LingXi-Image-MoE 1.0 is a text-to-image diffusion model with 470 million parameters, released on July 14, 2026, by LingXi Qihang Intelligent Technology and Hugging Face user shyai. Its core innovation lies in its Full-Layer MoE design, which extends beyond applying Mixture-of-Experts (MoE) to only select blocks by implementing the MoE architecture across all layers. This can be understood as a mechanism that selects the most suitable expert for each input during processing from among multiple specialized generation pathways, akin to a studio that calls upon different specialist painters for each scene rather than having one painter handle all subjects. The model’s weights and the complete training and inference code have been made public, making it particularly suitable for research aimed at directly investigating model architecture and execution processes, rather than merely using the generated images. While many existing open-source image generation models provide only completed checkpoints and basic inference examples, they often do not fully disclose which experts were selected or how routing evolved during training. A key differentiator of LingXi-Image-MoE is that it provides both training logs and Expert routing logs. This offers a transparent experimental setup for observing internal decision-making pathways, rather than a "black box" approach limited to comparing generation quality. Furthermore, as a model collected for an independent single-GPU research project, it provides a distinct starting point for reproducibility studies compared to models designed for large-scale institutions. However, details such as available GPU types, actual training duration, dataset composition, specific routing algorithm formulas, and quality benchmarks cannot be verified from the current Discovery information alone; therefore, verification via the official repository and model card is required. Biotech researchers are advised to utilize this model not as a direct diagnostic tool, but as a foundation for research-oriented synthetic image generation and MoE structure analysis. For example, by inputting prompts describing microscope images or laboratory scenes and preserving both the generated results and Expert routing logs, researchers can investigate which words related to cell types, staining methods, or imaging conditions correlate with specific expert selections. Additionally, based on the publicly available training code and weights, researchers can evaluate the potential for fine-tuning on life science image data and compare expert activation patterns and output variations under fixed seeds and identical prompt conditions. Generated images must be clearly distinguished from actual observational data and subjected to expert review; currently, there is no evidence supporting their direct use in clinical judgment or quantitative biological analysis. From a research reproducibility perspective, the published training logs can serve as a baseline for analyzing routing bias, over-selection of specific experts, and path variations caused by changes in input representations. By repeatedly executing the same set of prompts and managing generated images, seeds, checkpoints, and routing logs together, researchers can create experimental materials for comparing the internal operations of MoE image models. However, specific installation commands, supported resolutions, inference steps, samplers, precision settings, and quantitative performance metrics are not included in the provided information; thus, the official README and Hugging Face model card should be consulted prior to conducting reproducibility experiments.

LingXi-Image-MoE 1.0 is a text-to-image diffusion model with 470 million parameters, collected and released on July 14, 2026, by LingXi Qihang Intelligent Technology and Hugging Face user shyai. Its core innovation lies in adopting a Full-Layer MoE design that extends beyond applying Mixture-of-Experts (MoE) to only select blocks, instead deploying the MoE architecture across all layers. Among multiple expert generation pathways, it selects the experts best suited to the input and processing stage.

🔬 Expert Routing Analysis by Life Science Prompt: Collect generation results and Expert routing logs using the same set of life science prompts and a fixed seed, then compare changes in expert selection based on expressions indicating cell, tissue, and staining conditions. Supported resolution and inference parameters need to be verified from the official documentation.