SyncID-Pipe: Video Face Swapping Data Pipeline

• SyncID-Pipe is a data curation and supervision pipeline that generates paired training data, crucial for guiding video face swapping.
• It integrates an Identity-Anchored Video Synthesizer with an image face swapping model to transfer high-fidelity identity details across video frames.
• The framework employs bidirectional ID quadruplets and a synthetic-to-real curriculum to ensure temporal consistency, identity fidelity, and background realism.

SyncID-Pipe is the data curation and supervision pipeline proposed within the DreamID-V video face swapping framework. It is not a network by itself, but a training data generation and organization framework that pre-trains an Identity-Anchored Video Synthesizer (IVS), combines that IVS with a strong Image Face Swapping (IFS) model, and packages the resulting data into bidirectional ID quadruplets $\{I_r, V_r, I_g, V_g\}$ for explicit supervision of a Diffusion Transformer video model. Its stated purpose is to bridge the image-to-video gap: image face swapping models such as DreamID provide very high identity fidelity and attribute preservation on images, whereas direct frame-wise application to video produces temporal flicker and inconsistent motion (Guo et al., 4 Jan 2026).

1. System role and motivation

Within DreamID-V, SyncID-Pipe sits before the main DiT training stage. It produces paired training triples and quadruplets that feed the Modality-Aware Conditioning DiT model, and it also supplies the pose-aware prior used inside the Structural Guidance Module. The pipeline therefore serves two functions simultaneously: it constructs supervision data and it furnishes a reusable motion-conditioned prior through IVS (Guo et al., 4 Jan 2026).

The motivating problem is specific to Video Face Swapping (VFS). Existing methods struggle to maintain identity similarity and attribute preservation while preserving temporal consistency, even though IFS models already provide strong per-image editing quality. SyncID-Pipe addresses that mismatch through a three-step mechanism: learning a video generator that follows pose dynamics while preserving identity, using IFS as a “gold-standard” identity editor on key frames, and turning the result into paired supervision for a video diffusion transformer. The pipeline is therefore a transfer mechanism from image-domain identity editing to temporally coherent video synthesis rather than a standalone inference model (Guo et al., 4 Jan 2026).

A common misunderstanding is to treat SyncID-Pipe as a neural architecture equivalent to DreamID-V itself. The source description is explicit that it is a data pipeline and supervision framework. DreamID-V is the Diffusion Transformer-based video face swapping model; SyncID-Pipe is the mechanism that generates the data and priors used to train it (Guo et al., 4 Jan 2026).

2. Core architecture and generation workflow

SyncID-Pipe contains three main technical components: the Identity-Anchored Video Synthesizer, integration with the DreamID IFS model, and two auxiliary modules, Expression Adaptation and Enhanced Background Recomposition (Guo et al., 4 Jan 2026).

The IVS is a pose-driven, keyframe-anchored video generator. Its backbone is a First-Last-Frame video foundation model implemented as a latent-space Diffusion Transformer trained with Flow Matching. During IVS training, the clean latent video is the VAE encoding of a real portrait video $V_r$, noise is sampled as $\epsilon \sim \mathcal{N}(0,I)$, and the interpolated latent is

$\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$

The training objective is

$$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$

where $y$ includes the first frame, last frame, and pose sequence. The backbone is kept mostly frozen, and IVS adds a lightweight pose adapter through an Adaptive Pose-Attention mechanism. A small ConvNet Pose Guider converts per-frame pose representations into pose features $\mathbf{P}$, and pose features and latent tokens share the same RoPE indices so that pose and latent remain aligned in space and time (Guo et al., 4 Jan 2026).

In the pose-augmented attention mechanism, the frozen DiT provides

$$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$

while the trainable pose adapter provides

$$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$

The resulting attention adds a pose-conditioned term weighted by $\lambda$, so pose injection is explicit and tunable. IVS is trained on a large portrait dataset of 1000 hours and then frozen for subsequent data generation (Guo et al., 4 Jan 2026).

