No Separate Fake Score Network Needed: Condition Shifting Enables Few-Step Streaming Video Generation

Condition shifting replaces the fake score network for lighter few-step streaming video generation.
The study proposes using "condition shifting" to estimate the generated distribution's velocity field inside the generator, removing the separate fake score network required by Distribution Matching Distillation (DMD). Extending APEX to streaming video generation, it uses the original text condition for the generation path and a shifted condition for the fake estimation path, sharing model parameters and retaining components like the frozen reference teacher, Self-Replay Forcing, and perceptual regularization. On 180-second videos, APEX slightly outperforms S2-Avatar in instruction satisfaction and visual aesthetics, with a +30.0% net preference for instruction following in human studies, though motion quality favors S2-Avatar. Training peak memory drops 4.4% and average iteration time shortens 9.2%, while engineering complexity also decreases.
Few-step sampling is a key path to reducing the cost of streaming video generation, but the mainstream Distribution Matching Distillation (DMD) method relies on a separate fake score network to supervise the few-step generator. This introduces extra training memory and compute overhead, and the fake score network must continuously track the generator's evolution. A new study proposes using "condition shifting" to estimate the velocity field of the generated distribution inside the generator itself, thereby eliminating the separate fake score network.

How Condition Shifting Replaces the Fake Score Network
The research extends the APEX method to the streaming video generation setting. The core idea is: use the original text condition to select the "generation" path, and use a shifted condition to select the "fake estimation" path, with both sharing the same set of model parameters. This means the generator itself can estimate the velocity field of the generated distribution, and there is no longer a need to separately train a fake score network.
This design retains components from S2-Avatar such as the frozen reference teacher, Self-Replay Forcing, and perceptual regularization; the changes are concentrated in the source of the supervision signal. For streaming video generation, training-stage memory pressure and iteration time often directly determine whether a solution can be deployed, so the benefits of removing the separate network are intuitive.
Distribution Matching Distillation (DMD) is a class of distillation methods that compress the generative capability of multi-step diffusion/flow models into a few-step generator. Its core is to make the output distribution of the few-step generator approximate the teacher model's distribution, and the key to achieving this is estimating the score of the "generated distribution" (i.e., the log-density gradient, or equivalently the flow-field velocity). DMD typically introduces an additional fake score network to fit the current generator's distribution, and trains it continuously as the generator updates. This auxiliary network not only consumes extra memory and compute, but must also evolve in sync with the generator; otherwise the supervision signal becomes distorted. The condition-shifting idea bypasses this auxiliary network: by feeding different conditions to the same model (original text condition vs. shifted condition), the model plays the roles of generator and fake score estimator under different condition contexts, thereby obtaining an estimate of the generated distribution's velocity field under parameter sharing.
Experimental Setup and Evaluation Scale
The study evaluates APEX on 180-second videos, with three backbone evaluations per block and a single-step spatial Refiner. The evaluation consists of two parts: a 70-case human study and a 320-case StreamAV-Bench comparison. This evaluation scale is relatively solid for streaming video generation research, including both human subjective ratings and systematic benchmark-level comparisons.
Streaming video generation refers to models that continuously generate video frames in an autoregressive or block-wise manner, rather than generating an entire video at once. It requires the model to handle long temporal dependencies under limited memory and maintain inter-frame consistency. In the text, "three backbone evaluations per block" means generation is divided into several blocks, and each block requires multiple calls to the backbone network to complete denoising or flow-field integration; the single-step spatial Refiner performs a lightweight correction of each frame's spatial details beyond the temporal dimension, to compensate for the quality loss caused by few-step sampling. StreamAV-Bench is a benchmark for streaming audio-video generation, covering dimensions such as instruction following, visual quality, motion quality, and audio-video synchronization; the 320-case comparison provides a systematic quantitative basis. Combining such benchmarks with human studies reflects both objective metrics and subjective perception.
