Source-linked AI summary
LaDe: Unified Multi-Layered Graphic Media Generation and Decomposition
Vlad-Constantin Lungu-Stan, Ionut Mironica, Mariana-Iuliana Georgescu
TL;DR
Existing media-design methods limit layer counts or constrain layers to spatially continuous regions, restricting flexible editable composition. LaDe uses prompt expansion, latent diffusion with 4D RoPE, and an RGBA VAE to generate full designs and layers jointly. It supports text-to-image, text-to-layers, and decomposition, achieving state-of-the-art text-to-layers performance and competitive decomposition quality, while relying on stochastic LLM planning and substantial VRAM for many layers.
Problem
Existing layered-design methods use fixed layer counts or spatial-continuity constraints, limiting flexible generation of semantically meaningful editable layers.
Method
LaDe combines LLM prompt expansion, a latent diffusion transformer with 4D RoPE, and an RGBA VAE to jointly generate full designs and constituent layers.
Results
LaDe achieves state-of-the-art text-to-layers performance and competitive decomposition quality, with better decomposition for 2 and 3 layers and a PSNR of 32.65 for 2 layers.
Takeaways & Limitations
A single unified model supports text-to-image generation, text-to-layers design generation, and media-design image decomposition with flexible layer counts.
Takeaways & Limitations
LaDe depends on stochastic LLM prompt expansion and can require high VRAM when generating many layers, restricting scalability on limited-GPU systems.
Abstract
from arXiv · showhide
Media design layer generation enables the creation of fully editable, layered design documents such as posters, flyers, and logos using only natural language prompts. Existing methods either restrict outputs to a fixed number of layers or require each layer to contain only spatially continuous regions, causing the layer count to scale linearly with design complexity. We propose LaDe (Layered Media Design), a latent diffusion framework that generates a flexible number of semantically meaningful layers. LaDe combines three components: an LLM-based prompt expander that transforms a short user intent into structured per-layer descriptions that guide the generation, a Latent Diffusion Transformer with a 4D RoPE positional encoding mechanism that jointly generates the full media design and its constituent RGBA layers, and an RGBA VAE that decodes each layer with full alpha-channel support. By conditioning on layer samples during training, our unified framework supports three tasks: text-to-image generation, text-to-layers media design generation, and media design decomposition. We compare LaDe to Qwen-Image-Layered on text-to-layers and image-to-layers tasks on the Crello test set. LaDe outperforms Qwen-Image-Layered in text-to-layers generation by improving text-to-layer alignment, as validated by two VLM-as-a-judge evaluators (GPT-4o mini and Qwen3-VL).
2. Related Work
Diffusion-based design systems have progressed from flat-image generation toward layered representations, but existing approaches remain limited in controllability, layer count, or spatial organization. LaDe addresses these constraints with flexible, unified layered generation from short text prompts.
- Diffusion image-editing methods still operate on flat representations, preventing isolation and manipulation of individual compositional elements.
- Layer decomposition enables classic document-editing operations by separating an image into independently manipulable components.
- LaDe generates layered designs directly from text prompts, unlike decomposition-only approaches.
- LaDe supports a flexible number of layers, groups related elements despite spatial separation, and avoids ART’s external bounding-box planner.
- LaDe unifies media design generation, layered generation, and image decomposition in one end-to-end model.
3. Method
LaDe is a unified latent-diffusion framework that jointly generates editable RGBA media designs and their layers, while supporting generation and decomposition. It combines prompt expansion, layer-aware diffusion modeling, RGBA decoding, and training strategies for flexible layered outputs.
- Overall System: LaDe jointly generates a full media design and its n constituent RGBA layers, with text-to-image obtained by setting n to 0.The full design is the first generated image; remaining layers compose it through alpha blending.
- RGBA VAE: The RGBA VAE encodes four-channel images into latent embeddings and decodes them back to RGBA outputs using RGB, alpha, and perceptual losses.The loss weights RGB and alpha reconstruction separately and applies LPIPS to a gray-alpha-blended RGB version.
- Prompt Processing: Prompt expansion structures the input into scene, per-layer, and style descriptions, providing precise layer content while leaving layout decisions to the model.The expanded prompt is encoded with FlanT5 XXL before conditioning diffusion.
- Diffusion Model: The diffusion model combines prompt and image-layer embeddings, then uses full attention to denoise visual information while treating text as conditioning.Inputs are aligned through a linear adapter before concatenation.
- Diffusion Model: LaDe uses 4D RoPE over spatial coordinates, layer index, and token role to link prompt segments with their target layers and distinguish generation from decomposition.The role coordinate marks prompt, denoisable, and frozen tokens differently.
- Diffusion Model: Training randomly treats layers as outputs or input conditions, enabling generation and decomposition within the same model.Denoising is disabled for conditioned layers during decomposition by setting their timestep to zero and marking them non-denoisable.