The IFS component uses DreamID as a frozen off-the-shelf face swapper. Given a real source video $V_r$0 with identity A, a real source image $V_r$1 of identity A, and a target image $V_r$2 with identity B, DreamID swaps identity B onto the first and last frame of $V_r$3, producing $V_r$4 and $V_r$5. These frames act as temporally sparse but identity-accurate anchors. IVS then propagates that identity over the full sequence under the pose dynamics of $V_r$6, yielding a coherent synthetic video $V_r$7 (Guo et al., 4 Jan 2026).

Expression Adaptation decouples identity from expression. A 3D face reconstruction model extracts identity coefficients from $V_r$8 and expression and pose coefficients from each frame of $V_r$9, recombining them as

$\epsilon \sim \mathcal{N}(0,I)$0

The reprojected 2D landmarks replace the original pose sequence used by IVS. This retargeted sequence preserves motion and expression from $\epsilon \sim \mathcal{N}(0,I)$1 while preventing identity leakage from identity A into the generated video (Guo et al., 4 Jan 2026).

Enhanced Background Recomposition is designed to make supervision strictly aligned in background and general environment. For a forward-generated triple $\epsilon \sim \mathcal{N}(0,I)$2, SAM2 produces foreground masks on $\epsilon \sim \mathcal{N}(0,I)$3 and $\epsilon \sim \mathcal{N}(0,I)$4, MiniMax-Remover removes foreground from $\epsilon \sim \mathcal{N}(0,I)$5 to create a clean background video $\epsilon \sim \mathcal{N}(0,I)$6, the foreground of $\epsilon \sim \mathcal{N}(0,I)$7 is pasted onto $\epsilon \sim \mathcal{N}(0,I)$8 frame by frame to obtain $\epsilon \sim \mathcal{N}(0,I)$9, and edges are feathered. The supervision video remains the real $\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$0; the recomposed $\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$1 is used only as an input condition (Guo et al., 4 Jan 2026).

3. Bidirectional ID quadruplets as the central supervision unit

The core data unit of SyncID-Pipe is the bidirectional ID quadruplet

$\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$2

Element	Role
$\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$3	Reference image of real identity A
$\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$4	Real portrait video of identity A
$\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$5	Reference image of identity B
$\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$6	Synthetic video of identity B

From this quadruplet, SyncID-Pipe derives two paired directions for training. The first is the forward-generated paired data $\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$7. The second is the backward-real paired data $\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$8, also described in the main text through equivalent triplet formulations involving $\mathbf{z}_t = (1-t)\mathbf{z}_0 + t\epsilon .$9 and $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$0. The key distinction is that one direction uses synthetic ground truth generated by IVS, while the other uses real ground truth in a curriculum stage that adjusts realism and background fidelity (Guo et al., 4 Jan 2026).

The sampling strategy begins with source videos from OpenHumanVid after filtering for same-ID pairs. For each video $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$1, a reference image $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$2 of the same identity is sampled, and a different identity image $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$3 is sampled from a large identity pool such as VGGFace2. Pose and expression are extracted frame-wise from $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$4, retargeted to identity B, DreamID generates the first and last key frames $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$5, and IVS uses those key frames plus the retargeted pose sequence to synthesize $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$6 (Guo et al., 4 Jan 2026).

These quadruplets provide fully supervised training data for the video DiT. The ground-truth latent $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$7 is the VAE encoding of either the synthetic video $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$8 or the real video $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t,\mathbf{z}_0,\epsilon}\left[ \left\|(\mathbf{z}_0 - \epsilon) - \mathbf{v}_\theta(\mathbf{z}_t,t,y)\right\|^2 \right],$$9, and the conditions include identity features, pose guidance, and spatio-temporal context. The supervision is therefore explicit at the video level rather than implicit through masked self-reconstruction alone (Guo et al., 4 Jan 2026).

An important misconception concerns the word “quadruplet.” SyncID-Pipe does not introduce an explicit contrastive, triplet, or margin-based quadruplet loss. “Bidirectional quadruplets” refers to the data structure. Identity supervision is implicit through Flow Matching regression to a video that already has the correct identity and through the later Identity-Coherence Reinforcement Learning stage (Guo et al., 4 Jan 2026).

