Source
Liu, Gong and Liu (2022), LIT-636 — Rectified Flow, Table 1a.
The claim
Rectified Flow’s title promises generation that is “straight and fast”, and its abstract reports high quality from a single Euler step. That result needs two further procedures. Reflow retrains the model on pairs it generated itself, which straightens its paths. Distillation trains a one-step student. Table 1a separates them, on CIFAR-10 with the same architecture. FID at one Euler step:
| Model | Undistilled | Distilled |
|---|---|---|
| 1-rectified flow (no reflow) | 378 | 6.18 |
| 2-rectified flow (one reflow) | 12.21 | 4.85 |
Both procedures, and the table, come from LIT-636, which proposes reflow
as the way to straight paths and uses reflow followed by distillation for its
headline one-step result. The ordering recommended here, distillation first
and reflow optional, is a reading of that paper’s Table 1a, not advice the
paper gives. That is why LIT-636 is this practice’s source and not its
origin: the paper supplies the numbers and argues for the opposite pipeline,
and the record has found no paper that makes the recommendation, so
introduced_by is left empty (ADR-053).
Distillation does most of the work. It takes the base model from 378 to 6.18 on its own. Reflow before distillation improves that to 4.85, at the cost of generating a paired dataset and training a second time.
Reflow costs quality at many steps. With full RK45 sampling, FID worsens from 2.58 for the base model to 3.36 after one reflow and 3.96 after two. The paper says reflow “worsens the results in the large step regime” (§5.2).
So: distil first. Reflow if the last increment of one-step quality is worth another training pass, and keep the pre-reflow model when you need many-step quality. That also follows SOTA-206, which says to keep a multi-step option in a few-step model.
Contested at scale, from a different starting point
InstaFlow (LIT-796) runs the comparison this practice’s promote_when describes, at text-to-image scale and matched budget, and finds the opposite order. On Stable Diffusion 1.4 with the same U-Net, batch, 100K training steps and 3.2M generated pairs, distilling SD directly to one step reaches FID 40.9 on COCO-2017-5k and 34.6 on COCO-2014-30k. Spending half the budget on one reflow and half on distillation reaches 31.0 and 20.0 (its Table 1, App. C). The tuning, if anything, favours the direct arm, which got a nine-cell learning-rate grid. It is one seed at 0.9B.
It differs from this practice’s case in two ways, and both are why it narrows the claim rather than refuting it. Its starting point is a diffusion model’s curved probability-flow ODE, so the reflowed model is the first rectified flow in its pipeline. And its distiller is an LPIPS regression onto one teacher output per noise, not distribution matching. For regression distillation from a curved teacher, reflow is not optional. When the teacher is already a rectified flow, the only measurement is still Rectified Flow’s CIFAR-10 table, where reflow adds 6.18 → 4.85 rather than deciding success.
It agrees with the practice’s last clause. Its reflowed model loses a little many-step quality, 21.5 against SD 1.5’s 20.1 FID-5k at 25 steps, so keeping the pre-reflow model for many steps still holds. A second reflow did not reliably help (its Table 4).
If you reflow, reflow once, and train it like a distillation. Lee et al. (LIT-791) refine the reflow stage rather than test the ordering. Starting from EDM, one round trained with a U-shaped timestep distribution and an LPIPS-Huber loss reaches one-step FID 3.07 on CIFAR-10 and 4.31 on ImageNet-64 with no separate distillation stage, and they argue further rounds only add error. Their own table has DMD, a distribution-matching distiller with no reflow, ahead on ImageNet-64 at 2.62. They run no distillation-alone arm from the same teacher and no matched many-step comparison, so they do not meet promote_when.
Adjacent evidence for distilling without reflow. Shortcut Models (LIT-787, Table 1) start from a flow-matching model, a 1-rectified flow, on DiT-B at 256 px. Progressive distillation alone beats one reflow at one step on both datasets, 14.8 against 23.2 on CelebA-HQ and 35.6 against 44.8 on ImageNet, at roughly matched compute. There is no reflow-then-distil arm, so it does not test the ordering either. It does say that from a straight-path teacher, distillation without reflow beats reflow without distillation, which is this practice’s direction.
Conditions
- CIFAR-10 only. The comparison is at 32×32 on one dataset. The paper’s larger-image results are qualitative.
- Training budgets are not matched. Reflow-plus-distillation costs more training than distillation alone, and the table does not normalize for it.
- The distillation uses an LPIPS loss for the one-step student (App. A). Whether the balance holds with other distillation objectives, such as the distribution matching the video line uses, is untested. InstaFlow’s test uses LPIPS regression too.
- The teacher has to be a rectified flow already. From a curved diffusion teacher with regression distillation, the one test at scale finds reflow first better by 9.9 FID-5k (LIT-796, above). The practice does not cover that case. InstaFlow was listed here as an implementation until the correction pass after #395, because it implements the opposite pipeline; it is now filed as the paper that contests this one, and on 2026-10-03 its listing among Rectified Flow’s implementations was removed for the same reason.
- The theory is about the reflowed coupling. The straightness theorems apply to the model after reflow, not to the single-pass model most people train (SOTA-266).