4. Experiments
LaDe is evaluated on layered media generation and decomposition using Crello data, VLM-as-a-judge scores, pixel-level metrics, ablations, and qualitative comparisons. It achieves the best reported text-to-layers performance across tested layer counts and improves layer organization through variable-layer training and Eq. 2.
- Experimental setup: The experiments use layered media samples with meaningful grouped elements and evaluate generation with GPT-4o mini and Qwen3-VL judges.The Crello test set is used for comparisons, while decomposition additionally uses PSNR and RGB L1.
- Qualitative evaluation: Qualitatively, LaDe produces homogeneous RGBA layers and meaningful decompositions, while competing methods duplicate text or hallucinate and fail to reconstruct occluded content.LaDe’s one-step generation groups similar information into structured layers and avoids duplicated content in decomposition.
- Media design generation: LaDe obtains the best VLM-as-a-judge performance for media design layer generation across 2, 3, 4, and 5 requested layers.The comparison is against Qwen-Image-T2I plus Qwen-Image-Layered-I2L on the Crello test set.
- Media design generation: Qwen-Image-T2I plus Qwen-Image-Layered-I2L remains around 2.6-2.8 with GPT-4o mini and 2.3-2.5 with Qwen3-VL across tested layer counts.Both judges show the same trend favoring LaDe’s qualitative layered media designs.
- Media design decomposition: For decomposition, LaDe outperforms Qwen-Image-Layered-I2L on 2- and 3-layer cases, including a PSNR score of 32.65 for 2 layers.RGB L1 favors LaDe for 2 and 3 layers but Qwen-Image-Layered-I2L for 4 and 5 layers; Qwen’s Crello training makes its evaluation in-distribution.
- Ablations: Variable-layer training improves the ablation score from 3.78 to 3.85, while Eq. 2 improves layer-prompt adherence by reducing duplicated information across layers.RGBA VAE effects are especially visible for shadows and soft alpha-blended edges, even when VLM evaluation changes little.
5. Conclusion
LaDe is a unified latent diffusion framework for generating layered media designs and single images from short prompts, while also supporting media design decomposition. It achieves state-of-the-art text-to-layers performance and competitive decomposition quality, with limitations from stochastic prompt expansion and high VRAM use.
- Conclusion: LaDe combines LLM prompt expansion, a Latent Diffusion Transformer with 4D RoPE, and an RGBA VAE to generate flexible semantically meaningful layers.The framework also generates single images and supports media design decomposition.
- Conclusion: LaDe achieves state-of-the-art text-to-layers performance and competitive decomposition quality under VLM and pixel-level evaluations.Text-to-layers is assessed with GPT-4o mini and Qwen3-VL-30B-A3B-Instruct; decomposition uses pixel-level metrics and VLM-as-a-judge.
- Limitations: LaDe’s prompt expansion depends on stochastic LLM outputs, and generating many layers can consume substantial VRAM.These constraints may affect prompt quality and scalability on systems with limited GPU memory.
6. Implementation details
LaDe extends a converged text-to-image model into layered media processing through multi-phase training and conditioning on layer inputs. Its implementation includes random layer conditioning, optional automated layer-count prediction, and training across multiple task modalities.
- Training: LaDe is initialized from an already converged text-to-image model to accelerate convergence for layered media tasks.The stated focus is layered media design processing as an extension of standard image generation.
- Training: Multi-phase training allocates two-thirds of GPUs to the layered tasks while retaining a scaled legacy image-generation task.The first phase trains designs, images, and vectors equally for 70k steps, with layer conditioning weighted equally in layered training.
- Training: Image conditioning is sampled with 30% probability using a random number of input layers between one and N-1.The full media design is forced as the first condition 30% of the time to represent decomposition training.
- Prompt expansion: LaDe’s prompt expansion uses GPT-4o mini, although the number of layers must be set in advance for text-to-layers generation.The paper suggests an LLM could automate layer-count prediction from the user input, aspect ratio, and media-design type.
7. Qualitative Evaluation
The qualitative evaluation section points to additional LaDe text-to-image, text-to-layers, and image-to-layers examples in Figures 2–4. These examples supplement the reported quantitative and qualitative comparisons.
- Qualitative examples: Additional LaDe text-to-image, text-to-layers, and image-to-layers samples are presented in Figures 2, 3, and 4.The section directs readers to further qualitative examples across all three supported generation or decomposition tasks.
8. Prompts
This section presents the prompts used for Figures 11–15 and provides structured design-description instructions for organizing designs by layers.
- The prompts used in Figures 11, 12, 13, 14, and 15 are presented in this section.
- Design descriptions should begin with an overall concept before listing individual layers.
- Each design description should use a clear layered format, with every layer presented separately.