Source
Yin et al. (2024), LIT-631 — CausVid, Table 4.
The claim
A streaming video generator has to be causal: each chunk is generated from the ones before it, with a KV cache. It does not follow that its teacher has to be causal. Distil the causal student from a bidirectional teacher, which sees the whole clip and generates it well, rather than first making the teacher causal.
The recommendation is CausVid’s (LIT-631). Its method, asymmetric distillation, trains a block-causal 4-step student with distribution matching against a bidirectional teacher of the same architecture. It compares the two at the same ODE-regression initialization, architecture, data and 4 sampling steps (Table 4), scored as temporal quality / frame quality / text alignment:
| Teacher | Student score |
|---|---|
| Bidirectional | 94.7 / 64.4 / 30.1 |
| Causal | 91.9 / 61.7 / 28.2 |
The same table shows the student initialized by ODE regression beats one without it: 94.7 / 64.4 / 30.1 against 93.4 / 60.6 / 29.4.
Conditions
- The result does not isolate causality. The causal teacher is also the weaker model before any distillation: 92.4 / 60.1 / 28.5 against the bidirectional teacher’s 94.6 / 62.7 / 29.6 at 100 steps (Table 4). The authors attribute the gap to the causal teacher’s error accumulation passing into its student (Fig. 8). That is an explanation, and the ablation cannot tell it apart from “the better teacher makes the better student”.
- One run per arm and no confidence intervals.
- The student is worse than its teacher on two axes. It “performs worse in temporal flickering and output diversity” (§5.2), and the diversity loss is “characteristic of reverse KL” (§6). The practice picks the teacher. It does not make distillation free. The diversity cost now has one measurement, from image distillation. sCM (LIT-790, Fig. 7) runs one-step VSD, which is DMD’s distribution-matching gradient without its regression loss, against two-step sCD on one EDM2-M backbone at ImageNet-512. As guidance rises VSD’s precision rises and its recall falls, ending in “severe mode collapse”, while sCD stays close to the teacher. It is curves with no tabled values. The figure also plots one-step sCD, and at guidance 1.0 its recall is about 0.70 against VSD’s 0.65, so part of the gap is the objective’s, at matched step count and before guidance is raised; guidance widens it. rCM (LIT-783) makes the same claim for video from samples only.
- Distribution matching is the distillation method throughout. DMD (LIT-643) and DMD2 (LIT-646) are its image-domain sources. Whether the result carries to other distillation objectives is untested.