<h1 id="heading-207">What is PanGu Draw SDXL V3？</h1><p></p><p>“PanGu-Draw” is a novel latent diffusion model for text-to-image synthesis that adeptly accommodates multiple control signals: a diffusion-based generative model trained by HUAWEI.<a target="_blank" rel="ugc" href="https://pangu-draw.github.io">(learn more)</a></p><p></p><h1 id="heading-207">User Guide</h1><p></p><ol><li><p>This model and its running program are licensed under Apache 2.0. Please use this model and program in compliance with the license.</p></li><li><p>This model consists of two sub-models, one is the high_timestep model and the other is the low_timestep model. The running program needs to use both models simultaneously.</p></li><li><p>The running program of this model is still under development. It can only implement the text-to-image function for now. Please wait for further development if you need more functions.</p></li><li><p>This model has 5B parameters, and the running program is poorly optimized. Please make sure you have at least 20G of VRAM to run this program.</p></li><li><p>Comfyui and webui <strong>do not support</strong> this model，So I created a comfyui node to run this model(<a target="_blank" rel="ugc" href="https://github.com/ultranationalism/ComfyUI_Pangu_Draw_V3">here</a>)</p></li></ol><p></p><h3 id="heading-208">PanGu Draw 3.0 PanGu Draw 3.0 models implemented with MindSpore.I migrated it from MindSpore to Torch and implemented it to run on ComfyUI.</h3><p><a target="_blank" rel="ugc" href="https://github.com/mindspore-lab/mindone/tree/master/examples/pangu_draw_v3">Original Github rope</a></p><p><a target="_blank" rel="ugc" href="https://github.com/ultranationalism/ComfyUI_Pangu_Draw_V3">comfyui nodes</a></p><h3 id="heading-209">Features</h3><p>In contrast to version 2.0, Pangu Draw 3.0 has been subject to experimentation and updates across various aspects, including multi-language support, diverse resolutions, improved image quality, and model scaling. This includes:</p><ul><li><p>The current industry's largest 5-billion-parameter Chinese text-to-image model.</p></li><li><p>Supports bilingual input in both Chinese and English.</p></li><li><p>Supports output of native 1K resolution images.</p></li><li><p>Outputs images in multiple size ratios.</p></li><li><p>Quantifiable stylized adjustments: cartoon, aesthetic, photography controller.</p></li><li><p>Based on Ascend+MindSpore for large-scale training and inference, using a self-developed MindSpore platform and Ascend 910 hardware.</p></li><li><p>Utilizes self-developed RLAIF to enhance image quality and artistic expression.</p></li></ul><h3 id="heading-210">Author's Note:</h3><pre><code>This comfyui plugin is a program that I developed as a side project while learning stable diffusion. I am not a professional, and the program may have bugs. I apologize for my lack of technical skills.</code></pre>

Files with SHA256: 2c5a03a517341f6a915c58d114f9ad149b65bcf7e1480ece2036ea001d74f7df