Alice v1: Distillation-Enhanced Video Generation Surpassing Closed-Source Models

Abstract: Wepresent Alice v1, a 14-billion parameter open-source video generation model that achieves state-of-the-art quality through consistency distillation with score regularization (rCM). Contrary to conventional distillation-which trades quality for speed-we demonstrate that rCM-based distillation can exceed teacher model quality. We attribute this to three mechanisms: (1) the score regularization term acts as a mode-seeking objective that concentrates probability mass on high-quality outputs rather than covering the full teacher distribution, (2) our targeted synthetic data pipeline with hard example mining provides training signal specifically for failure modes (physics, hands, faces) that the teacher handles inconsistently, and (3) consistency enforcement acts as implicit regularization, eliminating "lucky path" dependence on specific noise samples. Alice v1 generates 5-second 720p videos at 24fps in 4 denoising steps (~8 seconds on H100), a 7x speedup over the 50-step teacher while improving VBench score from 84.0 (Wan2.2) to 91.2. This surpasses both the teacher and closed-source systems including Veo3 (~90) and Sora2 (~88) on automated benchmarks, with competitive results in human preference studies. We release all model weights, training code, synthetic data pipelines, and evaluation scripts to advance open research in video generation.

• The paper introduces Alice v1, a 14-billion-parameter open-source video model that combines continuous-time consistency distillation with mode-seeking score regularization to achieve 91.2 VBench, surpassing its Wan2.2 teacher’s 84.0 score.
• The method generates five-second, 720p clips in four denoising steps—about eight seconds on an H100—while targeted synthetic-data filtering and hard-example mining improve physics, hand, and facial consistency.
• The paper shows that staged training, curated teacher outputs, and a custom FlashAttention-2 JVP kernel make high-quality few-step generation practical, though comparisons with closed-source models remain uncertain because their scores are estimated.

Overview

Alice v1 is a 14-billion parameter open-source text-to-video model built on the Wan2.2 architecture and trained via score-regularized continuous-time consistency distillation (rCM). The paper's central claim is that distillation, conventionally treated as a quality-for-speed trade-off, can instead improve generation quality when combined with mode-seeking score regularization, targeted synthetic data curation, and consistency enforcement. The reported headline result is a VBench score of 91.2 from a Wan2.2 teacher scoring 84.0, obtained at 4 denoising steps (~8 seconds for a 5-second 720p clip on an H100), a 7× speedup over the 50-step teacher (2605.08115). The authors further claim this exceeds estimated closed-source scores for Veo3 (~90) and Sora2 (~88), though those figures are approximations derived from API outputs rather than controlled evaluations.

Architecture

Alice v1 retains the Wan2.2 component stack: a frozen umT5-XXL text encoder (5.7B parameters), a frozen causal 3D VAE (8× spatial, 4× temporal downsampling to 16-channel latents), and a dense 14B diffusion transformer with 40 blocks, hidden dimension 4096, and AdaLN timestep conditioning. Two architectural departures from the teacher are notable. First, the MoE design of Wan2.2—separate high-noise and low-noise expert networks—is replaced by a single dense transformer; the authors report that distillation internalizes the expert specialization, suggesting MoE structure matters more for training stability than inference quality. Second, because the architecture matches the teacher exactly, comparisons are directly attributable to the training objective rather than capacity differences.

Training objective: rCM

The method combines two complementary losses. The continuous-time consistency loss ($\mathcal{L}_{\text{sCM}}$) enforces that $F_\theta(x_t, t)$ maps any point on a probability-flow ODE trajectory to the same clean prediction, using an EMA target and a Jacobian-vector product (JVP) through the student. The score distillation term ($\mathcal{L}_{\text{DMD}}$) evaluates student-generated samples against the frozen teacher's score function, implementing a reverse-KL, mode-seeking pressure.

The paper frames the complementarity in terms of divergence behavior: sCM alone is mode-covering and produces blurry averages; DMD alone is mode-seeking but unstable and prone to collapse. Ablations support this strongly: consistency-only reaches 82.1 VBench (below the teacher's 84.0), DMD-only reaches 83.5 with instability, while the combination reaches 91.2. The balancing weight $\lambda = 0.1$ was selected by ablation, with $\lambda > 0.3$ causing prompt-specific collapse and $\lambda < 0.05$ producing blur.

A practical contribution is a custom FlashAttention-2 JVP kernel enabling forward-mode differentiation at ~1.8× overhead versus ~3× for naive computation, which the authors state is what makes sCM training feasible on 10B+ video models.

Why the student can exceed the teacher

The paper offers three mechanisms for the counterintuitive quality gain:

1. Mode-seeking concentration: reverse KL penalizes probability mass where the teacher assigns low density, so combined with retaining only the top 30% of teacher generations, the student learns an implicit "best-of-$n$" policy distilled into weights.
2. Hard example mining: systematic failure categories (physics, hands, faces) are oversampled 5×, with failures included at 0.2× weight to provide contrastive signal about success/failure boundaries.
3. Consistency as regularization: trajectory-invariant predictions eliminate dependence on "lucky" noise samples, interpreted as implicit ensemble distillation toward the consensus output.

