Flowception: Temporally Expansive Flow Matching for Video Generation

TL;DR

Flowception is a variable-length, non-autoregressive video generation framework that interleaves continuous flow-matching denoising of existing frames with discrete, learned frame insertions during sampling. It is the direct video analog of OneFlow’s text+image Edit-Flow + Flow-Matching pairing (same Meta FAIR group), and reframes long-form video generation as a coupled ODE–jump process where the model itself decides when and where to insert new frames. The interleaved insertion–denoising schedule averages ~1/3 of the attention FLOPs of a full-sequence model during training (≤2/3 during sampling) and the same network natively handles I2V, video interpolation, V2V, and scene completion by toggling which input frames are “active” (can induce insertions to their right) vs. “passive”. Improves FVD across Kinetics-600 / Tai-Chi-HD / RealEstate10K against both full-sequence and AR baselines, with the AR-baseline gap on RealEstate10K particularly large (21.80 vs 47.48 FVD).

Key claims

A non-AR video model can natively support variable-length generation by coupling continuous flow-matching denoising of existing frames with a learned per-frame Poisson insertion rate that triggers stochastic insertion of new noise frames during sampling [§3.1, §3.2].
The model has two prediction heads on a DiT backbone: a dense velocity field over existing frames and a per-frame insertion rate token (learnable token concatenated, MLP projection, exponential activation) [§3.4].
Per-frame time conditioning is enabled by changing AdaLN modulation to operate per-frame, so different frames can be at different noise levels — including frames that don’t yet exist (extended time t̄ < 0 = “deleted state”) [§3.1, §3.3, Fig. 3].
Training samples noisy sequences by drawing a global extended time τ̄, then assigning per-frame extended times via a uniform offset under the linear scheduler, which exactly matches the generation-time distribution of visible frames and their times [§3.3, Eqs. 7–10].
The insertion head is supervised by the negative log-likelihood of a Poisson distribution over the number of missing frames to the right of each position [§3.3, Eq. 11].
Under a linear scheduler, the expected attention FLOPs during training average to ~1/3 of a full-sequence flow-matching model at the same final length; during sampling the bound is ~2/3 of full-sequence [§3.4].
The same model handles I2V, V2V, video interpolation, and scene completion by passing different sets of context frames as either active (insertions allowed to their right) or passive (no insertions), specified only by relative order — no need to declare gap sizes [§3.4, Fig. 4].
Classifier-free guidance can be applied independently to the insertion rate (κ_R) — increasing κ_R biases the model toward longer videos, and prevents under-insertion “choppy transitions”; motion smoothness rises and dynamic degree falls as κ_R increases [§4.3, Fig. 7, Table 3].
Flowception inserts distant frames first and progressively fills in, which empirically makes it far more amenable to local-attention windows than full-sequence baselines [§4.3, Fig. 8].
Beats full-sequence and AR baselines on FVD across Kinetics-600 (164.73 vs 201.34 / 204.65), Tai-Chi-HD (25.21 vs 25.30 / 27.30), and RealEstate10K (21.80 vs 47.48 / 26.17) [Table 1].
Learned insertion ordering beats fixed insertion orders even at matched final length: data-driven > left-to-right (FVD 21.80 vs 23.61) > hierarchical (23.94) > random (25.03) on RealEstate10K [Table 2].
Left-to-right insertion ordering inside Flowception still drastically beats true AR baselines (FVD 23.61 vs 47.48 causal / 45.13 non-causal AR) — the gain comes from denoising new frames concurrently with not-yet-clean prior frames, not from the ordering itself [§4.3, Table 4].

Method

Flowception generates videos as variable-length sequences of frames, each frame carrying its own time value. Generation starts with a fixed small number of pure-noise frames at t=0 and a global time τ that advances from 0 to 1. At each step, the model (a) applies a flow-matching velocity update to every existing frame, and (b) for each position, samples whether to insert a new noise frame to the right with probability set by the insertion rate λ_i and the scheduler’s instantaneous insertion mass μ̇(τ)/μ(τ). Inserted frames start at noise level 0 and are denoised in subsequent steps in the context of partially-denoised neighbors. The process ends when every per-frame time reaches 1.

The architecture is a 2.1B DiT (38 blocks, hidden dim 1536, 24 heads × 64) with VideoROPE positional embeddings and full bidirectional MMDiT-style attention. Two changes vs. a standard DiT: AdaLN is per-frame so each frame’s modulation reflects its own time; and a learnable insertion-rate token is concatenated to each frame’s tokens, projected via MLP and exponentiated to give the non-negative rate. Conditioning frames are passed via a doubled input-channel scheme (first half noisy frames, second half clean conditioning or zero padding).