4. Training curriculum and conditioning interface

SyncID-Pipe enables the two-stage Synthetic-to-Real Curriculum and the subsequent IRL stage. IVS is first pre-trained using only real portrait videos: given $y$0, its first and last frames, and its pose sequence, the model learns pose-conditioned reconstruction under Flow Matching. This establishes a pose-conditioned video prior before any identity transfer is introduced (Guo et al., 4 Jan 2026).

The transfer of IFS advantages to video occurs through IVS distillation. DreamID produces identity B on a small number of key frames; IVS propagates that identity across the full pose trajectory to create $y$1; DreamID-V is then trained to map from a reference video $y$2 with identity A plus an ID image $y$3 of B to the synthetic target video $y$4. This is described as a form of knowledge transfer or distillation from image to video without an explicit distillation loss. The transferred supervision is effective in part because IVS outputs lie in the same latent distribution as the DiT prior (Guo et al., 4 Jan 2026).

The Synthetic-to-Real Curriculum has two stages. Synthetic Training uses forward synthetic pairs where the ground truth is $y$5. The appendix reports that the VAE latent space of synthetic $y$6 overlaps strongly with the base DiT’s outputs in t-SNE, which makes optimization easier and allows the model to reach very high identity similarity quickly. Real Augmentation Training then fine-tunes on backward-real paired data $y$7, using recomposed synthetic input with background matching $y$8 while keeping real $y$9 as ground truth. This improves realism, especially background statistics, while preserving the identity quality learned during synthetic training (Guo et al., 4 Jan 2026).

Identity-Coherence Reinforcement Learning is not strictly part of SyncID-Pipe, but SyncID-Pipe makes it possible by providing paired identity-aware supervision. The frame-wise identity score is based on ArcFace cosine similarity, and the corresponding Q-value is

$\mathbf{P}$0

with chunk-wise averages used to weight the Flow Matching loss: $\mathbf{P}$1 This reweights learning toward hard frames and reduces temporal flicker, particularly in pose-extreme segments (Guo et al., 4 Jan 2026).

SyncID-Pipe also defines the conditioning interface used by DreamID-V’s Modality-Aware Conditioning. The Spatio-Temporal Context Module uses a reference video and a dilated face mask, concatenated along the channel dimension with the noisy latent video. The Structural Guidance Module uses the same Pose-Attention mechanism as IVS, with parameters initialized from pre-trained IVS. The Identity Information Module extracts an ID embedding from $\mathbf{P}$2 and concatenates identity tokens with patchified video tokens along the token dimension. Abstractly,

$\mathbf{P}$3

which summarizes how context, pose, and identity are injected into the DiT (Guo et al., 4 Jan 2026).

5. Empirical effects, implementation details, and limitations

The most direct evidence for SyncID-Pipe appears in the DreamID-V ablation study. Removing the quadruplet mechanism yields substantially worse identity transfer: the setting “w/o Quadruplet” reports ID-Arc $\mathbf{P}$4, Var $\mathbf{P}$5, Pose $\mathbf{P}$6, Expr $\mathbf{P}$7, and FVD $\mathbf{P}$8, whereas the full model reports ID-Arc $\mathbf{P}$9, Var $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$0, Pose $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$1, Expr $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$2, and FVD $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$3. Training without the synthetic stage gives ID-Arc $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$4, Var $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$5, Pose $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$6, Expr $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$7, and FVD $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$8. Training without real augmentation gives ID-Arc $$\mathbf{Q}=\mathbf{Z}\mathbf{W}_q,\quad \mathbf{K}=\mathbf{Z}\mathbf{W}_k,\quad \mathbf{V}=\mathbf{Z}\mathbf{W}_v,$$9, Var $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$0, Pose $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$1, Expr $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$2, and FVD $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$3. Training without IRL gives ID-Arc $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$4, Var $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$5, Pose $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$6, Expr $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$7, and FVD $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$8. The pattern indicates that the quadruplet structure is critical for identity similarity, the synthetic stage is critical for transferring IFS-level identity quality, and real augmentation is critical for realism and background fidelity (Guo et al., 4 Jan 2026).

