Real-time talking avatar generation requires low latency and minute-level temporal stability. Autoregressive (AR) forcing enables streaming inference but suffers from exposure bias, which causes errors to accumulate and become irreversible over long rollouts. In contrast, full-sequence diffusion transformers mitigate drift but remain computationally prohibitive for real-time long-form synthesis. We present AvatarForcing, a one-step streaming diffusion framework that denoises a fixed local-future window with heterogeneous noise levels and emits one clean block per step under constant per-step cost. To stabilize unbounded streams, the method introduces dual-anchor temporal forcing: a style anchor that re-indexes RoPE to maintain a fixed relative position with respect to the active window and applies anchor-audio zero-padding, and a temporal anchor that reuses recently emitted clean blocks to ensure smooth transitions. Real-time one-step inference is enabled by two-stage streaming distillation with offline ODE backfill and distribution matching. Experiments on standard benchmarks and a new 400-video long-form benchmark show strong visual quality and lip synchronization at 34 ms/frame using a 1.3B-parameter student model for realtime streaming. Our page is available at: https://cuiliyuan121.github.io/AvatarForcing/
Figures & tables
Figure 2: Windowed denoising with dual anchors and two-stage distillation. AvatarForcing performs one-step denoising over a fixed local-future window with heterogeneous timesteps (cleaner on the left and noisier on the right). At each step, the window is jointly updated under bidirectional attention to emit the leftmost clean block, slide the window, and append fresh noise. Long-horizon stability is supported by a dual-anchor KV cache, including a RoPE re-indexed style anchor (with anchor-audio zero-padding) and a temporal anchor constructed from recent clean blocks. Real-time one-step inference is achieved by distilling a global bidirectional teacher into a streaming student via two-stage training with offline ODE backfill and DMD post-training on student rollouts.
Figure 3: Long-form qualitative comparison. We compare AvatarForcing with representative autoregressive and diffusion-based talking-avatar models on long-form generation, with an emphasis on temporal stability, identity preservation, and audio–visual synchronization.
Model
FID ↓
FVD ↓
CSIM ↑
Sync-C ↑
Sync-D ↓
Latency (s/frame) ↓
EchoMimic [ 42 ]
63.26 / 147.13
1115.86 / 2102.42
0.69 / 0.79
4.72 / 3.44
9.55 / 10.91
2.25
Hallo3 [ 43 ]
47.40 / 122.64
488.50 / 1371.98
0.85 / 0.83
2.67 / 3.90
10.29 / 10.42
9.60
HunyuanAvatar [ 44 ]
43.42 / 76.58
445.02 / 870.16
0.87 / 0.91
4.92 / 2.71
10.01 / 9.34
14.40
FantasyTalking [ 20 ]
43.14 / 144.71
483.11 / 1649.23
0.85 / 0.76
3.15 / 4.30
9.69 / 13.18
4.20
OmniAvatar [ 45 ]
41.48 / 81.12
330.76 / 1430.79
0.89 / 0.86
5.58 / 4.21
9.01 / 10.82
6.21
MultiTalk [ 46 ]
44.31 / 106.04
419.19 / 1788.19
0.88 / 0.76
5.88 / 5.26
9.58/ 9.39
3.75
Table 1: Quantitative comparison on CelebV-HQ and long-form videos. Values are reported as CelebV-HQ / long-form, except latency (s/frame). CelebV-HQ clips average 5 seconds and long-form videos average 2 minutes. Best and second-best long-form results are highlighted in deep and light green. For pipelined multi-GPU systems (e.g., LiveAvatar [ 40 ] ), latency (s/frame) denotes the steady-state output interval (approximately 1/FPS ).
Figure 4: Latency vs. window length. Inference latency increases with both window length and the number of denoising steps. Window length is shown in frames; with block size B=4 , a window of L blocks spans BL frames.
Quantity
Definition
Default
Steady-state latency
Per-frame wall-clock time in steady state (audio encoder + DiT + KV update + VAE); excludes initialization, data loading, disk I/O, and post-processing.
Measured wall-clock delay between receiving the audio embedding for frame t and emitting the aligned video frame t in a real-time streaming simulation (includes look-ahead buffering and steady-state compute).
0.51 s
Emission unit
One clean block per step ( B frames); all timings are normalized per frame.
B=4
Table 2: Steady-state latency vs. end-to-end delay protocol. Default values use L=4 , B=4 , and 25 FPS.
Figure 5: Window length L vs. denoising steps N . Ablation on window lengths L and denoising steps N for BL,N .
Denoising step N\ block length L
1
2
4
8
1
0.07/0.10/31.4
0.73/3.23/19.5
0.90/1.06/34.14
0.91/0.84/59.67
2
0.52/8.12/18.66
0.83/2.25/33.19
0.80/1.94/69.45
0.75/0.72/119.32
4
0.32/8.12/28.19
0.87/2.23/78.94
0.90/0.75/166.42
0.82/0.83/284.79
8
0.74/2.24/38.2
0.91/1.10/153.07
0.81/1.20/321.98
0.84/0.87/545.62
Table 3: Grid ablation of window length L and denoising steps N . Entries report CSIM ↑ / Adj-LPIPS ↓ / latency (ms) ↓ ; Adj-LPIPS is reported in 10−2 . Cells with the same compute budget L⋅N share the same background color (blue gradient).
Model
Latency (s/frame) ↓
CSIM ↑
ΔE2000↓
Adj-LPIPS ↓
Teacher
16
0.88/0.84
18.51/24.18
0.18/0.28
Student
4.8
0.86/0.84
17.77/25.86
0.09/0.15
AvatarForcing
0.034
0.90/0.86
8.71/11.39
0.0106/0.0189
Table 4: Ablation on distillation backbone and anchors. Left: Teacher/Stud- ent/ Ours. Right: removing either anchor. Latency is reported in s/frame. Long-form stability metrics are reported as Total / Last over 1-minute rollouts.
Variant
FID ↓
FVD ↓
CSIM ↑
Sync-C ↑
Sync-D ↓
ΔE2000↓
Adj-LPIPS ↓
w/o anchor-audio zero padding
58.53
796.35
0.89/0.84
3.43
11.23
8.91/13.87
0.0110/0.0191
w/o RoPE re-indexing
72.43
1382.63
0.67/0.49
2.23
19.23
13.56/33.86
0.018/0.032
Full
56.22
737.92
0.90/0.86
5.64
9.26
8.71/11.39
0.0106/0.0189
Table 5: Anchor-audio zero padding and RoPE re-indexing. Ablation on anchor-audio zero padding and RoPE re-indexing. Metrics follow Tab. 6 .
Method
FID ↓
FVD ↓
CSIM ↑
Sync-C ↑
Sync-D ↓
Latency (s/frame) ↓
ΔE2000↓
Adj-LPIPS ↓
Self-Forcing (1-step) [ 17 ]
88.52
1127.43
0.40/0.07
1.34
19.53
0.031
27.91/40.63
0.05/0.10
Causal ODE [ 33 ]
102.63
1382.63
0.58/0.32
0.12
24.21
0.029
16.73/35.92
0.08/0.12
AvatarForcing (Ours)
56.22
737.92
0.90/0.86
5.64
9.26
0.034
8.71/11.39
0.0106/0.0189
Table 6: One-step causal baselines on long-form videos. Baselines are adapted to one-step inference. Note that Causal Forcing is the cascade of Causal ODE initialization and DMD; here we only use Causal ODE as a baseline rather than the full method. FID/FVD/Sync are reported on long-form videos (Sec. 4 ); long-horizon stability metrics are reported as Total / Last over 1-minute rollouts.
Figure 6: Teacher/student and anchor ablation. Left: Teacher/Student/Ours. Right: removing either anchor.
Figure 7: Qualitative comparison to one-step causal baselines. Self-Forcing (1-step) denotes our one-step DMD baseline and shows noticeable color drift, while Causal ODE denotes our one-step ODE baseline and produces blurred frames with noise-level-dependent sharpness. AvatarForcing better preserves appearance consistency and sharp motion under long rollouts, and the ODE-only blur motivates the two-stage design (ODE initialization followed by DMD refinement).
Figure 8: Qualitative ablation on anchor-audio zero padding and RoPE re-indexing. Red boxes highlight mouth regions. Without anchor-audio zero padding, misaligned anchor audio contaminates the anchor pathway and induces mouth jitter and artifacts. Without RoPE re-indexing, the style anchor becomes phase-misaligned as the stream grows, leading to gradual appearance and color drift. The full model avoids both failure modes, consistent with Tab. 5 .