InnoAds-Composer: Efficient Condition Composition for E-Commerce Poster Generation

The paper proposes InnoAds-Composer, a single-stage framework that efficiently generates e-commerce posters by integrating subject, text, and style controls through an optimized token routing mechanism and a text enhancement module, while also introducing a new high-quality dataset and benchmark for this task.

The Gist

Imagine you are a graphic designer working for a massive online store like Amazon or JD.com. Every day, you need to create thousands of posters to sell different products. Each poster needs three things to be perfect:

1. The Product: A crisp, clear image of the item (like a pair of shoes or a bottle of perfume).
2. The Background: A beautiful, stylish scene that matches the product's vibe (like a sunny beach for sunscreen or a cozy living room for a lamp).
3. The Text: The product name and a catchy slogan, written perfectly in the right font and language (including tricky Chinese characters).

The Problem:
Until now, doing this with AI was like trying to build a house by hiring three different contractors who don't talk to each other.

• One contractor builds the house (the background).
• Another paints the furniture (the product).
• A third tries to write the address on the mailbox (the text).

Because they don't coordinate, the result is often a mess: the text looks like gibberish, the product looks blurry, or the background style clashes with the product. Also, trying to get all three to work together at once makes the computer so slow it takes forever to generate a single poster.

The Solution: InnoAds-Composer
The researchers at JD.com built a new AI system called InnoAds-Composer. Think of this system as a super-conductor who can orchestrate a symphony of three different instruments (Style, Product, and Text) simultaneously, perfectly in tune, and at lightning speed.

Here is how it works, using some simple analogies:

1. The "One-Stop-Shop" Approach (Single-Stage Framework)

Old methods were like a relay race where the baton gets dropped between runners. InnoAds-Composer is like a solo artist who can paint the background, place the product, and write the text all in one single, fluid motion. This ensures everything fits together perfectly from the very first brushstroke.

2. The "Smart Text Teacher" (Text Feature Enhancement Module)

Writing text in images, especially in Chinese, is notoriously hard for AI. It often looks like a child's scribble.

• The Old Way: The AI tried to guess the shape of the letters based on a blurry picture.
• The New Way: The system uses a special "Text Feature Enhancement Module" (TFEM). Imagine this as a two-teacher classroom:
  • Teacher A looks at the whole word to understand the general shape and flow.
  • Teacher B zooms in on every single letter, checking its exact position, size, and local details.
  • They combine their notes to ensure every character is sharp, clear, and spelled correctly, even if it's a complex Chinese character.

3. The "Smart Traffic Controller" (Importance-Aware Injection)

This is the magic trick that makes the system fast.
Usually, when you ask an AI to do three things at once, it tries to process all the instructions for every single step of the creation process. It's like a chef trying to chop onions, stir the soup, and bake a cake simultaneously at full speed the whole time. It's exhausting and wasteful.

InnoAds-Composer acts like a smart traffic controller:

• It analyzes the process and realizes: "Hey, the background style is most important at the beginning, but the text details only matter at the very end."
• So, it turns off the background instructions once the style is set, and turns off the text instructions until the final touches.
• By only listening to the instructions that matter right now, the computer doesn't have to do unnecessary work. This cuts the processing time significantly without losing quality.

4. The "Decoupled Attention" (The Quiet Sidekick)

In normal AI, the "instructions" (like the text or style) constantly chat with the "image being made," which creates a lot of noise and confusion.
InnoAds-Composer separates them. Imagine the Image is the main actor on stage, and the Instructions are the director in the booth.

• The actor listens to the director.
• But the director doesn't need to listen to the actor's every move; they just give their notes once and let the actor work.
• This "one-way street" saves a massive amount of energy and memory, making the system run much faster.

The Result

The team didn't just build the engine; they also built a giant training library (80,000 examples) specifically for e-commerce posters, teaching the AI exactly what a "good" poster looks like.

In summary:
InnoAds-Composer is a smart, fast, and single-step AI that can take a product photo, a style idea, and some text, and combine them into a professional, high-quality advertisement poster. It fixes the messy text, keeps the product looking real, and does it all much faster than previous methods by only paying attention to the instructions that matter at the exact moment they are needed.

Technical Summary

1. Problem Statement

E-commerce product poster generation requires synthesizing a single image that integrates three distinct elements: a subject (product), text (glyphs/promotional copy), and a designed background style. While recent diffusion models have advanced image synthesis, existing solutions face three critical limitations in this specific domain:

• Pipeline Fragmentation: Most systems rely on multi-stage pipelines (e.g., generating the scene first, then adding text), which leads to style inconsistency, loss of subject fidelity, and inaccurate text rendering.
• Text Rendering Deficiencies: Single-stage approaches often struggle to render complex scripts (especially Chinese characters) and small glyphs with high fidelity, resulting in blurry or illegible text.
• Computational Inefficiency: Simultaneously controlling subject, text, and style via naive token concatenation in diffusion transformers (DiT) introduces quadratic computational overhead due to the increased sequence length, making inference slow and resource-intensive.
• Data Scarcity: There is a lack of high-quality datasets that jointly provide fine-grained annotations for subject, text, and style conditions.

