Anchor Forcing in Streaming Video Diffusion

• Anchor Forcing is a cache-centric framework for interactive streaming video diffusion that preserves semantic context and long-horizon motion during prompt switches.
• The method utilizes anchor memory and anchor-guided re-cache to integrate persistent sink, local, and junction caches, thereby improving prompt adherence and quality scores.
• Tri-region RoPE re-aligns bounded positional indices across sink, junction, and local caches, maintaining motion dynamics within a pretrained temporal regime.

Anchor Forcing is a cache-centric framework for interactive streaming video diffusion in which a long video is generated under prompt switching while preserving perceptual quality, prompt adherence, and long-horizon motion. In its canonical form, the method addresses two failure modes specific to interactive streaming generation: prompt-switch cache failure, in which ordinary re-cache procedures cannot simultaneously preserve prior semantic KV evidence and boundary-adjacent latent cues, and RoPE positional shift, in which unbounded streaming indices drift outside the pretrained temporal positional regime. The framework introduces two coupled mechanisms—anchor-guided re-cache based on anchor memory, and tri-region RoPE with RoPE re-alignment distillation—to stabilize post-switch conditioning and retain pretrained motion priors over long horizons (Yang et al., 12 Mar 2026).

1. Definition, scope, and problem setting

Anchor Forcing is defined for interactive long video generation with prompt switching in a streaming autoregressive video diffusion regime. A long video is partitioned into temporal segments, each associated with a prompt, and generation proceeds with causal temporal attention and rolling transformer KV caches rather than full-history dense attention. The method is built on a distilled streaming video diffusion pipeline initialized from CausVid, with a Wan2.1-T2V-1.3B backbone and Distribution Matching Distillation (DMD) as the distillation basis (Yang et al., 12 Mar 2026).

The motivating failure at prompt boundaries is cache-theoretic. Prior streaming systems support prompt switching by re-cache, but the paper states that these methods “rebuild the cache by evicting prior KV states,” thereby losing both semantic context from the previous stage and recent latent evidence near the switch boundary. The second failure is positional: naive streaming assigns temporal RoPE indices by a global frame count, but the backbone was initialized from a model trained with ODE trajectories capped at 21 latent frames. This yields a positional distribution shift that weakens long-horizon motion dynamics (Yang et al., 12 Mar 2026).

The term Anchor Forcing is specific in this literature. Several adjacent video-generation papers analyze anchor heads or global context anchors, but they do not define a standalone method with this name. In Head Forcing, the anchor-related idea is the treatment of anchor heads inside a broader head-wise KV-cache policy (Tian et al., 14 May 2026). In Pyramid Forcing, the closest analogue is the cache rule for Anchor Heads inside a head-aware pyramidal KV-cache framework (Chen et al., 13 May 2026). In Rolling Forcing, the anchor-like mechanism is the attention sink / global context anchor that preserves KV states of initial frames (Liu et al., 29 Sep 2025). By contrast, Anchor Forcing is itself a named method centered on anchor memory and tri-region RoPE (Yang et al., 12 Mar 2026).

2. Anchor memory and anchor-guided re-cache

The first core mechanism is anchor memory, a structured memory used to reduce evidence loss across prompt switches. The paper defines a prompt stream

$\mathcal{C}=\{c^{1},c^{2},\dots,c^{m}\}$

with interaction boundaries

$\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$

where $f_i$ is the frame index at which the prompt updates. For any frame $t$,

$$\pi(t)=\max\{\, i \mid f_i \le t,\ i\in[1,m-1] \,\}, \qquad s(t)=f_{\pi(t)}.$$

Here $\pi(t)$ indexes the active prompt segment and $s(t)$ is the most recent prompt-switch frame (Yang et al., 12 Mar 2026).

Anchor memory combines three cache regions. The sink cache $\mathcal{S}$ stores KV states from the earliest $N_S$ generated frames and acts as persistent global semantic context. The local cache is a rolling short-range memory over the most recent $W$ frames,

$\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$0

The junction cache stores prompt-boundary evidence. After each boundary at $\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$1, the first $\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$2 frames under the new prompt are stored as

$\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$3

These are the frames that first realize the new prompt and therefore encode the most fragile transition information (Yang et al., 12 Mar 2026).

The full anchor memory is

$\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$4

This activation rule avoids redundancy: while the first $\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$5 post-switch frames remain inside the local window, the junction cache is inactive; once they roll out of the local window, the junction cache is appended so that boundary evidence remains accessible (Yang et al., 12 Mar 2026).

