conv - id: film model_ref: conv model_updates: conditioning: film - id: input model_ref: conv model_updates: conditioning: input diagnostics: [] controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: time_parameterization category: time context_ref: textured budget_semantics: matched_steps hypothesis: Time information can matter differently for clean-image and noise prediction. varying_axes: - model.time_embedding - training.prediction_type choices: - id: x0_sinusoidal model_ref: conv - id: x0_none model_ref: conv model_updates: time_embedding: none - id: epsilon_sinusoidal model_ref: conv training_updates: prediction_type: epsilon diagnostics: - id: epsilon_none model_ref: conv model_updates: time_embedding: none training_updates: prediction_type: epsilon controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: prediction_parameterization category: objective context_ref: textured budget_semantics: matched_steps hypothesis: Equivalent representational parameterizations use different finite-budget losses; only sampled image energy is comparable. varying_axes: - training.prediction_type choices: - id: x0 model_ref: conv - id: epsilon model_ref: conv training_updates: prediction_type: epsilon - id: v model_ref: conv training_updates: prediction_type: v diagnostics: [] controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: noise_loss_weighting category: noise context_ref: textured budget_semantics: matched_steps hypothesis: Noise allocation and SNR weighting jointly change optimization emphasis. varying_axes: - training.noise_schedule - training.loss_weighting choices: - id: cosine_uniform model_ref: conv - id: linear_uniform model_ref: conv training_updates: noise_schedule: linear_beta - id: cosine_minsnr model_ref: conv training_updates: loss_weighting: min_snr diagnostics: - id: linear_minsnr model_ref: conv training_updates: noise_schedule: linear_beta loss_weighting: min_snr controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: norm_lr category: optimization context_ref: textured budget_semantics: matched_steps hypothesis: Normalization and learning rate should be separated by completed controls. varying_axes: - model.norm - optimizer.lr choices: - id: group_lr1 model_ref: conv - id: none_lr1 model_ref: conv model_updates: norm: none - id: none_lr3 model_ref: conv model_updates: norm: none optimizer_updates: lr: 0.003 diagnostics: - id: group_lr3 model_ref: conv optimizer_updates: lr: 0.003 controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: batch_fixed_samples category: optimization context_ref: textured budget_semantics: matched_samples hypothesis: At equal samples seen, batch size changes update count and noise. varying_axes: - training.batch_size - training.training_steps choices: - id: batch8 model_ref: conv training_updates: batch_size: 8 training_steps: 6000 - id: batch16 model_ref: conv - id: batch32 model_ref: conv training_updates: batch_size: 32 training_steps: 1500 diagnostics: [] controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: training_budget_small_data category: budget context_ref: small_data budget_semantics: per_choice_steps hypothesis: Longer training may fit finite appearance draws without improving conditional sample distribution. varying_axes: - training.training_steps choices: - id: steps700 model_ref: conv training_updates: training_steps: 700 - id: steps3000 model_ref: conv training_updates: training_steps: 3000 - id: steps6000 model_ref: conv training_updates: training_steps: 6000 diagnostics: [] controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: sampler_steps_parameterization category: sampling context_ref: textured budget_semantics: matched_steps hypothesis: Inference discretization and prediction parameterization interact; extra network evaluations do not guarantee sample quality. varying_axes: - training.sampling_steps - training.prediction_type choices: - id: x0_steps4 model_ref: conv training_updates: sampling_steps: 4 - id: x0_steps32 model_ref: conv training_updates: sampling_steps: 32 - id: v_steps32 model_ref: conv training_updates: prediction_type: v sampling_steps: 32 diagnostics: - id: v_steps4 model_ref: conv training_updates: prediction_type: v sampling_steps: 4 controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: sampling_stochasticity category: sampling context_ref: textured budget_semantics: matched_steps hypothesis: Stochastic reverse steps can alter diversity and condition accuracy together. varying_axes: - training.sampling_steps - training.sampling_eta choices: - id: det4 model_ref: conv training_updates: sampling_steps: 4 - id: det16 model_ref: conv - id: stoch16 model_ref: conv training_updates: sampling_eta: 1.0 diagnostics: - id: stoch4 model_ref: conv training_updates: sampling_steps: 4 sampling_eta: 1.0 controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: texture_condition_bridge category: context context_ref: clean budget_semantics: matched_steps hypothesis: The same coordinate/operator contrast is repeated under a no-texture conditional appearance distribution. varying_axes: - model.mode - model.coordinates choices: - id: conv_absolute model_ref: conv - id: conv_relative model_ref: conv model_updates: coordinates: relative - id: unet model_ref: unet diagnostics: - id: unet_relative model_ref: unet model_updates: coordinates: relative controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: position_condition_bridge category: context context_ref: shifted budget_semantics: matched_steps hypothesis: The same coordinate/operator cells bridge a broader held-out position distribution. varying_axes: - model.mode - model.coordinates choices: - id: conv_absolute model_ref: conv - id: conv_relative model_ref: conv model_updates: coordinates: relative - id: unet model_ref: unet diagnostics: - id: unet_relative model_ref: unet model_updates: coordinates: relative controls: - Identical materialized train/validation/test distribution within a question. - All unlisted settings inherit explicit defaults; measured parameter counts are shown. - Sample quality uses the same fixed conditional reference draws and initial-noise streams. shortcut_checks: - larger network - more sampling steps - matching generator geometry explanation_target: Use the isolated local controls and actual sampling/condition/diversity diagnostics; distinguish optimization from distribution quality. admission: Freeze discovery order before independent confirmation; retain rejected triples without replacement. - id: budget_parameterization category: budget_objective context_ref: textured budget_semantics: per_choice_steps hypothesis: A noise-robust conditional mean can learn quickly while a stochastic conditional distribution requires more updates; parameterization rankings can change with budget. varying_axes: - training.prediction_type - training.training_steps choices: - id: relative_x0_3000 model_ref: conv model_updates: coordinates: relative training_updates: prediction_type: x0 training_steps: 3000 - id: relative_v3000 model_ref: conv model_updates: coordinates: relative training_updates: prediction_type: v training_steps: 3000 - id: relative_v6000 model_ref: conv model_updates: coordinates: relative training_updates: prediction_type: v training_steps: 6000 diagnostics: - id: relative_x0_6000 model_ref: conv model_updates: coordinates: relative training_updates: prediction_type: x0 training_steps: 6000 controls: - Complete 2x2 clean/velocity target by 3000/6000 update grid. - Same data, model architecture, optimizer, mini-batch and diffusion-noise streams. - Prediction-target losses are not comparable; rank only actual sample energy score. shortcut_checks: - always prefer clean-image prediction - always prefer velocity prediction - more training steps explanation_target: Separate matched-budget target effects and same-target budget effects. Conditional silhouettes can improve while appearance diversity shrinks; inspect sample energy, geometry and variance together. admission: Freeze discovery order before independent confirmation; do not replace rejected cells. calibration_decision: dataset_seed: 851900 training_seed: 20 formal_seeds_never_used: true decision_before_formal_runs: true calibration_runs: - generation/calibration - generation/calibration_long - generation/calibration_v6000 decision: Use 3000-step x0 default after independent calibration showed useful geometry. Retain 700-step short budget and add full x0/v by 3000/6000 controls to include a stochastic generation anchor. No requirement that every candidate fit the distribution well; weak controlled variants are retained. limits: The v6000 relative-coordinate pilot is useful but remains above oracle energy score; image variance is not proof of correctly learned texture modes. The calibration test split is exploratory only and never used as formal evidence.