Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

arXiv:2605.15141 · cs.CV · Submitted 2026-05-14 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation".

Tom: Real-time interactive video generation requires low-latency, streaming, and controllable rollout, and this paper addresses the limitations of existing autoregressive diffusion distillation methods by proposing Causal Forcing++,

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So, we’ve covered that Causal Forcing++ is a principled pipeline designed to solve the initialization bottleneck when generating videos with only one or two sampling steps, by using causal consistency distillation. Jane That means the paper argues that existing strategies are either misaligned with the causal rollout process, or they just aren't scalable enough for this aggressive regime.

Lu: The central claim of Causal Forcing++ is that it proposes a way to initialize a few-step AR student using causal consistency distillation, which aims to learn the same AR-conditional flow map as causal ODE distillation <ref:2605.15141#pg1>.

Meng: That sounds like it’s trying to bridge the gap between two different initialization techniques by focusing on what they both want to achieve: understanding the underlying flow map of the teacher model.

Lalam: It seems like a clever way to get supervision that is inherently local and real-time, which is exactly what we need when dealing with interactive video where we can't afford massive offline processing.

Tom: Exactly, and the authors highlight that causal ODE distillation requires generating full multi-step PF-ODE trajectories for every training sample, which they found to be very costly to scale <ref:2605.15141#pg1>.

Jane: And Causal Forcing++ avoids that cost by obtaining supervision from a single teacher ODE step between adjacent timesteps on real videos, avoiding the need to precompute and store those full trajectories <ref:2605.15141#pg2>.

Lu: The paper sets up the pipeline by inheriting Stage one and Stage three from previous work, but fundamentally changes Stage two by substituting initialization with Causal Consistency Distillation for few-step AR student initialization <ref:2605.15141#pg2>.

Meng: So the pipeline structure itself is mostly borrowed, but the crucial innovation lies in replacing that specific step with this more efficient causal CD method for getting the student started <ref:2605.15141#pg2>.

Lalam: This points to a design philosophy where we can retain the main training stages but inject a more intelligent, locally consistent way of setting up the student model for deployment <ref:2605.15141#pg2>.

Tom: The authors emphasize that Causal CD is a principled substitute for causal ODE initialization because it targets the same learning target—the AR-conditional flow map—but obtains supervision in a different, more manageable way <ref:2605.15141#pg1>.

Jane: And they quantify the efficiency gains by showing that this method reduces training cost by four times and requires no extra data curation or auxiliary storage, which is a significant practical improvement <ref:2605.15141#pg2>.

Lu: The comparison in Figure one shows that Causal Forcing++ outperforms Self Forcing and the Costly ODE initialization across several metrics like initialization time, cost, and latency <ref:2605.15141#pg2>.

Meng: That quantitative evidence is what we need; seeing a four times reduction in cost on top of achieving better VBench scores is concrete proof that this isn't just theoretical work <ref:2605.15141#pg2>.

Lalam: For me, the biggest implication here is that we can design AI systems that are inherently more efficient from the ground up, focusing on local consistency rather than relying on massive, expensive global trajectory calculations during training.

Conclusion: Tom: So, looking at the full picture, Causal Forcing++ isn't just another tweak; it’s a principled way to handle initialization that is both correct—it targets the right AR flow map—and more efficient because causal CD entirely sidesteps the need for offline trajectory generation <ref:2605.15141#pg2>.

Jane: I think the authors have successfully shown that we can achieve performance comparable to, and in some cases better than, previous state-of-the-art methods when running in the demanding frame-wise two-step setting <ref:2605.15141#pg2>.

Lu: The authors distinguished their work from prior research by being theoretically sound about targeting the correct AR flow map while avoiding that offline trajectory generation bottleneck, which is a key technical hurdle <ref:2605.15141#pg2>.

Meng: From an engineering standpoint, this suggests that future work in generative AI should prioritize objectives like Causal CD when designing models intended for real-time applications where latency and resource usage are primary concerns <ref:2605.15141#pg2>.

Lalam: The broader impact is that it lowers the barrier to entry for building sophisticated, high-quality, real-time interactive video systems by making the initialization process significantly more efficient and scalable <ref:2605.15141#pg2>.

