Flowception: Non-Autoregressive Video Generation

• Flowception is a video generation framework that interleaves discrete frame insertions with continuous denoising via flow matching for enhanced temporal coherence.
• It overcomes autoregressive and full-sequence limitations by enabling variable-length outputs with reduced computational cost and improved streaming capabilities.
• The DiT-style transformer architecture supports diverse tasks such as image-to-video, interpolation, and scene completion through an efficient ODE–jump process.

Flowception is a non-autoregressive, variable-length video generation framework that interleaves discrete frame insertions with continuous frame denoising via flow matching. Designed to address critical limitations in autoregressive and full-sequence flow-based video models, Flowception achieves improved temporal coherence, reduced computational cost, and enhanced task generality by inducing an ODE–jump process over variable-length frame sequences. It is applicable not only to standard video synthesis, but also to image-to-video, video interpolation, and scene completion, utilizing a unified architecture and scheduling mechanism (Ifriqi et al., 12 Dec 2025).

1. Generative Video Modeling: Motivation and Limitations

The challenge in generative video modeling is to sample realistic, coherent sequences of arbitrary length. Prior approaches are dominated by two paradigms:

• Autoregressive (AR) Denoising: Each new frame is generated conditioned on previously sampled frames, enabling streaming inference but suffering from exposure bias and error accumulation. Training employs teacher forcing, presenting ground-truth history, whereas inference must condition on potentially imperfect prior generations, leading to error drift. Additionally, causal attention required for efficient key-value caching restricts contextual expressivity.
• Full-Sequence Flow-Based Denoising: Models such as full-sequence diffusion denoise all frames in parallel with bidirectional attention, yielding high fidelity and long-term consistency. This necessitates a fixed video length, precludes streaming output, and incurs quadratic attention complexity with respect to frame number.

Flowception seeks a middle ground: a non-autoregressive, stochastic process for variable-length video that (i) avoids AR exposure bias via parallel denoising in bidirectional context, (ii) does not require pre-specified sequence length, and (iii) substantially reduces average attention and computational requirements compared to full-sequence flows (Ifriqi et al., 12 Dec 2025).

2. Probability Flow with Discrete Insertions and Continuous Denoising

Flowception models video generation by alternately performing two atomic operations:

• Continuous flow matching (denoising): Each inserted frame $X^i$ maintains a local “time” $t_i\in[0,1]$, progressing from noise ($t_i=0$) toward clean data ($t_i=1$) under the evolution

$\frac{dX^i}{dt_i}=v_i^\theta(X, t),$

where $v_i^\theta$ is a learned velocity field.

• Stochastic frame insertion: At every generative step, for each frame $i$, Flowception predicts an insertion rate $\lambda_i^\theta(X, t)$, controlling the probability of introducing a new noise frame $\varepsilon\sim\mathcal{N}(0, I)$ immediately after $X^i$, which itself starts with $t_{new}=0$.

Through these tightly coupled processes, Flowception defines a probability path for video generation that combines both “jumps” (insertions) and “flows” (denoising) in arbitrary order, yielding a variable-length ODE–jump process. Marking frames as “active” or “passive” generalizes the model to multiple video synthesis tasks without architectural modification.

3. Mathematical Structure and Training Objective

Let $$\mathcal{X} = \bigcup_{n=0}^\infty \mathbb{R}^{n\times H\times W\times C}$$ be the space of all sequences of $n$ frames. Each frame $X^i$ is associated with local time $t_i$; insertions and denoising steps proceed in global time $t_g$:

• Insertion operator: For sequence $X = (X^1, \dots, X^n)$ and slot $i$,

$$I(X, i, \varepsilon) = (X^1, \dots, X^i, \varepsilon, X^{i+1}, \dots, X^n).$$

• Continuous flow ODE: Optimal velocity field:

$v_i^*(X, t) = \mathbb{E}[X^i_1 - X^i_0 \mid X, t],$

under linear coupling $X^i_{t_i} = t_i X^i_1 + (1-t_i) X^i_0$.

The velocity loss is

$$\mathcal{L}_{\rm vel} = \mathbb{E}_{\tau, X_0, X_1}\left[\mathbf{1}_{[0,1]}(\tau)\, \|v^\theta(X_t, t) - (X_1 - X_0)\|^2\right],$$

with $t = \mathrm{clip}(\tau, 0, 1)$.

• Joint ODE–jump process: At step size $h$, all $t^i\gets \min\{t^i + h, 1\}$; $t_g\gets t_g + h$.
  • For each active $i$, $X^i \leftarrow X^i + h\, v_i^\theta(X, t)$
  • For each slot $i$, insert with probability

$$p_i = h\,\frac{\dot\kappa(t_g)}{1-\kappa(t_g)}\;\lambda_i^\theta(X,t)$$

where $\kappa$ is a monotonic scheduler (typically $\kappa(t)=t$).

• Insertion rate loss: For ground truth insertion count $k^i$ per slot:

$$\mathcal{L}_{\rm ins} = \mathbb{E}_{\tau, t, X_0, X_1}\left[\sum_{i=1}^{\ell(X)} \big(\lambda_i^\theta(X, t) - k^i \log \lambda_i^\theta(X, t)\big)\right].$$

Total training loss: $$\mathcal{L} = \mathcal{L}_{\rm vel} + \mathcal{L}_{\rm ins}$$.

