# Few-Step Streaming Video Generation via Condition Shifting

**Authors:** Deyuan Liu, Jungang Li, Kaiwen Zheng, Kai Jiang, Jintao Chen, Yaole Wang, Ming Lin, Tao Lin, Jianfei Chen, Jun Zhu, Jintao Zhang  
**Lab:** Tsinghua University · **Published:** 10 Nov 2026 · **Upvotes:** 66 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.14033) · [PDF](https://arxiv.org/pdf/2611.14033) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.14033.html)

## Curator's notes

**TL;DR:** APEX removes the fake-score network for few-step streaming video generation by shifting conditions within the generator.

- **What's new:** It replaces the separate fake network with condition shifting, reducing memory and compute while keeping other regularizations. It also adds a one-step spatial Refiner.
- **Key result:** APEX improves instruction fulfillment (3.856 vs 3.831) and visual aesthetics (0.641898 vs 0.632606) over S2-Avatar, with 30% net human preference for instruction following.
- **Why it matters:** Builders can generate high‑quality streaming video faster and with less GPU memory.

## Key numbers

- **30.0%**: with +30.0% net preference
- **4.3%**: followed by consistency and video quality at +4.3% each
- **4.4%**: Separate measurements under matched training settings show 4.4% lower peak allocated GPU memory per GPU and…

## Abstract

Few-step sampling reduces the cost of streaming video generation. Distribution Matching Distillation (DMD) supervises few-step generators with a separate fake-score network, which adds training memory and computation and must track the evolving generator. We extend APEX to streaming video by using condition shifting to estimate the velocity field of the generated distribution within the generator. The original text condition selects generation, while a shifted condition selects fake estimation using the same parameters. This removes the independent fake network while retaining S2-Avatar's frozen reference teacher, Self-Replay Forcing, and perceptual regularization. We evaluate APEX on 180-second videos using three backbone evaluations per block and a one-step spatial Refiner, with a 70-case human study and a 320-case StreamAV-Bench comparison. Compared with S2-Avatar, APEX achieves higher video instruction fulfillment on the Progressive track (3.856 versus 3.831) and higher visual aesthetics across both tracks (0.641898 versus 0.632606). Human raters favor APEX most strongly on instruction following, with +30.0% net preference, followed by consistency and video quality at +4.3% each. Motion quality favors S2-Avatar, audiovisual synchronization is predominantly tied, and usability is unchanged. Separate measurements under matched training settings show 4.4% lower peak allocated GPU memory per GPU and 9.2% shorter mean iteration time than DMD.

## What you could build

1. Create a real‑time text‑to‑video chatbot using APEX.
2. Develop a low‑cost video summarization tool for long lectures.
3. Implement an interactive game cutscene generator with few‑step sampling.

## Quiz yourself

What component does APEX eliminate compared to Distribution Matching Distillation?

**Answer:** APEX eliminates the independent fake-score network, using the same generator parameters with a shifted condition for fake estimation.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.