Comparison Results with S2-Avatar
On the Progressive track, APEX achieves higher video instruction satisfaction than S2-Avatar (3.856 vs. 3.831); it also scores higher on visual aesthetics on both tracks (0.641898 vs. 0.632606). Human raters show the strongest preference for APEX on instruction following, with a net preference of +30.0%, and +4.3% each for consistency and video quality.
Disagreements also exist: S2-Avatar is preferred on motion quality, audio-video synchronization is mostly on par, and usability shows no change. This indicates that APEX's gains are not a comprehensive sweep, but are concentrated in instruction following and visual perception, while motion performance is a weakness that needs further observation.
S2-Avatar is a representative prior method for streaming video generation, employing training components such as a frozen reference teacher, Self-Replay Forcing, and perceptual regularization. Self-Replay Forcing means letting the model replay its own generated intermediate results and use them as conditions for continued generation, to alleviate the mismatch between training and inference distributions (exposure bias); the frozen reference teacher provides a stable supervision target, preventing the teacher from drifting during training. APEX retains these components and only replaces the source of the supervision signal from a separate fake score network with condition shifting, so the comparison results can relatively cleanly reflect the effect of "removing the fake score network." Understanding this helps judge that APEX's improvement in instruction following does not come from additional components, but from improved supervision signal quality and training stability.
Practical Benefits of Training Overhead
Measured separately under matched training settings, APEX reduces peak allocated memory per GPU by 4.4% and shortens average iteration time by 9.2% compared with DMD. The magnitude is not dramatic, but for streaming video tasks that require long training and continuous generator evolution, the dual savings in memory and time have cumulative effects. More importantly, removing the separate fake score network leaves the training pipeline with one fewer component that must evolve in sync with the generator, reducing engineering complexity.
Significance and Limitations of the Method
The condition-shifting idea essentially embeds the ability to "judge itself" into the generator, using the same set of parameters to play different roles under different condition contexts. This self-supervised-style design reduces dependence on auxiliary networks and provides a lighter-weight training path for few-step streaming video generation.
It should be noted that the current evaluation is concentrated on 180-second videos and a specific backbone, and the disadvantage in motion quality also suggests that the method may still have room for improvement in dynamic scenes. Whether condition shifting can be generalized to more generation architectures, and the issue of error accumulation in long-term streaming generation, are directions worth attention in future work.
One-line Summary
Condition shifting replaces the fake score network, enabling lighter-weight few-step streaming video generation.
Paragraph Summary
The study proposes using "condition shifting" to estimate the velocity field of the generated distribution inside the generator, thereby removing the separate fake score network that Distribution Matching Distillation (DMD) relies on. The method extends APEX to streaming video generation, using the original text condition to select the generation path and a shifted condition to select the fake estimation path, with both sharing model parameters, while retaining components such as the frozen reference teacher, Self-Replay Forcing, and perceptual regularization. On 180-second videos, APEX slightly outperforms S2-Avatar in instruction satisfaction and visual aesthetics, and the human study shows a +30.0% net preference for instruction following; however, motion quality favors S2-Avatar. On the training side, peak memory is reduced by 4.4% and average iteration time is shortened by 9.2%, with engineering complexity also decreasing.
Related articles

Gemini 4 Neon Leak: A Potential Strong Contender Against Opus 5.5
A Hacker News thread suggests Gemini 4 Neon looks like an Opus 5.5 competitor. This article unpacks the product positioning behind this community signal, shifts in the top-model competitive landscape, and what it means for developers' model selection strategy.

Claude Fable 5.2's Self-Authored Short Film 'Fable' Sparks Discussion
A short film 'Fable,' credited to Claude Fable 5.2 for script, visuals, and directing, appears on Hacker News. This article analyzes the significance of full-pipeline AI creation, community reactions, and what it means for content creators.

Anthropic IPO Pricing: A Signal of AI Giants Going Public
Anthropic's IPO pricing draws attention. This article analyzes its potential impact on AI industry valuation, commercialization paths, and the competitive landscape, and highlights public-market expectations and challenges for AI companies.