The ablation evidence is consistent with each mechanism contributing: excluding failures entirely reduces physics improvement by 40%, and equal failure weighting causes occasional reproduction of failure patterns. However, these mechanisms are presented largely as post-hoc interpretations; the paper does not isolate them experimentally beyond the aggregate loss and data ablations, so the relative attribution among the three remains an assumption supported indirectly.

Synthetic data pipeline

The pipeline generates 1M prompts via GPT-4 across actions, subjects, environments, physics scenarios, and edge cases, then runs the teacher at 50 steps/720p (~17,000 H100-hours). Multi-stage filtering (VBench scoring, optical flow plausibility, CLIP aesthetics, face consistency, physics flagging) retains the top 30% (~300K videos). Hard example mining uses four automated detectors—a ViT physics classifier (threshold 0.7), a MediaPipe-based hand scorer (0.6), ArcFace temporal identity drift (cosine distance > 0.3), and RAFT flow anomaly detection—with thresholds calibrated against human judgments on validation sets.

Real data (~130K videos) is mixed at a 70:30 synthetic-to-real ratio, deliberately inverting the ~62% real "golden ratio" reported for self-consuming training. The justification—that distillation draws synthetic data from a capable teacher rather than the student itself, avoiding recursive degradation—is plausible but rests on the assumption that teacher artifacts do not accumulate through the filtered subset. Notably, ~35K real videos are gameplay footage from physics-rich games (BeamNG.drive, Teardown, GTA V, sports simulators), motivated by ground-truth engine physics with no observability gaps. The authors report that physics understanding transfers to real-world prompts despite the domain gap, though this transfer claim is asserted empirically rather than analyzed mechanistically.

Data curation ablations show clear monotonic gains: unfiltered synthetic achieves 86.3 VBench, quality filtering adds +2.4, hard example mining adds another +2.5, with Physical Plausibility improving disproportionately (+0.14 over the teacher).

Training protocol

Training proceeds in four stages: (1) pure sCM at 480p/3s for 100K steps to establish stable few-step mapping before introducing DMD gradients; (2) full rCM with resolution scaling to 720p; (3) introduction of the 70:30 mixture plus perceptual losses; (4) DPO on 10K human preference pairs. Stage ordering is validated by ablation: skipping Stage 1 diverges within 10K steps, introducing real data early costs 2.1 VBench points, and omitting DPO leaves VBench nearly unchanged (90.8) but drops human win rate by 8%—indicating DPO affects perceptual alignment rather than benchmarked quality. Total compute is approximately 40,000 H100-hours, roughly 10× cheaper than training Wan2.2 from scratch.

Results

Model	Params	VBench	Steps	Time (H100)	Open
Mochi	10B	80.2	50	55s	✓
HunyuanVideo	8.3B	82.5	50	45s	✓
Wan2.2 (teacher)	14B active	84.0	50	60s	✓
Sora2 (est.)	~30B	~88	~30	~20s	✗
Veo3 (est.)	—	~90	~30	~15s	✗
Alice v1	14B	91.2	4	8s	✓

Human preference studies (1,000 pairs, 3 raters each) show 68–72% win rates against open-source baselines and 51–54% against Veo3 and Sora2 APIs. The step-count ablation shows diminishing returns beyond 4 steps (91.2 → 91.5 at 8 steps), confirming the few-step operating point. Speedups of roughly 7× hold consistently across H100, A100, and consumer GPUs.

Two caveats bear directly on the closed-source comparison: the Sora2 and Veo3 scores are estimates from API outputs under unspecified prompting conditions, and VBench is an automated metric known to correlate imperfectly with perceptual quality. The human evaluation partially mitigates this, but the margin over Veo3 (51% win rate with 22% ties) is within the range where evaluation methodology could matter.

Limitations and open questions

The paper concedes several limitations. Audio-visual joint generation is absent, unlike Veo3. Native training on 5-second clips limits long-form coherence; sliding-window inference extends to 60+ seconds but degrades beyond 30 seconds. Complex physics (fluids, deformables, multi-body interactions) remains difficult despite the +0.14 Physical Plausibility gain, and in-video text rendering is inconsistent. Beyond these stated limitations, three questions remain open: whether the mode-seeking-over-teacher-distribution argument holds for teachers with lower output variance than Wan2.2; whether the 70:30 synthetic ratio generalizes beyond this particular teacher-data regime; and whether the claimed superiority over closed-source systems survives standardized, blind evaluation protocols.

Conclusion

Alice v1 demonstrates that rCM-based distillation, combined with curated synthetic data and staged training, can raise a student above its teacher on both automated benchmarks and human preference—here by 7.2 VBench points at a 7× inference speedup and roughly one-tenth the teacher's training compute. The strongest empirical support lies in the loss-combination and data-curation ablations; the mechanism-level explanations are coherent but not individually isolated. The principal evidentiary weakness is the comparison against closed-source systems, which relies on estimated scores. The release of weights, code, pipelines, and evaluation scripts makes the central claims reproducible and testable by the community.