Semantic Packet Aggregation and Repeated Transmission for Text-to-Image Generation

Seunghun Lee, Jihong Park, Jin‐Ho Choi, Hyuncheol Park

Abstract

Text-based communication is expected to be prevalent in 6G applications such as wireless AI-generated content (AIGC). This paper addresses the challenges of transmitting text prompts over erasure channels for a text-to-image AIGC task by developing the semantic segmentation and repeated transmission (SMART) algorithm. SMART groups words in text prompts into packets, prioritizing the task-specific significance of semantics within these packets, and optimizes the number of repeated transmissions. Simulation results show that SMART achieves higher similarities in received texts and generated images compared to a character-level packetization baseline, while reducing computing latency by orders of magnitude compared to an exhaustive search baseline.

Introduction: The Challenge of Semantic Communication

In traditional wireless communication, every bit is typically treated with equal priority. However, as 6G networks move toward supporting Large Language Models (LLMs) and AI-generated content (AIGC)—such as sending a prompt from a mobile device to a remote Stable Diffusion server—this "bit-neutral" approach becomes highly inefficient. Over an unreliable erasure channel, losing a critical word like "tiger" completely changes the output, whereas losing a filler word like "a" might have no impact at all.

Overview of the SMART framework for semantic text-to-image communication, illustrating the flow from original image to generated image through a lossy channel.

Existing packetization strategies focus on optimizing packet length and sequential data, ignoring the underlying meaning of the payload. The authors propose a shift toward "Semantic Communication," where the goal is to package and transmit text prompts so that the core semantics survive packet loss, ensuring the receiving AI model generates an image faithful to the original intent.

The SMART Algorithm

The proposed SMART (Semantic Packet Aggregation and Repeated Transmission) algorithm introduces a two-pronged approach to semantic reliability:

  1. Semantic Packet Aggregation (SemPA): Instead of character-by-character splitting, SemPA groups words into packets to maximize the expected semantic similarity between the original and received prompts. It utilizes a pre-trained CLIP model to measure the cosine similarity of text features. To avoid over-concentrating importance in a single packet—which would create a single point of failure—a synergy baseline score balances semantic distribution across multiple packets.
  2. Semantic Repeated Transmission (SemRT): To combat loss, packets are retransmitted. However, finding the optimal retransmission count for each packet is a combinatorial problem that scales factorially. To solve this in real-time, the authors trained a lightweight neural network to predict optimal transmission counts based on semantic scores, bypassing the exhaustive search bottleneck.

Experimental Evaluation

The system was evaluated using 100 sentences from the MS-COCO dataset. CLIP was used to measure text and image similarity, while Stable Diffusion v2.1 generated images from the decoded prompts at the receiver end.

Key Results

The SMART algorithm demonstrated significant improvements over standard baselines:

  • Text Similarity: At a high packet error rate ($p = 0.3$), SMART achieved over 6x higher text cosine similarity compared to character-level transmission.
  • Image Quality: SMART resulted in a 48.3% higher image cosine similarity. Qualitative results showed SMART successfully preserving the "tiger" concept while the baseline produced corrupted, distorted images due to character-level errors.
  • Computational Efficiency: The ML-based SemRT method achieved 99.4% of the performance of an exhaustive search while operating up to $5 \times 10^5$ times faster (0.87 ms vs. 430 seconds for a 12-word prompt).
Quantitative performance comparison showing SMART's superior cosine similarity between texts compared to character-level transmission across various packet decoding error rates.

Conclusion and Future Directions

The SMART framework proves that prioritizing semantics over sequential bits leads to more robust AIGC applications in lossy environments. By treating text as a collection of meaningful concepts rather than just characters, the system achieves graceful degradation under network stress. Future work will look into moving beyond fixed dictionaries toward native tokenization used by modern LLMs to further align transmission strategies with generative model architectures.

Summary updated

Abstract

Text-based communication is expected to be prevalent in 6G applications such as wireless AI-generated content (AIGC). Motivated by this, this paper addresses the challenges of transmitting text prompts over erasure channels for a text-to-image AIGC task by developing the semantic segmentation and repeated transmission (SMART) algorithm. SMART groups words in text prompts into packets, prioritizing the task-specific significance of semantics within these packets, and optimizes the number of repeated transmissions. Simulation results show that SMART achieves higher similarities in received texts and generated images compared to a character-level packetization baseline, while reducing computing latency by orders of magnitude compared to an exhaustive search baseline.
BibTeX
@misc{lee2025semantic,
  author = {Seunghun Lee and Jihong Park and Jin‐Ho Choi and Hyuncheol Park},
  title = {{Semantic Packet Aggregation and Repeated Transmission for Text-to-Image Generation}},
  year = {2025},
  eprint = {2503.23734},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.23734},
  howpublished = {arXiv},
}
Figure