This structure underlies anchor-guided re-cache. At a prompt switch, the model does not rebuild cache from historical latents alone. Instead, local-cache recomputation under the new prompt is warm-started from the previous segment’s anchor memory, preserving three distinct types of evidence at once: persistent sink semantics, recent local context, and boundary-adjacent interaction cues. A simple cache flush improves only slightly, whereas the anchor-guided re-cache variant yields substantially larger gains in CLIP score, dynamic degree, and quality score, which the paper attributes to explicit integration of sink, clip/junction, and local cache information (Yang et al., 12 Mar 2026).

3. Tri-region RoPE and bounded temporal geometry

The second core mechanism is tri-region RoPE, introduced to reconcile streaming generation with the backbone’s bounded pretrained positional regime. The paper sets

$\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$6

matching the temporal range seen during the upstream ODE-trained backbone. Rather than using a single unbounded global temporal index, Anchor Forcing assigns region-specific bounded indices to the sink, local, and junction regions of memory (Yang et al., 12 Mar 2026).

For the sink region, the indices are fixed: $\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$7 For the local region, the intent is piecewise. Before the streaming index reaches the pretrained ceiling, the local window uses ordinary temporal indexing; after that point, the local window is remapped into a fixed bounded band at the top of the pretrained range. The paper states this as a bounded indexing rule with $\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$8 for local key index $\mathcal{F}=\{f_{1},f_{2},\dots,f_{m-1}\},$9 once $f_i$0. For the junction region, the junction positions are placed immediately before the local window by an offset rule

$f_i$1

The result is a stable relative geometry in which sink, junction, and local caches each occupy a bounded positional zone (Yang et al., 12 Mar 2026).

The practical significance is not merely positional clipping. The paper’s ablation on RoPE variants shows that Bounded RoPE yields only a small increase in dynamic degree, whereas tr-RoPE raises dynamic degree from 26.37 to 74.83 and quality score from 78.92 to 81.72. This suggests that the gain arises from region-aware bounded geometry, not from clipping alone (Yang et al., 12 Mar 2026).

The paper further describes RoPE re-alignment distillation as the training-time adaptation by which the streaming model is long-tuned under this bounded tri-region indexing scheme. No separate closed-form $f_i$2 is given in the provided text; the available formulation remains the DMD objective, while the re-alignment effect is attributed to training under the same tri-region positional geometry that will later be used at inference (Yang et al., 12 Mar 2026).

4. Training and inference pipeline

Anchor Forcing is built on a Wan2.1-T2V-1.3B backbone, initialized with causal attention masking following CausVid, and trained with prompts from filtered and LLM-augmented VidProM. The training recipe follows the streaming long-tuning style associated with LongLive: DMD training, short-window attention with sink tokens, and streaming continuation over self-generated rollouts up to 60 seconds (Yang et al., 12 Mar 2026).

The denoiser is written as

$f_i$3

where $f_i$4 is the current noisy latent, $f_i$5 the diffusion time, $f_i$6 the prompt embedding, and $f_i$7 the assembled memory from cached KV states. The DMD gradient approximation used in the paper is

$f_i$8

The accessible text does not provide an additional explicit RoPE-specific loss beyond this distillation formulation (Yang et al., 12 Mar 2026).

At inference, generation begins under prompt $f_i$9 and maintains the sink cache $t$0, rolling local cache $t$1, and, when activated, the current segment’s junction cache. For each new frame or latent chunk, the model assembles

$t$2

applies tri-region RoPE, and denoises conditioned on the active prompt. At a prompt switch $t$3, the prompt changes to the next segment condition, anchor-guided re-cache is performed, the first $t$4 post-switch frames are stored as $t$5, and streaming continues under the updated prompt (Yang et al., 12 Mar 2026).

The principal hyperparameters reported for the streaming cache are: $t$6 The model generates 5-second videos at $t$7 resolution, and the streaming long-tuning rollout extends to 60 seconds. Training uses AdamW with learning rate $t$8 for the actor and $t$9 for the critic, on approximately 24 hours on 56 A800 GPUs (Yang et al., 12 Mar 2026).

5. Empirical performance and ablation evidence

The paper evaluates Anchor Forcing on an interactive benchmark derived from 100 narrative scripts, each containing six consecutive 10-second prompts for a total duration of 60 seconds. The principal long-horizon metrics are subject consistency, background consistency, motion smoothness, dynamic degree, aesthetic quality, imaging quality, quality score, and clip-wise semantic score computed using ViCLIP (Yang et al., 12 Mar 2026).

On this interactive benchmark, Anchor Forcing reports:

• Dynamic Degree: 73.00
• Quality Score: 82.25

against the compared streaming baselines:

• Infinity-RoPE: Dynamic Degree 35.17, Quality Score 79.98
• MemFlow: Dynamic Degree 45.47, Quality Score 79.35
• LongLive: Dynamic Degree 26.37, Quality Score 78.92

The clip-wise semantic scores for Anchor Forcing remain relatively stable over the full 60-second rollout:

• 0–10s: 24.48
• 10–20s: 24.35
• 20–30s: 23.50
• 30–40s: 23.53
• 40–50s: 23.61
• 50–60s: 23.63

The paper interprets this as reduced late-segment semantic degradation under prompt switching (Yang et al., 12 Mar 2026).