The implementation schedule is equally specific. IVS pre-training uses 1000 hours of portrait videos. Synthetic Training for DreamID-V uses 100 hours of IVS-generated videos $$\mathbf{K}'=\mathbf{P}\mathbf{W}_k',\quad \mathbf{V}'=\mathbf{P}\mathbf{W}_v' .$$9. Real Augmentation Training uses 150 hours of mixed real $\lambda$0 and synthetic $\lambda$1 with background recomposition. IRL is performed on 10 hours of videos with high ID variance. The optimizer is AdamW with a constant learning rate of $\lambda$2 for all stages. Stage 1 uses 50k iterations with global batch size 16; Stage 2 adds 80k iterations with global batch size 32 (Guo et al., 4 Jan 2026).

Qualitatively, the reported figures and supplementary comparisons emphasize stable identity across frames, high identity similarity close to DreamID IFS but without flicker, and better expression and motion fidelity than image-based and video-based baselines under occlusions and complex expressions. The explanation given in the source is that synthetic IVS data supplies strong identity supervision under controlled pose, real augmentation via background recomposition improves realism, and explicit quadruplets align identity and attributes more tightly than inpainting-based or unpaired schemes (Guo et al., 4 Jan 2026).

The stated limitations are operational rather than conceptual. Expression Adaptation depends on 3D reconstruction quality, so extreme occlusion or very low resolution can degrade retargeted pose and cause imperfect motion or identity leakage. Background recomposition may introduce artifacts, even though the real training target remains $\lambda$3. Extreme motions and occlusions remain hard cases, which is precisely why IRL emphasizes low-similarity frames. The pipeline is built on portrait videos, so strong generalization to animation or unusual lighting does not eliminate the possibility of domain shift under more extreme non-human or stylized data (Guo et al., 4 Jan 2026).

6. Broader interpretations and adjacent synchronization frameworks

The established use of SyncID-Pipe is the DreamID-V data engine just described. At the same time, the surrounding literature suggests a broader interpretation of the term as a synchronization-and-identification pipeline operating over structured state histories. In that broader sense, a SyncID-Pipe-like design appears whenever a system creates explicit identifiers for evolving states, propagates them through calibrated stages, and uses those identifiers to recover consistent global semantics (Burgess et al., 2022).

One such interpretation is the hierarchical, policy-based data service described in “Continuous Integration of Data Histories into Consistent Namespaces,” which organizes handlers, parents, and roots into a hierarchy of calibrated data pipelines. There, fair interleaving, pull-based batching, and association tuples such as $\lambda$4, $\lambda$5, and $\lambda$6 define a versioned coordinate system or versioned namespace with “latest official version” semantics and self-protecting rate-limited versioning. This is not DreamID-V’s video pipeline, but it articulates a general model in which version identifiers are assigned by a controlled synchronization hierarchy rather than by a global clock (Burgess et al., 2022).

A second interpretation appears in the cross-chain state channel scheme Interpipe. There, recursive state synchronization propagates round-indexed state proofs $\lambda$7 across chains, while accumulator digests and membership witnesses bind particular cross-chain transactions to particular synchronization rounds. The resulting design behaves like a distributed identifier-driven pipe in which the pair $\lambda$8 functions as a state SyncID and per-transaction witnesses tie concrete events to those synchronized states. This is a blockchain synchronization framework rather than a multimedia training pipeline, but it shares the same abstract pattern of explicit synchronized identifiers plus staged propagation (Liang et al., 2024).

A third interpretation arises in symmetry reduction for turbulent pipe flow. There, the supplied exposition frames the method of slices as a way of “synchronizing and identifying” exact coherent structures by quotienting continuous symmetries. In the symmetry-reduced state space, travelling waves reduce to equilibria and relative periodic orbits reduce to periodic orbits, enabling a consistent identification of dynamically relevant states and their unstable manifolds. This again is not the DreamID-V system, but it suggests an analogous use of synchronization to produce stable identifiers for otherwise drifting trajectories (Willis et al., 2012).

Taken together, these adjacent usages suggest that the name “SyncID-Pipe” naturally denotes more than a single implementation. In its specific, published form it is DreamID-V’s bidirectional supervision pipeline for video face swapping. More generally, the term suggests a design pattern in which synchronization, explicit identifiers, and staged propagation are combined to make evolving data histories tractable, whether the underlying objects are face identities in video, versioned key-value histories, cross-chain channel states, or symmetry-reduced flow trajectories.