Preference pairs
One prompt, a human-preferred image and a rejected image. Preserve identities and hashes.
Structure-aware preference alignment
Beyond preference.
Inspect the constraints.
Flow-matching preference optimization with decomposed semantic and structural evidence.
Exploratory automated-signal ablations — not human-calibrated128Real preference pairs
5 / 5Completed training runs
640Recorded optimizer steps
4.127 GiBPeak allocated T4 memory
The research question
Implemented & pilot-trainedCan component-level constraints improve prompt and structural correctness over ordinary human-preference optimization?
Eight signals
Semantic, presence, color, attribute, count, relation, geometry and visual-quality proxies.
Frozen reference + LoRA
SD3 Medium flow backbone. A shared pooled OT schedule and human-gradient conflict protection.
Controlled ablations
Five objectives, identical starting adapters and 128 steps, followed by 360 matched generated outputs.
What is different in our design?
Instead of collapsing every signal into a single preference score, preserve separate signed constraint margins, applicability masks and weights.
- Ignore constraints that do not apply to the prompt.
- Use a minimum-margin deadband to suppress tiny score differences.
- Limit conflicting auxiliary influence so it cannot reverse the human preference direction.
- Keep reliability gates separate from conflict protection.
These constraints supervise a shared flow-preference logit; they are not independent detection heads or direct pixel-space losses. OT is the coupling used within training, not a separately trained model.
What this pilot establishes
- All five real-model training runs completed on a T4.
- Same data, initialization, schedule and base-model revision across runs.
- All 360 matched evaluation images and decomposed scores were produced.
It does not establish superiority over the base papers or a validated human-calibrated reward system. Most automatic metric differences remain uncertain.
Reliability weights are one for active signals in this campaign. The final reliability-gated method remains a later experiment.
Inspect a real training pair
All 128 frozen training pairs are available here. These are dataset images, not outputs from the new adapters.
Decomposed evidence
| Signal | Preferred | Rejected | Delta | Eligible |
|---|
Delta = preferred − rejected. A negative delta indicates evaluator disagreement with the human ranking, not a proven labeling error. Eligible means applicable and |delta| ≥ 0.02, before conflict protection.
Automatic scores are uncalibrated proxies, not correctness probabilities. Only 2 of 128 pairs activate count loss; this pilot cannot establish count-learning improvements.
Recorded training, inspectable step by step
Completed on T4 / 16 SeptemberThis is a replay of actual saved metrics, not a currently running training job.
128 / 128 steps
Loss
Gradient norm
Preference logit
Actual auxiliary activity
Steps with a nonzero logged contribution, out of 128.
Final checkpoint fingerprint
Different objectives have different loss terms, so lower total training loss is not evidence that one model is better. Preference logits also do not substitute for held-out image evaluation.
Controlled configuration
| Base | SD3 Medium, pinned revision; frozen reference behavior | Policy | LoRA rank 4, FP32 trainable weights |
|---|---|---|---|
| Schedule | 128 steps; checkpoint every 16 | Data / seed | Same 128 pairs / seed 42 |
| Learning rate | 0.00005 | Memory | FP16 base; 256 px; batch 1 |
| OT | Pool 8, identical assignment hash | Initialization | Restored before each run |
Controlled generated-image comparison
360 / 360 generatedSix outputs use the same prompt, seed, resolution and inference schedule. These are actual SD3 Medium and trained-LoRA outputs from the completed T4 evaluation.
Automatic scores are uncalibrated proxies. This run does not establish a human preference winner, and count conclusions are especially limited because only 2/128 training pairs activated count loss.
Verify the artifacts
This hosted copy displays recorded local build validation. Rechecking the research files requires the local Python server on the project computer.
Recorded local build validation passed. This hosted copy does not perform a fresh local-file integrity check. No credentials or human-calibration key are included.
Frozen experiment-plan hash
What the evaluation says
Exploratory result complete60 prompts × 6 models = 360 images
- All variants completed under the same inference protocol.
- The structural-only adapter improved the automatic geometry proxy over the preference baseline by 0.0187; exploratory 95% interval 0.0007 to 0.0378.
- Most aggregate differences remain uncertain, and visual-quality differences are not clearly separated.
- No blind human win rate is inferred from these automatic scores.
This supports technical feasibility and identifies a signal worth follow-up. It is not a claim of superiority or a human-calibrated result.