4. Model Architecture, Scheduling, and Efficiency

The Flowception architecture is built on a DiT-style transformer with 38 blocks of hidden size 1536 and 24 attention heads, using pretrained LTX autoencoder latents at $256\times256$ spatial resolution. Each frame is augmented with a learnable “rate token,” projected via an MLP and $\exp(\cdot)$ to yield nonnegative insertion rates $\lambda_i$.

Per-frame AdaLayerNorm (AdaLN) conditions each frame on its own local time $t_i$, decoupling denoising schedules across frames. Attention is by default fully bidirectional across visible frames. For long sequences, local windowing over $K$ frames is supported. Flowception exhibits improved robustness to small attention windows compared to full-sequence flows, as early in sampling, the sequence remains short and global attention is computationally feasible.

Video length emerges jointly with content via the insertion head; no explicit length predictor is required.

Computational complexity: For linear scheduling $\kappa(t) = t$, the expected visible frame fraction at global time $t$ is $t$, and the mean quadratic attention cost per step integrates to $1/3$ the cost of full-sequence flows. Flowception uses $\alpha$ times more steps (to allow late insertions to denoise), so total FLOPs are $\frac{\alpha}{3}$ times those of a full-sequence flow. With $\alpha\approx2$, this realizes a $1.5\times$ speedup in sampling and $3\times$ in training (Ifriqi et al., 12 Dec 2025).

5. Sampling, Inference, and Task Generality

Sampling proceeds by initializing $n_{\rm start}$ noise frames ($t^i=0$), then iteratively denoising all visible frames and probabilistically inserting new frames. Generation continues until all frames reach $t^i=1$.

initialize n_start noise frames X[1...n], t[i]=0, t_g=0 while min(t[i]) < 1: v, lambda = Model(X, t) for i in range(n): X[i] += h * v[i] t[i] = min(t[i]+h, 1) t_g += h for i in slots: with probability h * (d kappa(t_g)/(1-kappa(t_g))) * lambda[i]: insert noise frame after X[i] with t_new=0 return X

By marking frames as “active” or “passive,” the same model supports image-to-video, video-to-video, video interpolation, and scene completion without further modification.

6. Empirical Performance and Ablations

Experiments utilize the Tai-Chi-HD, RealEstate10K, and Kinetics-600 datasets, with 2.1B parameter models trained at $256\times256$ for 300k–400k iterations, producing sequences of length 145 at 16 fps. Evaluation metrics include Fréchet Video Distance (FVD) and VBench suite scores (imaging quality, background consistency, aesthetic quality, motion smoothness, subject consistency, dynamic degree).

Dataset	Full-Sequence FVD	Autoregressive FVD	Flowception FVD
Kinetics-600	204.65	201.34	164.73 (–19.5%)
Tai-Chi-HD	27.30	25.30	25.21 (–7.7%)
RealEstate10K	26.17	47.48	21.80 (–16.7%)

On RealEstate10K, Flowception achieves better FVD (21.80) than both full-sequence (26.17) and AR (47.48) baselines, and similar superiority in image quality (VBench Imaging 51.18 vs. 50.11 and 48.55). Qualitatively, Flowception maintains detail and sharpness to the end of long sequences.

Ablation studies confirm each component’s importance:

• Insertion rules: Learned Flowception $\lambda$ (21.80 FVD) outperforms random slot (25.03), hierarchical (23.94), and fixed left-to-right insertion (23.61).
• Guidance on $\lambda$: Raising classifier-free guidance bias $w_s=1\to5$ produces longer and smoother videos (motion smoothness 99.30 → 99.33).
• Local attention: Flowception performance remains within 10–20% of global attention with small window sizes $K$, unlike full-sequence flows where FVD deteriorates rapidly.
• Task abstraction: The architecture supports multiple video generation and interpolation modes solely via activation of frame slots.

7. Limitations and Prospects for Development

Flowception’s limitations include:

• Under-insertion: Insufficient frame insertion, if guidance is weak, causes choppy motion; tuning classifier-free guidance on $\lambda$ is critical.
• Catch-up for late frames: Frames inserted late in global time require sufficient denoising steps; current scheduling ($\alpha\approx2$) doubles total steps. More adaptive interleaving (e.g., per-frame power-law reparameterization) is a potential optimization.
• Long-horizon scalability: Ultra-high frame-rate or very long video sequences may require hierarchical token compression or more efficient sparse attention for tractability.
• Failure modes: Fast camera pans with fine detail and out-of-distribution context frames can cause misprediction of insertion timing and velocities.

Flowception delivers a unified video generation framework that integrates the streaming and variable-length capabilities of AR models with the fidelity and error resilience of bidirectional flows, all while reducing compute requirements.

---

For in-depth methodology, experimental setups, and open-sourced codebase, see "Flowception: Temporally Expansive Flow Matching for Video Generation" (Ifriqi et al., 12 Dec 2025).