Tom: So in short, Causal Forcing++ delivers strong results under aggressive low-step conditions while simultaneously slashing latency by fifty percent compared to earlier approaches <ref:2605.15141#pg2>.

Jane: It’s a significant step forward because it proves that we can get high fidelity in a very constrained, real-time environment without incurring the massive computational penalty of offline trajectory precomputation <ref:2605.15141#pg2>.

Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou, Bokai Yan, Xinyuan Li, Xiao Yang, Chongxuan Li, Jun Zhu

Tsinghua University · ShengShu · Renmin University of China

cs.CV

Submitted: 2026-05-14

Updated: 2026-10-04

Comments: fixed typos

Code: https://github.com/thu-ml/Causal-Forcing

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 82/100

The gist: Causal Forcing++ replaces causal ODE distillation with causal consistency distillation to initialize a few-step autoregressive student, enabling frame-wise 2-step video generation that beats prior

Key concepts

Causal Consistency Distillation (Causal CD): This is the core innovation. Instead of requiring expensive precomputed full trajectories, Causal CD learns the same AR-conditional flow map as causal ODE distillation but gets supervision from just one teacher ODE step between adjacent video frames. This makes initialization both more efficient and easier to optimize.
Few-step AR Students: These are the small autoregressive models that are initialized using Causal CD. They are used in real-time generation where only a few steps (like 1 or 2) of the diffusion process are needed to produce an output, making them crucial for low-latency video tasks.
AR Diffusion Distillation: This is the overall training strategy. It involves using a larger, more powerful teacher model to guide and train a smaller student model. The goal is for the student to learn the same complex patterns as the teacher in an autoregressive diffusion setting, which is vital for high-quality video generation.
Asymmetric DMD: This refers to Stage 3 of the pipeline, where asymmetric Discrete Moving Difference (DMD) is used with self-rollout. In this stage, both the teacher and critic models remain bidirectional, allowing for effective training while maintaining temporal consistency in the final generation phase.

Terminology

Distillation — 89× in this paper · explained in 2 episodes
Diffusion Model — 34× in this paper · explained in 8 episodes
Self Forcing — 31× in this paper
Consistency — 30× in this paper · explained in 3 episodes
Diffusion Models — 15× in this paper · explained in 6 episodes
World Model — 15× in this paper · explained in 7 episodes
Consistency Distillation — 11× in this paper · explained in 2 episodes
Action-Conditioned World Model — 10× in this paper · explained in 1 episode
Exposure Bias — 10× in this paper · explained in 2 episodes
Ablation — 6× in this paper · explained in 1 episode
Transformer — 6× in this paper · explained in 2 episodes
Benchmark — 4× in this paper · explained in 2 episodes
Generative Model — 4× in this paper · explained in 2 episodes
Core — 3× in this paper · explained in 2 episodes
Diffusion Process — 3× in this paper · explained in 2 episodes
Distribution Matching — 3× in this paper · explained in 2 episodes
Flow Matching — 3× in this paper · explained in 8 episodes
Generative Modeling — 3× in this paper · explained in 3 episodes
Generative Models — 3× in this paper · explained in 2 episodes
Ablation Studies — 2× in this paper · explained in 1 episode
Adaptation — 2× in this paper · explained in 1 episode
KV cache — 2× in this paper · explained in 3 episodes
Ablation study — 1× in this paper · explained in 1 episode
Aggregation — 1× in this paper · explained in 1 episode
Closed-Loop Control — 1× in this paper · explained in 3 episodes
EMMA — 1× in this paper · explained in 2 episodes
Quantization — 1× in this paper · explained in 12 episodes
Rectified Flow — 1× in this paper · explained in 3 episodes
Reinforcement Learning — 1× in this paper · explained in 4 episodes
Self-Distillation — 1× in this paper · explained in 7 episodes
Sparse Attention — 1× in this paper · explained in 5 episodes
Two-Stage Framework — 1× in this paper · explained in 2 episodes

