InnoAds-Composer: Efficient Condition Composition for E-Commerce Poster Generation
Authors: Yuxin Qin, Ke Cao, Haowei Liu, Ao Ma, Fengheng Li, Honghe Zhu, Zheng Zhang, Run Ling, Wei Feng, Xuanhua He, Zhanjie Zhang, Zhen Guo, Haoyi Bian, Jingjing Lv, Junjie Shen, Ching Law
Organizations: JD.com, Inc., Beijing, China · Chongqing University of Posts and Telecommunications, Chongqing, China · The Hong Kong University of Science and Technology, Hong Kong, China · Zhejiang University, Hangzhou, China
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.
Figures & tables
Figure 2: Case Examples and Dataset Construction Pipeline for E-commerce Poster Generation.
Figure 3: Overview of InnoAds-Composer. The framework comprises three modules: (1) Multi-Condition Tokenization , which maps heterogeneous controls into a shared token space and aligns them with the MM-DiT backbone; (2) Importance-Aware Condition Injection , which routes each control to its importance layers to improve efficiency while preserving controllability; and (3) Decoupled Attention , which allows the main stream to attend to condition cues while the condition branch performs self-attention only, removing the extra path to reduce cost and maintain training–inference consistency.
Figure 4: The importance heatmaps of the three conditions across timesteps and layers.
Figure 5: Qualitative results. Left: Input conditions, including C1-style images, C2-glyph images, and C3-subject images. Right: Results generated by different methods.
Models
Visual Text Quality
Subject Consistency
Style Consistency
General Quality
Sen. Acc ↑
NED ↑
DINO ↑
IoU ↑
CSD ↑
CLIP-I ↑
IR-Score ↑
FID ↓
Flux-inpaint [ 28 ]
-
-
0.853
0.877
-
-
0.822
79.41
Flux-Fill [ 28 ]
-
-
0.867
0.932
-
-
0.950
75.12
Flux-Kontext [ 28 ]
-
-
0.831
0.793
0.573
0.470
0.955
76.76
OmniControl2 [ 52 ]
-
-
0.812
0.685
0.692
0.477
0.930
80.13
OmniGen2 [ 64 ]
-
-
0.820
0.822
0.685
0.433
0.945
81.83
Table 1: Quantitative results. The best scores are bolded , while the second-best is underlined . Models marked with “*” in the table indicate closed-source models, while “-” denotes that the corresponding metric cannot be computed for the images generated by that model.
Figure 6: Comparison of image generation quality under different condition token pruning strategies.
Methods
Latency(s)
FLOPs(T)
Memery(G)
Flux-Kontext [ 28 ]
76.02
218.45
55.29
Ours (Stage I)
55.87
165.56
39.71
Ours (Stage II)
47.32
135.25
39.41
Table 2: Efficiency analysis across different training stages.
Figure 7: Ablation study of Text Feature Enhancement Module. Zoom in for details.
Figure 8: Additional qualitative results generated by our method.
Figure 9: Qualitative results under different condition token pruning ratios.
Pruning Ratio ( % )
NED ↑
MSE ↓
CLIP-I ↑
[0,0,0]
0.976
0.056
0.582
[50,20,30]
0.971
0.057
0.581
[70,40,50]
0.970
0.057
0.562
[80,50,60] (ours)
0.969
0.058
0.594
[90,60,70]
0.582
0.066
0.451
Table 3: Comparison of Stage II performance under different token pruning ratios. In the pruning ratio column, [x, y, z] denote the token pruning ratios for the glyph, subject, and style conditions, respectively.
Wuhan University · State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University · The Hong Kong University of Science and Technology (Guangzhou)