2. Methodology: InnoAds-Composer

The authors propose InnoAds-Composer, a single-stage framework built on a Multi-Modal Diffusion Transformer (MM-DiT) backbone. It addresses the challenges through three core innovations:

A. Unified Multi-Condition Tokenization

The framework maps heterogeneous inputs (style images, subject images, and glyph images) into a shared token space to enable joint inference:

• Style Control: Supports both text prompts and style images. Style images are encoded via a VAE and patchified into visual tokens.
• Subject Control: Uses a masked subject image (background filled with black) encoded by a VAE to ensure the product structure is preserved without background leakage.
• Text (Glyph) Control: Introduces a Text Feature Enhancement Module (TFEM) with a dual-branch architecture:
  1. Global Branch: Encodes the entire glyph image via VAE to capture visual structure.
  2. Local Branch: Processes single-glyph crops using an OCR backbone, injecting positional cues (absolute location, font size, local position).
  3. Fusion: A lightweight character encoder fuses these features to improve text sharpness, boundary integrity, and readability.

B. Importance-Aware Condition Injection

To mitigate the quadratic complexity of attention mechanisms caused by long token sequences, the authors analyze the influence of each condition across different layers and diffusion timesteps.

• Observation: They found non-uniform influence patterns:
  • Style: Dominates early layers and early timesteps.
  • Subject: Maintains high influence across mid-to-deep layers and most timesteps.
  • Glyph: Gains intensity in middle layers and later timesteps (refinement phase).
• Strategy: Instead of injecting all condition tokens at every step, the model routes each condition only to its "most responsive" layers and timesteps. This significantly shortens the active token sequence, reducing computational load while maintaining controllability.

C. Decoupled Attention Mechanism

Standard attention computes interactions between all tokens (condition and noisy latent), which is redundant as condition tokens evolve slowly.

• Optimization: The authors remove the cross-attention pathway from condition queries to noisy-latent keys ($Q_c \to K_n$).
• Result: The main stream attends to conditions, but the condition branch performs self-attention only. This allows condition features to be computed once and cached for all timesteps during inference, drastically reducing latency and memory usage.

D. Two-Stage Training Strategy

• Stage I: Trains the model with all condition tokens retained to learn robust priors.
• Stage II: Fine-tunes the network with the selected (pruned) tokens removed. Diffusion timesteps are sampled proportionally to their importance to align training with the inference strategy, minimizing performance degradation.

3. Key Contributions

1. InnoAds-Composer Framework: A single-stage, multi-condition generation model that achieves coordinated control over style, subject, and text without multi-stage pipelines.
2. Text Feature Enhancement Module (TFEM): A novel architecture that fuses global glyph structure with local semantic and positional cues, significantly improving Chinese and English text rendering accuracy.
3. Efficiency via Importance Analysis: The first method to reveal and leverage the non-uniform, complementary influence patterns of conditions across layers and timesteps, enabling selective token injection to curb quadratic complexity.
4. InnoComposer-80K & Bench: The creation of a new high-quality dataset (80k samples) and benchmark containing joint subject, glyph, and style conditions, filling a gap in e-commerce poster research.

4. Experimental Results

The method was evaluated against state-of-the-art baselines (including Flux, AnyText2, PosterMaker, and commercial models like Seedream 4.0) on the InnoComposer-Bench.

• Qualitative Performance: InnoAds-Composer produces posters with superior subject fidelity, consistent background styles, and highly accurate text rendering, outperforming models that suffer from "copy-paste" style effects or text hallucinations.
• Quantitative Metrics:
  • Text Quality: Achieved the highest Sentence Accuracy (0.857) and Normalized Edit Distance (0.976).
  • Subject Consistency: Led with a DINO score of 0.923 and IoU of 0.972.
  • Overall Quality: Achieved the best FID (54.39) and IR-Score (1.036).
• Efficiency Gains:
  • Compared to the baseline Flux-Kontext, Stage I reduced inference latency by 26.5% and FLOPs by 24.2%.
  • Stage II (with token pruning) further reduced latency by 37.8% and FLOPs by 38.1% while maintaining competitive generation quality.

5. Significance

This paper represents a significant step forward in commercial AI applications. By solving the "trilemma" of subject fidelity, text accuracy, and style consistency in a single efficient stage, InnoAds-Composer offers a practical solution for automated e-commerce design.

• Industrial Impact: It reduces the computational cost of generating high-quality marketing materials, making real-time or large-scale poster generation feasible.
• Scientific Contribution: The work provides a new paradigm for efficient conditional diffusion, demonstrating that "smart" token routing based on layer/timestep importance can outperform naive concatenation without sacrificing quality.
• Data Resource: The release of InnoComposer-80K and the benchmark sets a new standard for evaluating multi-modal generation tasks in e-commerce.