Summary

Causal Forcing++ replaces causal ODE distillation with causal consistency distillation to initialize a few-step autoregressive student, enabling frame-wise 2-step video generation that beats prior 4-step chunk-wise methods while halving first-frame latency and cutting Stage 2 training cost roughly 4×. The work targets real-time interactive video generation, where low latency, streaming rollout, and user-controllable interaction are essential <ref:2605.15141#pg1,low latency, streaming rollout, and user-controllable interaction are essential>. Prior autoregressive distillation methods typically rely on chunk-wise autoregression with 4-step sampling, which still falls short of real-time interaction due to coarse response granularity and non-negligible sampling latency <ref:2605.15141#pg1,chunk-wise autoregression with 4-step sampling, which still falls short of real-time interaction>. The authors therefore push AR diffusion distillation to a more aggressive and largely underexplored regime: frame-wise autoregression with only 1–2 sampling steps <ref:2605.15141#pg1,frame-wise autoregression with only 1–2 sampling steps>.

The initialization bottleneck

The paper identifies the initialization of a few-step AR student before asymmetric DMD as the key bottleneck in this aggressive regime <ref:2605.15141#pg1,the initialization of a few-step AR student before asymmetric DMD as the key bottleneck>. Three existing options fail in complementary ways:

ODE initialization with a bidirectional teacher, as in CausVid and Self Forcing, is architecturally misaligned with causal rollout: the teacher trajectory depends on future frames that are unavailable to an AR student <ref:2605.15141#pg1,the teacher trajectory depends on future frames that are unavailable to an AR student>.
Using a multi-step AR diffusion model directly, as in LiveAvatar and WorldPlay, avoids this mismatch but lacks few-step generation capability <ref:2605.15141#pg2,avoids this mismatch but lacks few-step generation capability>.
Causal ODE initialization, as in Causal Forcing, corrects the learning target by distilling from an AR teacher, but requires generating full multi-step PF-ODE trajectories for every training sample, making it costly to scale <ref:2605.15141#pg2,requires generating full multi-step PF-ODE trajectories for every training sample>.

How it works

The key observation is that causal ODE distillation and causal CD aim to learn the same object: the AR-conditional flow map (or namely the consistency function) of the teacher <ref:2605.15141#pg2,causal ODE distillation and causal CD aim to learn the same object>. They differ in supervision: ODE distillation needs full offline trajectories, whereas causal CD obtains supervision from a single online teacher ODE step between adjacent timesteps on real videos <ref:2605.15141#pg2,a single online teacher ODE step between adjacent timesteps on real videos>. The paper proves the optimal consistency model matches the teacher's flow map up to solver error, bounded as O((∆t) p) <ref:2605.15141#pg6,where ∆t denotes the maximum difference between adjacent timesteps>. Beyond efficiency, adjacent-timestep consistency yields a smaller per-step optimization gap than causal ODE distillation, which regresses noisy intermediate states directly to clean endpoints <ref:2605.15141#pg2,adjacent-timestep consistency yields a smaller per-step optimization gap than causal ODE distillation>. The pipeline keeps Stage 1 (teacher forcing AR diffusion training) and Stage 3 (asymmetric DMD with self-rollout) from Causal Forcing and replaces Stage 2 with causal CD <ref:2605.15141#pg7,replaces its Stage 2 with the causal CD objective of Eq. (3)>.

Results

On Wan2.1-1.3B at 480 × 832 with 81 frames, the frame-wise 2-step model reaches a VBench Total of 84.14 and Quality of 84.89, surpassing CausVid (81.33), Self Forcing (83.74), and Causal Forcing (84.04), while improving throughput by about 1.4× and reduces latency by 50% <ref:2605.15141#pg11,improves throughput by about 1.4× and reduces latency by 50%>. Ablations show causal CD matches or surpasses causal ODE across all step settings while dramatically reducing Stage 2 cost <ref:2605.15141#pg12,matches or surpasses causal ODE across all step settings>, cutting cost from roughly 11,600 to 2,900 A800 GPU-hours and eliminating 1,900 GiB of trajectory storage <ref:2605.15141#pg6,reduces the data curation and training cost from ∼11,600 to ∼2,900 A800-GPU hours>.

