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

E-commerce product poster generation aims to automatically synthesize a single image that effectively conveys product information by presenting a subject, text, and a designed style. Recent diffusion models with fine-grained and efficient controllability have advanced product poster synthesis, yet they typically rely on multi-stage pipelines, and simultaneous control over subject, text, and style remains underexplored. Such naive multi-stage pipelines also show three issues: poor subject fidelity, inaccurate text, and inconsistent style.

To address these issues, we propose InnoAds-Composer, a single-stage framework that enables efficient tri-conditional control tokens over subject, glyph, and style. To alleviate the quadratic overhead introduced by naive tri-conditional token concatenation, we perform importance analysis over layers and timesteps and route each condition only to the most responsive positions, thereby shortening the active token sequence. Besides, to improve the accuracy of Chinese text rendering, we design a Text Feature Enhancement Module (TFEM) that integrates features from both glyph images and glyph crops.

To support training and evaluation, we also construct a high-quality e-commerce product poster dataset and benchmark, which is the first dataset that jointly contains subject, text, and style conditions. Extensive experiments demonstrate that InnoAds-Composer significantly outperforms existing product poster methods without obviously increasing inference latency.

E-commerce product poster generation has emerged as a crucial task that aims to automatically synthesize a single image effectively conveying product information through the integration of subject, text, and a designed style. Recently, diffusion models have demonstrated fine-grained and efficient control over image synthesis; however, e-commerce poster generation remains relatively underexplored. Unlike free-form artistic layouts, retail posters must follow strict layout and branding rules while maintaining subject fidelity, style consistency, and text accuracy.

Current e-commerce poster generation systems still fall short in three respects. First, most do not offer end-to-end joint control of background style, subject fidelity, and text accuracy within a single model; multi-stage pipelines that compose the scene and render the text tend to be inaccurate, resulting in style inconsistency and loss of subject fidelity. Second, emerging single-stage approaches incorporate text control but struggle to render complex scripts and small glyphs with high fidelity. Third, the designed background style is often prompt-driven and may deviate from global style or semantic constraints. These limitations are exacerbated by the scarcity of training data with fine-grained, multi-condition annotations.

To address these challenges, we introduce InnoAds-Composer, a single-stage, multi-condition framework built on an MM-DiT backbone. A unified tokenization maps style, subject, and glyph conditions into the same token space. We propose TFEM with entire-glyph VAE tokens and single-glyph OCR crops with positional cues; conduct layer and timestep importance analysis for importance-aware injection; and use decoupled attention to remove costly low-value interactions. We also create InnoComposer-80K and InnoComposer-Bench.

We construct a high-quality bilingual e-commerce poster dataset through a structured pipeline. Each sample comprises three customized inputs: a background-style image conveying global aesthetics, a subject image specifying the product and its approximate location, and a glyph image providing the textual content to be rendered.

Images are grouped by SKU; annotators select visually similar pairs within each group. For one image in each pair, Qwen-Edit removes product content to obtain a clean style reference. The reference and its paired image form a diptych to fine-tune Qwen-Image, yielding clean background references. We synthesize style images in batches; unlike pipelines enforcing tight pixel-level match, our backgrounds are semantically aligned yet intentionally differ in local details. Glyph images (Chinese and English) are extracted via OCR; subjects are segmented with Grounded-SAM. After de-duplication and resolution normalization, we construct InnoComposer-80K (80,000 samples: prompt, subject, style, glyph). For evaluation, InnoComposer-Bench contains 300 held-out items ranked by product emphasis, style consistency, and text accuracy.

InnoAds-Composer adopts unified multi-condition tokenization on an MM-DiT backbone. A T5 encoder embeds the user prompt into text tokens $h^p$; the VAE encodes noise latents into patch tokens $h^z$.

Background style (Eq.1): $h^b = \mathcal{C}(h^p)$ if no style image ($m{=}0$); else $h^b = \mathcal{C}(h^i, h^{p_0})$ where $h^{p_0}$ is a fixed anchor prompt.

