LayerComposer: Multi-Human Personalized Generation via Layered Canvas

TL;DR

LayerComposer tackles the challenge of interactive, multi-human personalization by introducing a layered canvas where each subject is placed on its own RGBA layer, enabling occlusion-free composition and intuitive layout control. It pairs this representation with a diffusion-based pipeline and a transparent latent pruning mechanism to keep conditioning costs nearly independent of the number of subjects, along with a layerwise cross-reference training scheme to reduce copy-paste artifacts. Empirical results show superior spatial control, identity preservation, and prompt adherence across 4P, 2P, and 1P personalization benchmarks, plus favorable user study outcomes. The approach also supports background integration and generalizes to non-human content, suggesting practical impact for interactive group photo generation and future extensions with background and cross-modal guidance.

Abstract

Despite their impressive visual fidelity, existing personalized image generators lack interactive control over spatial composition and scale poorly to multiple humans. To address these limitations, we present LayerComposer, an interactive and scalable framework for multi-human personalized generation. Inspired by professional image-editing software, LayerComposer provides intuitive reference-based human injection, allowing users to place and resize multiple subjects directly on a layered digital canvas to guide personalized generation. The core of our approach is the layered canvas, a novel representation where each subject is placed on a distinct layer, enabling interactive and occlusion-free composition. We further introduce a transparent latent pruning mechanism that improves scalability by decoupling computational cost from the number of subjects, and a layerwise cross-reference training strategy that mitigates copy-paste artifacts. Extensive experiments demonstrate that LayerComposer achieves superior spatial control, coherent composition, and identity preservation compared to state-of-the-art methods in multi-human personalized image generation.

Figures (7)

• Figure 1: LayerComposer introduces an interactive personalization paradigm that enables a Photoshop-like experience for multi-human text-to-image generation. It allows users to place and resize subjects on a layered canvas, providing spatial control and reference human injection simultaneously. LayerComposer then harmonizes the layered canvas into a coherent, high-fidelity image that follows the prompt.
• Figure 2: Layerwise Cross-Reference Sampling Strategy. During training, LayerComposer constructs the layered canvas by sampling each layer from a random source image within the same identity group, where the target image is another different sample in that group. This strategy intentionally introduces mismatches both among input layers and between the inputs and the target, encouraging LayerComposer to learn adaptive alignment to the prompt and context during inference. In contrast, conventional single-reference training tends to overfit to pixel-level correspondences, often leading to outpainting or copy-paste artifacts (see examples in \ref{['fig:ablation']}).
• Figure 3: LayerComposer Pipeline. LayerComposer conditions a diffusion model on both a text prompt and a layered canvas, where each layer is first encoded into latents using a VAE. To distinguish multiple layers that may spatially overlap, each layer is assigned a unique layer index $j$, which is added to its 3D positional embeddings to be $[j, x, y]$. Finally, transparent latent pruning is performed to retain only the latents in non-transparent regions per layer, while discarding the others (gray boxes) for scalable personalized generation.
• Figure 4: Ablation study. Layerwise cross-reference training enables coherent, well-integrated multi-subject compositions by mitigating copy–paste artifacts and encouraging subjects to integrate naturally into the scene. The layered canvas effectively addresses occlusion (highlight in red boxes): in its absence, overlapping subjects (e.g., the pom-pom on the red hat of the left woman in $1^{\text{st}}$ row) completely disappear, and identity preservation deteriorates noticeably for occluded subjects.
• Figure I: Failure sample in complex reasoning scenarios. LayerComposer struggles when strong spatial reasoning is required. In this example, the subjects fail to be sitted naturally in the chairs, leading to unrealistic, copy-paste-like compositions that closely resemble the input images.