Continuous-Time Distribution Matching: Research Distillation Contract¶
Continuous-Time Distribution Matching is a research method for few-step diffusion distillation. The paper describes continuous-time distribution matching for the named experimental architectures, and links to code. It is evidence for reproducing or extending that release under its conditions, not a drop-in promise of a specific step count, speedup, quality level, GPU topology, or compatibility with another teacher.
Bind teacher, student, and trajectory¶
For every experiment, retain:
- paper version, code commit, checkpoint/artifact identifiers and digests, license/access terms, framework/dependency revisions, and distributed runtime/hardware record;
- teacher and student parameterization, conditioning path, prediction target, noise/trajectory schedule, loss branches/weights, data pipeline, seeds, and checkpoint/resume policy;
- a source-disjoint evaluation set with task quality, diversity, non-target preservation, safety/rights, reproducibility, and failure measurements;
- comparable baseline runs using the same prompt/input distribution, output resolution, inference environment, warm/steady-state latency, throughput, memory, and error policy; and
- serving artifact/version, rollback target, output provenance, monitoring, and release decision.
Changing the teacher, base architecture, scheduler, conditioning contract, or distributed implementation changes the distillation problem. Do not inherit an adapter interface, schedule, loss branch, or published result from a different model family without an explicit implementation and measured validation.
Evaluate acceleration separately from output trust¶
Few-step output must be compared with the declared teacher and baseline for the requested task. Report speed/cost separately from fidelity, diversity, text/geometry preservation, and artifact rate. A visually attractive sample cannot prove that a student preserves source facts, behaves consistently on a new prompt distribution, or is safe to serve.
Keep checkpoints and output lineage sufficient to rerun a failed evaluation or roll back a deployed student. A training run that stops early, loses provenance, or cannot reproduce its evaluation remains incomplete.
Failure boundary¶
If the teacher/student contract, code/artifacts, licensing, source-disjoint evaluation, or rollback path is missing, keep CDM in research state. Do not substitute a different distillation method, publish an unmeasured checkpoint, or advertise a faster sampler as an equivalent CDM result.