Alice v1: Distilled Text-to-Video Generation

• Alice v1 is a text-to-video generation model that utilizes score-regularized consistency distillation to produce high-quality 5-second 720p videos in only 4 denoising steps.
• It employs a 14B dense diffusion transformer and a frozen 5.7B text encoder to improve VBench scores and reduce runtime compared to its teacher model Wan2.2.
• The model leverages an elaborate synthetic data pipeline, hard example mining, and a four-phase curriculum to enhance temporal coherence and overall motion realism.

Searching arXiv for the primary paper and closely related distillation / consistency-model context. Searching arXiv for "Alice v1 Distillation-Enhanced Video Generation consistency distillation score regularization". Alice v1 is a 14-billion-parameter open-source text-to-video generation model presented as a distilled student of Wan2.2, designed to generate 5-second, 720p, 24fps videos in 4 denoising steps. Its defining claim is that consistency distillation with score regularization, denoted rCM, can improve both speed and quality relative to the teacher, rather than trading one for the other. In the reported setting, Alice v1 improves VBench from 84.0 for Wan2.2 to 91.2 while reducing inference from 50 denoising steps to 4 and runtime from about 60 seconds to about 8 seconds on an H100 for a 5-second 720p clip (Xiaoyu et al., 27 Apr 2026).

1. Definition and scope

Alice v1 is described as a text-conditioned latent video diffusion/consistency model derived from Wan2.2. The system is open-source, video-only, and trained natively for 5-second clips. The paper positions it as an open model that aims to close, and on automated benchmarks exceed, the quality gap to closed-source systems such as Veo3 and Sora2, while remaining substantially faster than its 50-step teacher (Xiaoyu et al., 27 Apr 2026).

The model’s scope is specific. It targets silent text-to-video generation rather than multimodal video generation with audio. It is optimized around a short-horizon regime: native training is limited to 5-second clips, and although the paper states that sliding-window generation can extend beyond 30 seconds and even 60+, coherence degrades past 30 seconds. The paper also states that complex physics remains difficult, especially fluids, deformable objects, and multi-body interactions, and that in-video text rendering remains inconsistent (Xiaoyu et al., 27 Apr 2026).

Conceptually, Alice v1 is important because it is not framed as ordinary acceleration distillation. The paper argues that the student can exceed the teacher because the objective does not attempt to reproduce the teacher’s full variable-quality distribution. Instead, it combines a consistency objective with a score-based regularizer, and it trains on filtered synthetic outputs plus hard examples, so that the student preferentially learns the teacher’s stronger modes (Xiaoyu et al., 27 Apr 2026).

2. Model architecture and inference path

Alice v1 uses a frozen 5.7B-parameter umT5-XXL text encoder that supports prompts up to 512 tokens and over 100 languages. Video is generated in the latent space of a frozen 3D causal VAE that applies $8\times$ spatial and $4\times$ temporal compression into 16-channel latent codes. The denoiser is a 14B dense diffusion transformer with 40 blocks, hidden size 4096, 32 attention heads, MLP width 16384, $2\times 2$ latent patching, SiLU activations, and AdaLN timestep conditioning (Xiaoyu et al., 27 Apr 2026).

A notable architectural distinction from Wan2.2 is the replacement of the teacher’s Mixture-of-Experts denoiser with a single dense transformer. The paper argues that distillation effectively internalizes the expert specialization that Wan2.2 separated into high-noise and low-noise experts. This suggests that the student’s simplicity is not merely an implementation convenience; it is part of the authors’ account of how the distilled model remains fast while retaining broad generative competence (Xiaoyu et al., 27 Apr 2026).

At inference time, the model starts from latent Gaussian noise corresponding to a 5-second, 720p, 24fps video in the compressed VAE space. The prompt is encoded by umT5-XXL, and the student applies 4 denoising or consistency steps rather than integrating a 50-step diffusion trajectory. The latent output is then decoded by the frozen 3D causal VAE into RGB video. The paper does not provide the exact classifier-free guidance scale or scheduler specification, so those details remain unspecified in the text (Xiaoyu et al., 27 Apr 2026).

3. Score-regularized consistency distillation

The technical core of Alice v1 is score-regularized consistency distillation, or rCM. The teacher is a frozen pretrained video diffusion model, specifically Wan2.2. The student is trained to map noisy latent inputs toward clean outputs while remaining consistent along the teacher’s probability-flow trajectory and simultaneously being regularized by a score-distillation term evaluated on student-generated samples (Xiaoyu et al., 27 Apr 2026).