Training samples a global extended time τ̄ ∈ [-1, 2], computes per-frame extended times as t̄_i = τ̄ - δ_i with δ_i uniform on [0,1], then clips to [0,1]. Frames with t̄_i < 0 are deleted (gives the Poisson supervision target for the insertion head), frames with 0 ≤ t̄_i < 1 get the flow-matching velocity loss, and frames with t̄_i ≥ 1 are frozen. The total loss is the Poisson NLL on insertion counts plus the standard flow-matching velocity MSE.

Results

Image-to-video, 145 frames, 16 FPS: FVD on Kinetics-600 164.73 vs. autoregressive 201.34 vs. full-sequence 204.65; on Tai-Chi-HD 25.21 vs. 25.30 / 27.30; on RealEstate10K 21.80 vs. 47.48 / 26.17 [Table 1].
VBench subject consistency: 94.74 / 94.43 / 87.02 (K-600 / Tai-Chi / RealEstate10K) — best or second-best across the board [Table 1].
AR-baseline gap is driven by AR, not by causal masking: AR causal 47.48 vs. AR non-causal 45.13 vs. Flowception 21.80 on RealEstate10K [Table 4]. The AR-vs-Flowception gap is much larger than the causal-vs-bidirectional gap inside AR.
Insertion-pattern ablation (RealEstate10K, fixed length): random 25.03 → hierarchical 23.94 → left-to-right 23.61 → learned 21.80 FVD [Table 2]. Even forced left-to-right insertion inside the Flowception framework massively beats the AR baselines (23.61 vs 47.48), demonstrating the win is concurrent denoising of in-progress frames, not the ordering.
Insertion guidance: increasing κ_R from 1.0 → 5.0 raises motion smoothness (99.30 → 99.33) and lowers dynamic degree (78.59 → 77.78); FVD degrades from 21.80 → 25.30 [Table 3]. The headline behavior is qualitative — guidance biases the model to insert more frames, producing longer videos from the same seed (Fig. 7).
Local attention amenability: with local-only attention windows (1 frame back / 1 frame forward), full-sequence quality collapses while Flowception degrades much less, attributed to its empirical coarse-to-fine insertion order [§4.3, Fig. 8].
Compute: ~1/3 attention FLOPs vs. full-sequence training, ≤2/3 during sampling, comparable to AR sampling but with more robustness at low NFEs [§3.4].

Why it’s interesting

Flowception is the cleanest video-side instantiation of the insertion-as-generative-primitive idea that OneFlow (OneFlow: Concurrent Mixed-Modal and Interleaved Generation with Edit Flows) demonstrated for text+image. Same Meta FAIR group (Berrada, Nguyen, Chen), same Edit-Flow + Flow-Matching machinery, same emergent ability to handle multiple tasks (here: I2V / interp / V2V / scene completion; there: text + variable-number-of-images interleaved). The fact that the recipe transfers from discrete-text + continuous-image to continuous-frames + discrete-frame-count is the strongest signal yet that “Flow + Flow” is a real fourth corner of the generative-model design space, alongside AR-everywhere, full-sequence-everywhere, and Transfusion-style AR+FM.

In the Autoregressive Video Generation cluster the wiki has been tracking, every filed approach (Self-Forcing / Causal Forcing / Context Forcing / PackForcing / PFP / MALT / LongVie 2) attacks the training-test exposure-bias problem by perturbing the AR conditioning to look like a rollout. Flowception sidesteps the problem entirely by refusing to commit frames before they’re fully denoised — its left-to-right ablation already beats AR-causal by 24 FVD points without any of the drift-correction machinery the AR lineage needs. This is a structural argument for non-AR streaming, not just a points-on-a-benchmark win.

The RoPE story in Context Length / Quality Trade-off in Video Generation gets a new wrinkle here: Flowception inherits VideoROPE and uses bidirectional attention, but its coarse-to-fine frame-count trajectory means most attention happens on short sequences. This is a different angle than the RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably / Bounded-PE workarounds — it doesn’t fix the RoPE failure mode, it just keeps the working sequence short enough that it doesn’t bite. The local-attention compatibility result is the same story: temporal locality of late insertions means distant context is only attended to early when there are few frames to attend to.

The insertion-CFG result is a new axis for Classifier-Free Guidance variants: alongside HuMo’s per-modality scales, ALG’s content schedule, and NAG’s attention-space negative guidance, Flowception adds per-prediction-head guidance — the velocity head’s CFG and the insertion-rate head’s CFG can be tuned independently, and κ_R is interpretable (more guidance → longer videos).