How LoRA Supercharges AI‑Generated Seasonal Poetry Posters

1. Introduction

AIGC tools have revolutionized the design industry by dramatically improving creation efficiency and turning designers' ideas into creative images, but current AI‑generated images suffer from randomness and homogeneity. Training a proprietary LoRA model is proposed to improve image quality, precise prompt expression, and stylistic stability.

2. Project Exploration

2.1 Project Background

The AI‑native product "Wen Xiaoyan" includes many AI features. In the already launched "24 Solar Terms – Poem to Paint" bot, the goal is to generate elegant acrostic‑poem posters that match each solar term. Previously only the poem was AI‑generated while the background required manual design. By applying LoRA‑based image generation, the team aims to create ink‑wash style posters for all 24 solar terms, from the beginning of spring to the end of winter.

2.2 Gameplay Innovation

Integrating LoRA into the bot enables rapid generation of multiple posters with a consistent style, greatly increasing creative variety and enhancing the AI‑driven gameplay experience. A configuration platform allows simple seasonal prompt changes, streamlining the setup process. The accumulated training set forms a complete ink‑wash style system that satisfies the LoRA training requirements.

2.3 Clear Direction

Maintain traditional ink‑wash style : Ensure LoRA emphasizes atmospheric brush strokes and poetic ambience, immersing viewers in a scroll‑like landscape.

Emphasize top blank space : Preserve blank areas typical of ink‑wash art to accommodate acrostic poems, creating a harmonious visual‑textual blend.

Support flexible seasonal switching : Enable a single LoRA model to adapt to all 24 solar terms by understanding seasonal aesthetics and applying targeted training strategies.

2.4 Model Training

After defining the training directions, the team launched the LoRA training pipeline, addressing challenges step by step.

1) Dataset Completion

Since the existing seasonal images did not cover all 24 terms, missing images were supplemented using Midjourney with prompts emphasizing ink‑wash ambience and blank space, producing concise, high‑quality visual cues.

2) Dataset Processing

Because AI generation is highly random, the team refined the training images in Photoshop, adjusting layout and creating description templates that highlight the blank area, which is essential for both ink‑wash aesthetics and poem placement.

3) Training and Evaluation

Season‑aware classification of the training set was applied to avoid feature contamination between seasons. Adjusted LoRA parameters and high‑quality prompt templates enabled the model to generate images that faithfully reflect seasonal characteristics.

4) Deploying Generation

The final step combined the AI‑generated images with AI‑crafted acrostic poems, producing a series of culturally rich seasonal posters that showcase traditional Chinese aesthetics through modern technology.

3. Benefits

Integrating LoRA‑driven image generation into the "24 Solar Terms – Poem to Paint" bot achieves fully AI‑powered visual creation, attracting user attention, stimulating creative interaction, and enhancing the product’s AI attributes and brand influence. Ongoing LoRA updates will continue to improve the image creation experience.