The paper begins from the diffusion noising relation

$$x_t = \alpha_t x_0 + \sigma_t \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I),$$

and uses a teacher-defined probability-flow ODE

$\frac{dx_t}{dt} = v_\phi(x_t,t).$

The student learns a consistency map

$F_\theta : (x_t,t) \mapsto x_0.$

The combined training objective is

$$\mathcal{L}_{\mathrm{rCM}}(\theta)=\mathcal{L}_{\mathrm{sCM}}(\theta)+\lambda\,\mathcal{L}_{\mathrm{DMD}}(\theta),$$

with main training using $\lambda = 0.1$ (Xiaoyu et al., 27 Apr 2026).

The paper interprets the two terms asymmetrically. $\mathcal{L}_{\mathrm{sCM}}$ is described as behaving like a forward-divergence, mode-covering term, whereas $\mathcal{L}_{\mathrm{DMD}}$ is described as a reverse-divergence-style, mode-seeking score regularizer. In the authors’ account, forward-only consistency training tends toward blur or mode averaging, while score regularization pushes student samples back toward high-density, higher-quality teacher regions. This is the basis for the paper’s claim that distillation can be quality-enhancing rather than merely quality-preserving (Xiaoyu et al., 27 Apr 2026).

The paper attributes Alice v1’s teacher-surpassing behavior to three mechanisms. First, score regularization acts as a mode-seeking objective that concentrates probability mass on the teacher’s higher-quality outputs. Second, the synthetic data pipeline focuses training on failure modes such as physics, hands, and faces. Third, consistency enforcement acts as implicit regularization by eliminating what the paper calls “lucky path” dependence on specific noise samples. The paper interprets this as a kind of trajectory invariance: the student is forced to produce stable clean predictions across points on the same teacher trajectory (Xiaoyu et al., 27 Apr 2026).

4. Synthetic data pipeline, hard example mining, and training curriculum

The training data pipeline is unusually elaborate. The authors generate 1 million prompts using GPT-4, with coverage over actions, subject types, environments, styles, physics scenarios, and edge cases such as multiple characters, complex camera motion, text rendering, and mirrors. Prompts are described as detailed, averaging 50–100 words. Wan2.2-14B is then run on all prompts with 50 denoising steps at 720p, 5 seconds, 24fps, requiring about 17,000 H100-hours and producing 1 million synthetic videos (Xiaoyu et al., 27 Apr 2026).

These synthetic videos are filtered by VBench-based criteria for temporal consistency, motion smoothness, and aesthetics, plus optical-flow analysis, CLIP aesthetic scoring, face-quality checks, and physics-plausibility checks. Only the top 30%, roughly 300,000 videos, are retained. The paper then adds hard example mining targeted at four failure modes: physics violations, hand quality, face consistency, and motion coherence. Prompts in hard categories are oversampled with $4\times$0 additional teacher samples. Both successful and failed examples are included, but failed examples are downweighted to $4\times$1 probability. The paper states that removing failures entirely reduces physics improvement by 40%, while equal weighting causes occasional failure imitation; a 1:5 failure:success ratio gave the best boundary signal between success and failure (Xiaoyu et al., 27 Apr 2026).

The final training mixture is 70% synthetic and 30% real. The real portion totals about 130,000 videos and includes approximately 40K stock footage videos from Pexels, Pixabay, and filtered WebVid; 35K video game recordings; 25K CGI/VFX clips; 15K robotics simulation videos; and 15K scientific visualization clips. The paper emphasizes the game footage as especially useful because it offers deterministic physics, perfect observability, and controllable variation (Xiaoyu et al., 27 Apr 2026).

Training is staged as a four-phase curriculum. Stage 1 runs for 100K steps at 480p and 3 seconds, with batch size 256, learning rate $4\times$2, synthetic data only, pure consistency loss, and target 8-step generation. Stage 2 runs 50K steps, scales from 480p to 720p, uses batch size 128 and learning rate $4\times$3, stays on synthetic data, and introduces full rCM with 4 target steps. Stage 3 runs 20K steps at 720p and 5 seconds, with batch size 64 and learning rate $4\times$4, using the 70:30 synthetic:real mixture plus LPIPS and VGG feature matching on top of rCM. Stage 4 runs 10K steps with batch size 32 and learning rate $4\times$5, and applies DPO using 10,000 human A/B preference pairs collected on Stage 3 outputs. The paper states that skipping Stage 1 causes divergence within 10K steps, introducing real data from Stage 1 lowers VBench by 2.1, and skipping Stage 4 leaves VBench nearly unchanged at 90.8 versus 91.2 but reduces human preference win rate by 8% (Xiaoyu et al., 27 Apr 2026).

