DBL-Diffusion

Green tiles denote the explicit RGBA foreground layer — pulled the way a compositor keys a matte against a green screen.
It carries both the object and the effects it induces (shadows, reflections).

Video Object Insertion

"Add Wooden Boat."

Row 1 is each method's composite output, row 2 its explicit foreground layer. VACE-Inp / VACE-C predict a composite directly and never expose a separate layer; VACE-F and DBL-Insert both learn a foreground layer, so their composite shown here is obtained by alpha-blending that layer onto the background.
VACE-Inp cannot synthesize object-induced effects beyond the masked region, while VACE-C and VACE-F show unstable effects and background appearance changes.
DBL-Insert jointly synthesizes explicit foreground layers and scene-consistent composites for more reliable integration.

Video Layer Decomposition

"Decompose Sheep."

Row 1 is each method's recovered background, row 2 its recovered foreground layer — all three methods produce both. Optimization-based (Gen-Omnimatte) and feed-forward (OmnimatteZero) methods often leave semi-transparent or incomplete foregrounds around fine boundaries and complex effects; DBL-Decompose keeps sharper boundaries and more complete effect separation, at a modest trade-off in pixel-wise reconstruction metrics.

While reconstruction-oriented methods such as Gen-Omnimatte and OmnimatteZero achieve higher PSNR/SSIM/LPIPS scores, this is expected since they are explicitly designed to reconstruct the input video by optimizing pixel-level fidelity or preserving latent representations through attention masking.
In contrast, DBL-Decompose leverages diffusion priors to generate perceptually plausible and visually coherent background layers rather than directly reproducing the ground-truth frames. As shown in above results, our method produces cleaner and more realistic backgrounds while maintaining high-quality foreground layers with accurate object boundaries and scene-dependent effects.

LoRA vs. DoRA training strategy

The RGB branch stays close to the pretrained latent space, so LoRA is enough; the RGBA branch has to learn an out-of-domain concept (transparency), where DoRA's decoupled magnitude/direction update adapts more stably.

Video Layered Object Insertion — BG → FG + Composite

Video Layer Decomposition — Composite → BG + FG

(1) Fréchet Inception Distance(↓)

DBL-Decompose achieves the superior perceptual fidelity for both foreground and background, as evidenced by consistently lower FID and more visually plausible results.

(2) Human preference study — video object insertion

DBL-Insert is consistently preferred over all competing methods across every evaluation criterion, demonstrating superior insertion quality and better alignment with the intended editing objective.

(3) Human preference study — video layer decomposition

DBL-Decompose receives the highest preference for both background and foreground reconstruction by a large margin, indicating that our explicit layer modeling produces more visually plausible decomposed layers than existing methods.

(4) Supplementary video

The supplementary video showcases more qualitative examples and ablations, along with various applications including layer-based video editing, scene-aware video editing, and other applications enabled by our layered video representation.

@article{han2026explicitlayer,
  title     = {Explicit Layer Modeling for Video Object Insertion and Video Layer Decomposition},
  author    = {Han, kyujin and Shin, seungjoo and Cho, sunghyun},
  journal   = {arxiv},
  year      = {2026},
}