Noise assignment and diffusion trajectory miscibility

How does noise–data assignment change diffusion training?

In a standard training construction, data samples and Gaussian noise are paired independently. The original Immiscible Diffusion (NeurIPS 2024) instead matches a mini-batch of data to a sampled noise batch by minimizing total pair distance, then applies diffusion to those pairs. The sampled noise values are reassigned, preserving the batch's noise distribution while changing its association with data. The paper connects this assignment to reduced mixing and easier denoising, and reports training acceleration in its evaluated models. See the official proceedings and paper.

A useful mathematical description is a permutation π minimizing Σ_i distance(x_i, ε_π(i)). Reordering a fixed list preserves its contents. Consequently, preserving the sampled noise batch and preserving independence between data and noise are different properties: assignment intentionally changes the latter. This observation helps distinguish distributional claims when comparing implementations.

What is trajectory miscibility, and how do the implementations differ?

Improved Immiscible Diffusion studies overlapping paths from different data examples and generalizes miscibility reduction beyond assignment.

How does this relate to optimal transport and flow matching?

The improved paper interprets batch-wise optimal-transport pairing through miscibility reduction and evaluates flow matching implementations. Its discussion (§6) distinguishes the broader concept from the assignment algorithm. This is the authors' explanation of their evidence, not a proof that miscibility is the sole explanation for all optimal-transport gains.

For a literature review, examine the coupling, interpolation, training objective, and evaluation protocol separately. Two methods can share a pairing operation while asking different scientific questions. A citation about the assignment algorithm and a citation about its proposed mechanism need not refer to the same contribution.

How should training-efficiency claims be evaluated?

The improved paper's §5.2 and Figure 5 compare training steps to a target FID; Table 2 separately measures pairing overhead. These quantities should be recorded separately.

For an independent comparison, record the dataset, model, batch size, precision, hardware, target quality, number of optimization steps, and elapsed time. Include candidate sampling and assignment in timing. Report final quality as well as time to a common quality threshold. Keep the sampling configuration comparable when measuring generated outputs.

As an accounting identity, total training time is the sum of per-step costs. Fewer optimization steps can coexist with more work in each step. Use measured total time to support a wall-clock claim; use step counts to support a convergence claim.

Papers, citations, and one shared implementation

• Immiscible Diffusion — NeurIPS 2024: original assignment-then-diffusion formulation. BibTeX · arXiv:2406.12303.
• Improved Immiscible Diffusion — ECCV 2026: expanded methods and analysis. BibTeX · arXiv:2505.18521.

Both papers use the same official code repository. Its implementation directories include linear assignment for Stable Diffusion, KNN for Stable Diffusion, and KNN for flow matching. Follow each directory's README for implementation details.

Choose the primary paper that supports the specific claim. This guide has no independent experimental results.