Controllable Layered Image Generation for Real-World Editing (LASAGNA)

TL;DR

LASAGNA is a unified diffusion framework from Adobe Research + UC Santa Cruz that jointly generates a composite image together with its constituent layers — a photorealistic background and a transparent RGBA foreground with physically-grounded visual effects (shadows, reflections). A single attention-based model handles three workflows — Text2All (text → composite + BG + FG), FG_Gen (BG → FG + composite), and BG_Gen (FG → BG + composite) — by labeling each visual latent (noise / BG / FG / mask) with a learnable role-embedding. The authors release LASAGNA-48K, a real-image (not synthetic) dataset of background + RGBA foreground pairs built by running LayerDecomp over COCO-style imagery with a VLM-based quality filter, and LasagnaBench, the first benchmark for layer editing.

Key claims

LASAGNA generates an image jointly with a photorealistic background and a transparent foreground that carries physically-grounded visual effects (shadows, reflections), addressing the standard layered-generation failure mode where foreground layers lack realistic compositing cues [Abstract, Fig. 1].
A single model is trained as a layer-conditional denoiser over a unified token sequence; learnable role embeddings tag each latent (noise / BG / FG / mask) so one attention-based network handles Text2All, FG_Gen, and BG_Gen without per-mode heads [project page, §3].
Mask-guided spatial conditioning gives explicit foreground-placement control, which prior unified layer generators (e.g. ART) lack [§2, contrastive discussion].
LASAGNA-48K is built from real-world imagery (e.g. COCO) decomposed via LayerDecomp, with an InternVL2.5-8B-based curator scoring BG fidelity as a proxy for overall layer quality at ~72.3% precision/recall on a 1K human-annotated test set, followed by Qwen2.5-VL-32B for FG-artifact removal [§Data, §Data filtering].
LasagnaBench is positioned as the first benchmark targeting layer-editing (identity + visual-effect preservation across layers), filling a gap left by composite-only benchmarks [Abstract, §Benchmark].
Versus prior work: ART (2025) supports layout conditioning but is limited in editing versatility; LayerDecomp (2024) and Qwen-Image-Layered decompose but cannot synthesize novel content — LASAGNA claims to be the first unified, controllable generator producing transparent FGs with physically-grounded effects [§Related, §2].

Method

LASAGNA treats layered image generation as a layer-conditional denoising problem with a shared token vocabulary across all three workflows. The token sequence concatenates four latent streams — noisy target latent, background latent, foreground (RGBA) latent, and a location mask — each tagged with a learnable role embedding so the model knows which latent is being conditioned on and which is being denoised. The same attention backbone is then run under three masking regimes:

Text2All: only text is given; the model denoises all three of {composite, BG, FG} from noise.
FG_Gen: BG and (optionally) a location mask are given; FG and the composite are denoised.
BG_Gen: FG is given; BG and composite are denoised.

This unifies tasks that would normally require separate networks. Training data comes from LASAGNA-48K, built by running LayerDecomp-style decomposition over real photographs and filtering with two VLMs: InternVL2.5-8B scores background fidelity (used as a proxy for joint BG+FG quality, since LayerDecomp’s consistency loss couples them), and Qwen2.5-VL-32B does a second pass to flag residual artifacts in the extracted RGBA foregrounds.

Results

The paper’s empirical headline is qualitative: LASAGNA produces composites whose BG and FG layers can be directly re-edited into new coherent scenes (Fig. 1b) while preserving object identity and visual effects. Quantitative numbers live in LasagnaBench, which the project page does not fully expose — the abstract claim is that LASAGNA “excels in generating highly consistent and coherent results across multiple image layers simultaneously” relative to prior layered generators (ART) and decomposition methods (LayerDecomp, Qwen-Image-Layered) [Abstract]. The data-curator’s 72.3% precision/recall on human-annotated quality labels is the most concrete number surfaced in the public materials [§Data filtering].

Why it’s interesting

This is the image-domain counterpart to the per-pixel video understanding work the wiki has been tracking. Where OmnimatteZero decomposes an existing video into BG + FG + effects training-free, LASAGNA synthesizes an image and its layers jointly, and is trained — but trained over real photographs decomposed by LayerDecomp, which makes LASAGNA-48K a derivative of that decomposition pipeline. Two design choices are worth flagging for Luma: (1) the role-embedding trick that lets a single attention backbone serve three different conditioning regimes is reusable for any multi-modal joint-generation setup (e.g. video + audio + depth); (2) “VLM-as-data-curator” with a cheap proxy metric (BG fidelity standing in for joint BG+FG quality) is a recurring pattern in the recent crop of large-scale dataset releases.