Why not score distillation

Causal DMD, a teacher-forcing score-distillation variant, produces sharper early frames but suffers from substantially stronger exposure bias and is therefore unsuitable as a few-step initialization <ref:2605.15141#pg9,suffers from substantially stronger exposure bias and is therefore unsuitable as a few-step initialization>. Because reverse-KL is mode-seeking, once the shifted mode moves into the poor-quality region, most generated samples tend to follow this erroneous mode <ref:2605.15141#pg9,once the shifted mode moves into the poor-quality region>, whereas mode-covering CD retains mass in good-quality regions.

World-model extension

The pipeline also distills a camera-pose-conditioned generator, injected via PRoPE, into an interactive action-conditioned world model in the spirit of Genie3 <ref:2605.15141#pg7,distilling this bidirectional model with Causal Forcing++ into an interactive action-conditioned world model>.

Improvements for AI systems

Replace Stage 2 causal ODE trajectory curation with online causal CD: Swap the offline 48-step PF-ODE trajectory generation in Causal Forcing for the paper's causal CD objective, which obtains supervision from a single online teacher ODE step between adjacent timesteps on real videos, cutting Stage 2 from ∼11,600 to ∼2,900 A800-GPU hours (∼4× speedup) and auxiliary storage from ∼1,900GiB to zero.
Use adjacent-timestep local pairing instead of t→0 regression: Train the few-step AR student by pairing x t i with a single Euler step to x̂ t−Δt i under the AR teacher conditioned on ground-truth prefix, since causal CD reduc[es] the per-step gap to Δt and yield[s] an easier optimization target, avoiding Causal ODE's large jumps from t to 0.
Adopt frame-wise 2-step AR generation with the ASD first-frame trick: Generate frame-wise with 2 steps (t = 1, 0.8333) for the 20 subsequent latent frames while keeping the first latent frame at 4 steps, per ASD, delivering 0.27 s first-frame latency and 14.1 FPS versus 0.60 s / 10.4 FPS for chunk-wise 4-step baselines.
Reject causal DMD (score-distillation) initialization in favor of causal CD: Do not initialize via teacher-forcing DMD, because mode-seeking DMD is more sensitive to accumulated history errors, causing exposure bias to rapidly amplify in later frames, whereas mode-covering CD maintains a more dispersed distribution and keeps mass in good-quality regions under history shift.
Reject Self-Forcing-style ODE init from a bidirectional teacher: Never initialize the AR student from bidirectional-teacher ODE trajectories, which violate[] the frame-level injectivity required by an AR student and "collapse[] toward the conditional expectation E[x 0 i x t i, x t<i, t]," producing VBench Total below 80 and Dynamic Degree ≤ 2 in frame-wise settings.
Reject direct multi-step AR diffusion initialization: Do not skip few-step adaptation, since asymmetric DMD acts as a refiner rather than a complete trainer; direct AR init collapses at 1-step to 0 Dynamic Degree, 1.101 VisionReward, and even a -14 Instruction Following score.
Keep Stage 1 teacher-forcing AR training and Stage 3 asymmetric DMD with self-rollout unchanged: Inherit both stages from Causal Forcing, using a Wan2.1-1.3B AR teacher/student and a Wan2.1-14B real-score model, so the pipeline stays within the asymmetric-DMD framework and supports temporal KV-cache inference.
Extend to action-conditioned world modeling via PRoPE camera-pose conditioning: Build a camera-pose-annotated dataset with WorldPlay, finetune Wan2.1-1.3B into a bidirectional pose-conditioned model by injecting pose via PRoPE, then distill with Causal Forcing++ into an interactive AR world model supporting actions like moving forward continuously and move forward first, then tilt the camera downward.
Select the step count per latency budget using the reported trade-off: Choose 2-step for best VBench Total (84.14) and VisionReward (6.661), 4-step for best Quality (84.94) and Dynamic Degree (71), or 1-step for maximum throughput (20.7 FPS), noting all three share the same 0.27 s first-frame latency due to the ASD trick.

Abstract

Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose Causal Forcing++, a principled and scalable pipeline that uses causal consistency distillation (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline,, surpasses the SOTA 4-step chunk-wise Causal Forcing under the frame-wise 2-step setting by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50% and Stage 2 training cost by about 4 times. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM.