Self-Forcing++: Towards Minute-Scale High-Quality Video Generation

TL;DR

Self-Forcing++ extends the Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion streaming-AR recipe past its ~5 s teacher horizon without any long-video teacher or long-video dataset. The student is rolled out beyond the teacher’s window (up to 100 s), noise is re-injected into its own degraded long rollout, and DMD is computed on a uniformly sampled teacher-length window — teaching the student to recover from its own error-accumulated states. With a rolling KV cache used at both train and inference and an optical-flow-reward GRPO post-stage, the resulting Wan2.1-T2V-1.3B model generates coherent video up to 4 min 15 s (99.9% of the base’s positional-embedding budget, ~50× the Self-Forcing baseline) at the same 17 FPS throughput.

Key claims

The dominant failure mode of streaming-AR distillation past 5 s is compounding autoregressive error, not a breakdown of the AR mechanism itself — long rollouts retain structural coherence but degrade into motion stagnation, scene freezing, and exposure drift [§3.2 Motivation].
A short-horizon (5 s) teacher implicitly captures the data distribution of a longer “world”; any short contiguous window of a generated long video can be treated as a sample from the teacher’s distribution, justifying sliding-window DMD on long student rollouts [§3.2 Extended DMD].
Extended DMD: roll the student out to T̃ frames (T̃ ≥ T_teacher), uniformly sample a contiguous window of length T_teacher from the rollout, and apply DMD between student and teacher distributions only inside that window [§3.2, Eq. 2].
Backwards noise initialization (re-noising the student’s own long rollout instead of starting from independent Gaussian noise) is essential — it preserves temporal dependencies with the preceding frames so the teacher distribution is evaluated on a temporally-consistent trajectory [§3.2, Eq. 1].
A rolling KV cache used during both training and inference eliminates Self-Forcing’s first-chunk-mask train/inference mismatch and removes the need for CausVid’s overlapping-frame recomputation [§3.2 Training with rolling KV cache].
An optical-flow-magnitude reward under GRPO (per-step importance weights computed from the re-noising step, sequence-level log-prob summed over the AR rollout) suppresses spikes in inter-frame optical flow, fixing abrupt scene transitions caused by sliding-window memory loss [§3.3].
VBench is biased toward over-exposed and degraded long videos because of outdated evaluation models; the authors replace it with a Gemini-2.5-Pro-judged “Visual Stability” score that explicitly rates over-exposure and error accumulation [§3.4].
On 5 s clips, the model is competitive with Self-Forcing (VBench Total 83.11 vs. 83.00; Semantic 80.37 vs. 80.14) — no regression from training only on long rollouts [Table 1].
At 50 s, Visual Stability jumps from 40.12 (Self-Forcing) and 40.47 (CausVid) to 90.94, with Dynamic Degree 55.36 vs. 34.35 / 37.35 and Temporal Quality 91.03 vs. 88.17 / 89.34 [Table 1].
At 100 s, the model still posts Visual Stability 84.22 (vs. Self-Forcing 32.03, CausVid 39.21), Dynamic Degree 54.12 (vs. 26.41 / 34.60), and Text Alignment 26.04 (vs. 22.00 / 24.41) — +104.9% Dynamic Degree and +56.4% over the two baselines respectively [Table 2, §4.2].
Shortening the training attention window (a naive alternative fix for Self-Forcing’s train/inference mismatch) produces only modest gains — Visual Stability rises from 40.12 → 52.50 at window=9, vs. 90.94 with the full Self-Forcing++ recipe [Table 3].
Training-budget scaling unlocks horizon: at 1× budget the model produces a coherent 5 s; at 4× extends to longer horizons with consistent subjects; at 20× yields stable 50 s+; at 25× reaches the 4 min 15 s / 1023-latent-frame ceiling of Wan2.1-T2V-1.3B’s positional embedding [§4.4].

Method

The base is Wan2.1-T2V-1.3B (5 s @ 16 FPS, 480×832) initialized to a few-step AR student via the standard CausVid/Self-Forcing ODE-distillation init. The novel training stage is extended-rollout DMD with backwards noise initialization on a rolling KV cache. A training step:

Roll the student out chunkwise (3 latent frames per chunk) to a horizon T̃ ∈ {21, …, 100 s of latent frames} using a rolling KV cache that evicts oldest chunks — identical to inference.
Take the clean rollout x_0^{1..T̃}, re-inject noise according to the diffusion schedule σ_τ to get x_τ^{1..T̃} (Eq. 1). This noise-re-injected long trajectory is the starting state for DMD.
Uniformly sample a starting index t* ∈ [1, T̃ − T_teacher], extract the window x_τ^{t*..t*+T_teacher}, and compute the asymmetric DMD KL between student and teacher distributions on that window only (Eq. 2). The teacher only ever sees teacher-length windows, but the windows themselves can be sampled from arbitrary points in a long student rollout, so the student is taught to recover from error-accumulated states.

