Description:

⚠️ EXPERIMENTAL MERGE / TEST BUILD

This model is a direct weight merge of Baidu's ERNIE-Image and Alibaba's Z-Image Turbo.

This is a proof of concept to test the compatibility of DiT architectures. Due to technical nuances of merging different DiT checkpoints, the model exhibits specific behaviors. It is not a polished final release.

🔥 Key Features:

• ⚡ Lightning Fast: Runs perfectly on 8 Steps (Turbo mode).

• 📝 Incredible Text Rendering: Inherits ERNIE's ability to render legible text (Cyrillic, Latin, Kanji) and follow complex instructions.

• 🎨 High Fidelity: Combines Z-Image Turbo's photorealism and lighting with ERNIE's composition logic.

⚙️ Recommended Settings:

• Sampler: Euler

• Scheduler: Beta / Simple (Beta schedule alpha: 0.6, beta: 0.6)

• Steps: 8 - 10

• CFG Scale: 1.0 (CRITICAL: Do not increase, or the image will degrade/burn)

• Resolution: 1024x1024 or 1024x1536

⚠️ Known Limitations & Quirks:

1. Facial Bias: The model inherits a strong Asian/East Asian prior from the ERNIE base weights.

   • Solution: To generate non-Asian faces, you MUST use explicit physical descriptors in your prompt (e.g., Caucasian man, blue eyes, African woman, dark skin, European features). Generic terms like "man" or "woman" will likely result in Asian faces.

2. Architecture Mismatch: This is a direct merge of weights, not a trained hybrid. Some layer alignments may be imperfect, leading to occasional inconsistencies in complex scenes.

🚀 Future Roadmap: We are currently analyzing the latent space alignment. The next version will likely utilize LoRA-based merging or adapter training to fix the architectural inconsistencies and neutralize the facial bias for a more universal model.

Please generate and share your results! We are looking for feedback on prompt adherence and image quality.