Summary
The AI Search channel puts Ernie Image — a newly released open-source image generator from Baidu — head-to-head against Zage, the previous leading open-source model, across a diverse battery of prompts. Tests cover photorealistic scene generation, complex multi-element compositions (recursive imagery, miniature cityscapes, crowded interior scenes), in-image text rendering with multi-sentence passages, and a range of artistic styles including impressionist oil painting, Chinese watercolor, minimalist flat illustration, and stipple dot art.
Ernie Image outperforms Zage in most categories, particularly in prompt adherence for long, detail-heavy descriptions and in text rendering accuracy — demonstrating correct reproduction of paragraph-length text within generated images, a historically weak point for diffusion models. Zage retakes the lead in anatomical accuracy: a challenging yoga-pose prompt revealed significant distortion from Ernie while Zage handled the pose correctly. The video is candid about both models’ shared weakness in generating highly abstracted painterly styles.
The host also walks through the local installation process for Ernie Image, enabling offline, unlimited generation without API costs — a practical consideration for developers and creatives evaluating open-source alternatives to commercial tools like Midjourney or DALL-E 3. A sponsored segment covers Gamma Imagine, a new AI-powered design tool that converts text prompts into fully formatted posters, infographics, and social media visuals. For anyone tracking the current frontier of local open-source image generation, this side-by-side provides a practical reference across real-world use cases.
📺 Source: AI Search · Published April 16, 2026
🏷️ Format: Comparison