In non-interactive settings, the method also improves single-prompt streaming generation. On the 5s benchmark, it reports:

• Dynamic: 73.61
• Total: 83.99
• Quality: 84.84
• Semantic: 80.57

On the 30s benchmark, it reports:

• Total Score: 83.25
• Quality Score: 84.40
• Semantic Score: 78.67
• Dynamic Score: 79.17
• Imaging Quality: 69.37
• Temporal Style: 24.17

These results are presented as the highest overall VBench score among the compared streaming autoregressive baselines (Yang et al., 12 Mar 2026).

The component ablations are especially informative. Starting from the Baseline:

• CLIP: 23.58
• Dynamic Degree: 26.37
• Quality Score: 78.92

adding + tr-RoPE yields:

• CLIP: 23.71
• Dynamic Degree: 74.83
• Quality Score: 81.72

and adding + tr-RoPE + AR yields:

• CLIP: 23.85
• Dynamic Degree: 73.00
• Quality Score: 82.25

This supports the paper’s division of labor: tri-region RoPE drives the dominant motion improvement, while anchor-guided re-cache further improves quality and prompt adherence (Yang et al., 12 Mar 2026).

The replacement ablations reinforce the same point. For re-cache strategies:

• Baseline: 23.58 / 26.37 / 78.92
• Flush: 23.71 / 27.60 / 79.09
• AR: 24.22 / 50.73 / 79.78

For RoPE variants:

• Baseline: 23.58 / 26.37 / 78.92
• Bounded: 23.58 / 29.09 / 79.38
• tr-RoPE: 23.71 / 74.83 / 81.72

A plausible implication is that prompt-switch cache preservation and positional re-bounding solve different, complementary failure modes: AR stabilizes evidence at boundaries, whereas tri-region RoPE preserves motion priors over long horizons (Yang et al., 12 Mar 2026).

6. Terminological boundaries, misconceptions, and limitations

A recurring misconception is that Anchor Forcing denotes any video-generation method that uses an anchor-like memory. The literature is more specific. Rolling Forcing also preserves early-frame information through a global context anchor, but its primary mechanism is rolling-window joint denoising with an attention sink (Liu et al., 29 Sep 2025). Head Forcing treats anchor heads as one profiled attention-head category and assigns them a specific KV-cache policy, but it explicitly does not define a standalone method called Anchor Forcing (Tian et al., 14 May 2026). Pyramid Forcing similarly studies Anchor Heads inside a head-aware pyramidal KV-cache policy rather than a separate algorithm with this name (Chen et al., 13 May 2026). Anchor Forcing is therefore best understood as the specific cache-and-position redesign introduced for interactive streaming video diffusion with prompt switching, not as a generic label for all anchor-based video methods (Yang et al., 12 Mar 2026).

The method’s limitations are largely architectural and deployment-oriented. The paper implies extra memory and storage overhead from maintaining junction caches and anchor memory, extra re-cache complexity at prompt boundaries, and dependence on discrete prompt-switch structure, since junction caches are defined relative to explicit boundary frames rather than continuous editing signals. It also notes that hyperparameters such as $$\pi(t)=\max\{\, i \mid f_i \le t,\ i\in[1,m-1] \,\}, \qquad s(t)=f_{\pi(t)}.$$0, $$\pi(t)=\max\{\, i \mid f_i \le t,\ i\in[1,m-1] \,\}, \qquad s(t)=f_{\pi(t)}.$$1, $$\pi(t)=\max\{\, i \mid f_i \le t,\ i\in[1,m-1] \,\}, \qquad s(t)=f_{\pi(t)}.$$2, and $$\pi(t)=\max\{\, i \mid f_i \le t,\ i\in[1,m-1] \,\}, \qquad s(t)=f_{\pi(t)}.$$3 remain design choices, and that the benefits of tri-region RoPE rely on streaming long-tuning/distillation under the same bounded positional scheme used at inference (Yang et al., 12 Mar 2026).

A further point of interpretation concerns the word forcing. The paper does not formulate Anchor Forcing as a hard optimization constraint analogous to teacher forcing or classifier-free guidance. Instead, the forcing is cache-centric and positional: the method forces streaming inference to preserve prompt-boundary evidence through anchor memory and forces temporal attention to remain within a pretrained positional regime through tri-region RoPE. In that narrower and method-specific sense, Anchor Forcing is a redesign of memory assembly and temporal indexing for interactive long video generation (Yang et al., 12 Mar 2026).