There is no overlap-frame recomputation, no latent-frame masking, no long-video teacher, and no long-video training data — only the same Wan2.1-T2V-1.3B teacher Self-Forcing uses. The rolling cache is identical at training and inference, so the train/inference distribution mismatch that Self-Forcing’s first-chunk-mask was patching is gone by construction.

After extended DMD converges, an optional GRPO stage with an optical-flow-magnitude reward fine-tunes for smoother long-range transitions (§3.3). Per-step importance weights are read from the same re-noising operation in Eq. 1, and the AR rollout’s total log-probability is the sum of per-step log-probs.

For evaluation, Gemini-2.5-Pro rates each generated long video along over-exposure and error-accumulation axes, aggregated to a Visual Stability score (§3.4, Fig. 3) — designed to replace VBench Imaging/Aesthetic Quality on long videos, which the paper shows favors over-exposed CausVid and dark/static Self-Forcing outputs.

Results

Throughput stays at 17 FPS / 0.69 s first-frame latency — same as Self-Forcing and CausVid on the same base.

5 s clips: VBench Total 83.11 (vs. Self-Forcing 83.00, CausVid 82.46, MAGI-1 79.18, NOVA 80.12, Pyramid Flow 81.72), Quality 83.79, Semantic 80.37. The model is not specifically trained on 5 s but matches the prior streaming-AR SOTA [Table 1].

50 s clips: Visual Stability 90.94 (vs. CausVid 40.47, Self-Forcing 40.12, SkyReels-V2 60.41, MAGI-1 51.25, NOVA 45.94); Dynamic Degree 55.36 (vs. CausVid 37.35, Self-Forcing 34.35); Temporal Quality 91.03 (vs. 89.34 / 88.17); Text Alignment 26.37 (vs. 25.25 / 24.77). Framewise quality 60.82 is matched with the baselines (the paper flags framewise quality as unreliable on long videos) [Table 1].

75 s clips: Visual Stability 86.10, Dynamic Degree 55.62, Text Alignment 26.31 — vs. Self-Forcing 35.00 / 29.15 / 23.39 and CausVid 39.84 / 35.82 / 24.76 [Table 2].

100 s clips: Visual Stability 84.22, Dynamic Degree 54.12, Text Alignment 26.04 — vs. Self-Forcing 32.03 / 26.41 / 22.00 and CausVid 39.21 / 34.60 / 24.41. The +56.4% Dynamic Degree vs. CausVid and +104.9% vs. Self-Forcing are the headline numbers [Table 2, §4.2].

Failure modes at long horizons that this method avoids: Self-Forcing’s progressive darkening (motion collapse + exposure drift), CausVid’s over-exposure trend, and SkyReels’ sporadic noise-collapse recovery into low-fidelity content [§4.2 qualitative].

Ablations:

Shrinking Self-Forcing’s attention window (15 → 12 → 9 latent frames) modestly improves Visual Stability (44.69 → 42.19 → 52.50) but never approaches Self-Forcing++‘s 90.94 [Table 3] — the train/inference mismatch is not the dominant cause; the lack of explicit long-rollout supervision is.
GRPO with optical-flow reward suppresses optical-flow spikes between consecutive frames, eliminating abrupt scene transitions visible in the no-GRPO variant [§4.3.2, Fig. 5].
Training-budget scaling: 1× → 5 s coherent; 4× → consistent multi-second subjects; 8× → detailed backgrounds + accurate subjects; 20× → stable 50 s+; 25× → 4 min 15 s (1023 latent frames, the positional-embedding ceiling) [§4.4, Fig. 6].

Why it’s interesting

This is the direct continuation of Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion and the missing entry from the streaming-AR lineage already on the wiki — Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation and Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation fix the initialization stage of Self-Forcing (the bidirectional → AR ODE distillation step); Context Forcing: Consistent Autoregressive Video Generation with Long Context adds long-context memory and Bounded Positional Encoding; Self-Forcing++ attacks the third orthogonal axis: what does the training loop look like when the student rolls out past the teacher’s horizon? The answer — re-noise the student’s own degraded long rollout, then DMD only on a teacher-sized window sampled from it — is a clean restatement of Self-Forcing’s “simulate inference at training time” principle one level deeper: now training also simulates the error accumulation that happens at inference. It’s the cleanest published evidence so far that the long-AR-video Pareto from Context Length / Quality Trade-off in Video Generation can be shifted dramatically without changing the cache structure or backbone — only the loss-sampling protocol. The Visual-Stability vs. VBench discrepancy the paper documents is also a useful warning for VLM-as-Evaluator: outdated VBench imaging metrics actively reward over-exposed and dark/static long videos.

Self-Forcing++ is also already referenced by FlowAct-R1: Towards Interactive Humanoid Video Generation (FlowAct-R1) as the inspiration for memory-error injection — filing the original now makes that citation resolvable.