Subject: non-subject regions are filled with black; the masked image is VAE-encoded and patchified into $h^s$.

Glyph + TFEM (Eq.2): branch 1 encodes the entire glyph image → $h^{c1}$; branch 2 processes single-glyph OCR crops with absolute position, font-size code, and local position → $h^{c2}$; $h^c = \mathrm{GlyphEnc}(h^{c1}, h^{c2})$.

Importance-Aware Injection: guided by importance curves, we retain by default 40% of style tokens, 50% of subject tokens, and 20% of glyph tokens. In selected layers, corresponding condition tokens are concatenated with noisy-latent tokens $h^z$; non-selected tokens are omitted at that layer.

Decoupled Attention (Eq.3–5): $O_n = \mathrm{Attn}(Q_n, [K_n; K_{ci}], [V_n; V_{ci}])$; $O_{ci} = \mathrm{Attn}(Q_c, K_{ci}, V_{ci})$ — removing the $Q_c \!\to\! K_n$ pathway. Condition activations at each block can be computed once and cached across timesteps during inference.

Two-Stage Training: Stage I retains all condition tokens to train a fully conditioned generator; Stage II removes selected tokens and fine-tunes, sampling timesteps in proportion to the global importance map.

We measure condition importance via attention preference weights. At layer $b$ and timestep $t$, $A^{(b,t)} = \mathrm{Softmax}(Q^{(b,t)} {K^{(b,t)}}^\top / \sqrt{d})$. For each condition $ci$, we extract its subgraph, apply the condition mask, and compute the mean to obtain $S_{ci}(b,t)$.

From the heatmaps (Fig.4): background style dominates in early layers and early timesteps but decays rapidly. The subject condition forms a persistent high-intensity band in mid-to-deep layers across most timesteps. The glyph condition gradually increases in intensity in middle layers and later timesteps, corresponding to stroke refinement. The conditions display a non-uniform and complementary relationship across timesteps and model depth.

Table presents quantitative comparison. In Stage I, our method achieves the best performance in nearly all aspects: Sen. Acc 0.857, NED 0.976, DINO 0.923, IoU 0.972, CSD 0.729, IR-Score 1.036, FID 54.39. After Stage II, performance remains competitive (Sen. Acc 0.847, DINO 0.914, FID 55.24) while improving efficiency via importance-aware token routing.

Leveraging decoupled attention, Stage I reduces latency by 26.5% and FLOPs by 24.2% vs Flux-Kontext. Stage II achieves further reductions of 37.8% in latency and 38.1% in FLOPs without compromising generation quality.

Flux-Fill and USO cannot generate visual text, making them unsuitable for poster scenarios requiring integrated textual elements. PosterMaker generates subject and text but struggles to maintain background style consistency. Qwen-Image-Edit and Seedream 4.0 show good subject consistency and text rendering, yet often produce redundant or mismatched text, and style transfer tends to exhibit a copy-and-paste effect. InnoAds-Composer achieves strong consistency across text, subject, and background style.

Ablation on TFEM: without it, generated images exhibit noticeable textual errors; incorporating it improves rendered text quality, with an approximate 5% improvement in Sen. Acc (Fig.7).

We presented InnoAds-Composer, a single-stage, multi-condition framework for e-commerce poster generation that delivers simultaneously subject, text, and designed style. By unifying subject, style, and glyph tokens in the same space and preserving the priors of an MM-DiT backbone, our approach enables joint inference without multi-stage pipelines. The proposed TFEM fuses single/entire glyphs with positional information, substantially improving text sharpness, boundary integrity, and readability. Complementing this, we perform importance-aware condition injection and decoupled attention to reduce redundant interactions and shorten the activated sequence, leading to improved inference efficiency. We constructed InnoComposer-80K and InnoComposer-Bench for learning and fair comparison.