Source
Zheng, Wang, Ma, Chen, Zhang, Balaji, Chen, Liu, Zhu and Zhang (2025), LIT-783 — rCM, §3.3, §4.1, Tables 1–2 and Fig. 7.
Lu and Song (2024), LIT-790 — sCM, Fig. 7.
What to do
Distil with the continuous-time consistency loss of sCM as the main objective. Add DMD’s distribution-matching loss as a regularizer:
- L = L_sCM + λ·L_DMD, with λ = 0.01.
- Compute L_DMD on the student’s own samples, from a rollout of a random 1 to 4 steps, not on noised real data.
- Train a fake-score network alongside, as in DMD, and update it several times per student update (5, and 10 at 14B). No GAN term is needed.
- Keep sCM’s tangent normalization. At 10B parameters and above, and for video, keep the full JVP rather than a finite difference for the time derivative, and run the time-embedding layers in FP32.
The weighting matters in both directions. The consistency term follows the teacher’s ODE and keeps diversity. The distribution-matching term pulls the student’s samples toward the teacher’s distribution and restores detail. Make the second just strong enough to fix detail.
Evidence
Pure sCM fails at scale (LIT-783, §3.3, Fig. 3, App. E). On Cosmos-Predict2 at 2B and 14B and on Wan2.1 at 1.3B, four-step sCM students are sharp and close to the teacher on ordinary prompts. They break on small text. In video, textures blur and objects interpenetrate across frames. Scaling the model does not fix it. The authors attribute this to error accumulation: the self-feedback term in the JVP dominates at large t, where the teacher’s supervision vanishes, and BF16 makes the JVP’s error much larger than the output’s (Fig. 11). This evidence is images and videos, with no table.
The regularizer fixes it, and its weight has a floor (Fig. 7, Wan2.1-1.3B, 10K iterations at batch 64, four steps):
| λ | 1 | 0.1 | 0.01 | 0.001 |
|---|---|---|---|---|
| VBench | 84.32 | 84.57 | 84.43 | 82.68 |
Quality holds from λ = 1 down to 0.01 and drops at 0.001. The authors read the samples, five seeds per λ, as trading diversity for quality as λ rises, and pick 0.01 as “the smallest scale to preserve good quality”.
At full scale it matches DMD2 on quality (Tables 1–2). Four-step GenEval on Cosmos-Predict2 is 0.79, 0.81 and 0.83 at 0.6B, 2B and 14B, against teachers at 0.81, 0.83 and 0.84 and a re-implemented DMD2 at 0.77 and 0.80 (0.6B and 2B). Four-step VBench on Wan2.1 is 84.43 at 1.3B, against DMD2’s 84.56 and the teacher’s 83.02, and 84.92 at 14B. One student also serves one and two steps.
Why keep consistency as the main term (LIT-790, Fig. 7). On one EDM2-M backbone at ImageNet-512, sCM ran one-step VSD, which is DMD’s distribution-matching gradient without its regression loss, against two-step sCD. As guidance rises, VSD’s precision rises and its recall falls, ending in “severe mode collapse”. sCD’s precision and recall stay close to the teacher’s. Read from the plot (no values are tabled): one-step sCD is plotted too, and at guidance 1.0 its recall is about 0.70 against about 0.65 for one-step VSD, so the gap is there at matched step count and before guidance is raised. The arm that sums the two losses at equal weight tracks VSD’s recall, not sCD’s. That is the record’s one measurement of the diversity argument, it comes from a different group from rCM, and it is why the DMD term is added with a small weight rather than at parity.
The loss being added is DMD’s (LIT-643): the gradient is a frozen real score minus an online fake score, evaluated on noised generator samples, normalized by the mean absolute denoising error. DMD needed a regression loss onto teacher ODE outputs to train stably. DMD2 (LIT-646) replaced that with several fake-score updates per generator step and recovered 2.61 FID on ImageNet-64 without it (Table 3 there). rCM keeps DMD2’s update ratio and drops its GAN.
Conditions
- The comparison that matters is not tabled. No λ = 0 row exists, and pure sCM against rCM is shown only in samples. The λ = 0.001 row (82.68) is the closest number, and it is a regularized run.
- The diversity claim is from samples. rCM’s “notable advantages in diversity” over DMD2 rest on five videos per method (Fig. 1) and five seeds per λ (Fig. 7). No diversity metric is reported. sCM’s Fig. 7 is the measured support for the direction, read from curves; its one-step sCD arm makes the comparison like for like.
- The DMD2 baseline is a re-implementation, with a discriminator branch on the fake-score network and no configuration or budget given. On 1.3B VBench it is slightly ahead of rCM.
- It is not free of tuning. It keeps DMD2’s fake-score network and its 5 to 10 critic updates per student step. Learning rates and the sampling σ_max differ by model.
- A rival distiller beats it with people. DMAD (LIT-770) uses rCM’s Wan2.1 rows as baselines, and its human study prefers DMAD to rCM on 63.8% of prompts, ties counted as non-wins. That compares two distillers. It does not test whether the DMD term helps sCM.
- Only at large scale. Every rCM result is on 0.6B–14B text-conditioned teachers. Whether pure sCM’s detail failures appear on small class-conditional models, where sCM reaches 1.88 FID at two steps on ImageNet-512 (LIT-790), is not tested.
Beside the record’s other practices
SOTA-394’s diversity condition cites both sources: distribution matching on its own costs diversity. This practice is the response that keeps distribution matching for detail while letting a forward-direction term carry diversity. SOTA-206 also cites rCM: one rCM student serves one, two and four steps, and video needs two.
Known implementations
- Cosmos-Predict2 and Wan2.1 T2V students distilled with rCM, the source’s released models.