5. Empirical performance and evaluation

The paper evaluates Alice v1 primarily with VBench and human preference studies. The headline comparison is summarized below.

Model	VBench	Steps / runtime
Wan2.2	84.0	50 / 60s
Sora2	$4\times$6	$4\times$7 / $4\times$8s
Veo3	$4\times$9	$2\times 2$0 / $2\times 2$1s
Alice v1	91.2	4 / 8s

The open-model comparison is direct and favorable to Alice v1, while the closed-model numbers are explicitly described by the paper as estimates from API outputs rather than fully controlled evaluations. This matters for interpretation: the Wan2.2 comparison is central and direct, whereas the Veo3 and Sora2 comparisons are indicative rather than definitive (Xiaoyu et al., 27 Apr 2026).

The detailed VBench breakdown shows improvements across all 16 reported dimensions, with the largest gain in Physical Plausibility, from 0.72 to 0.86, a $2\times 2$2 of +0.14. Other reported increases include Subject Consistency from 0.78 to 0.86, Human Actions from 0.75 to 0.84, Overall Consistency from 0.79 to 0.87, and Imaging Quality from 0.84 to 0.90. The paper presents these gains as evidence that the combination of rCM, quality filtering, hard example mining, and curriculum training improves not only speed but also temporal coherence and motion realism (Xiaoyu et al., 27 Apr 2026).

The step-count ablation is also important. Reported VBench values are 78.4 for 1 step, 85.2 for 2 steps, 91.2 for 4 steps, and 91.5 for 8 steps, with corresponding runtimes of 2, 4, 8, and 16 seconds. This indicates that 4 steps is the practical operating point: 8 steps yields only a marginal +0.3 VBench at double the runtime (Xiaoyu et al., 27 Apr 2026).

Human preference studies use 1,000 video pairs, 3 independent raters per pair, and 500 unique participants total. Alice v1 wins 68% of comparisons against Wan2.2 with 15% ties, 72% against HunyuanVideo with 14% ties, 51% against Veo3 with 22% ties, and 54% against Sora2 with 20% ties. The paper presents these results as decisive against open baselines and competitive against the closed systems included in the study (Xiaoyu et al., 27 Apr 2026).

6. Limitations, caveats, and significance

The paper is explicit about several limitations. Alice v1 is video-only and does not generate audio. Native training is limited to 5-second clips. Sliding-window generation can extend clip length, but coherence degrades past 30 seconds. Complex physics remains difficult, especially fluids, deformable objects, and multi-body interactions, and in-video text rendering is still inconsistent. These limitations bound the scope of the claimed improvements: Alice v1 is presented as a strong short-form video generator, not as a general solution to long-horizon world simulation (Xiaoyu et al., 27 Apr 2026).

The evaluation also carries caveats. Comparisons to Veo3 and Sora2 are not fully controlled, because their VBench scores and runtimes are estimated from API outputs. VBench is the dominant quantitative metric, but it is only one benchmark. The preference study is meaningful but limited in scale relative to deployment claims. In addition, the paper does not specify all inference details, notably the exact classifier-free guidance scale and scheduler. Safety considerations are described as not being deeply discussed, despite the fact that the release includes model weights, training code, synthetic data pipelines, and evaluation scripts (Xiaoyu et al., 27 Apr 2026).

Within those limits, Alice v1 is significant for two reasons. First, it advances an explicit methodological claim: consistency distillation with score regularization can be quality-enhancing when paired with filtered synthetic data and hard example mining. Second, it is presented as a complete open release, including weights and training infrastructure, which the paper frames as a contribution to open research in video generation. A plausible implication is that the work shifts attention from scaling teacher inference alone toward distillation objectives and data curation strategies that explicitly favor the teacher’s higher-quality modes over faithful reproduction of its full output distribution (Xiaoyu et al., 27 Apr 2026).