CADS3.0XL-DiT Workflow - Commercial Design and Advertising Assistance System v3

AiARTiST SDXL Commercial Design and Advertising Material WorkflowCFG++: Manifold-constrained CFG for Diffusion Models <h1 style="text-align: center;">Classifier Free Guidance</h1>Original Source: <a href="https://arxiv.org/html/2406.08070v1" target="_blank">https://arxiv.org/html/2406.08070v1</a>Classifier Free Guidance (CFG) is a fundamental tool for text-guided generation in modern diffusion models. Although CFG is effective, it also has notable drawbacks. For example, CFG-adapted DDIM lacks reversibility, leading to complex image editing; moreover, high guidance scales crucial for high-quality output often result in pattern collapse issues. Contrary to the common belief that these are inherent limitations of diffusion models, research reveals that these problems actually stem from non-manifold phenomena associated with CFG, rather than the diffusion models themselves. More specifically, inspired by the recent advances in the Diffusion Inversion Solver (DIS) based on diffusion models, the project team redefines text-guidance as an inverse problem with text-conditioned score-matching loss and develops CFG++, a new method that addresses non-manifold challenges inherent in traditional CFG. CFG++ features surprisingly simple fixes to CFG but offers significant improvements, including better text-to-image generation sample quality, reversibility, smaller guidance scales, reduced pattern collapse, and more. Additionally, CFG++ seamlessly interpolates between unconditional and conditional sampling at lower guidance scales, consistently outperforming traditional CFG at all scales. Experimental results confirm that our approach significantly enhances text-to-image generation, DDIM inversion, editing, and solving inverse problems, demonstrating the wide-ranging impact and potential applications of our method in various domains utilizing text guidance.<img src="https://liblibai-online.liblib.cloud/img/d6178593b5834f9fa487bca389ce5d83/7e4e8f9d2b6dcf0d19c1ab693869ad848ec0dd1b339071c357fa23e80ab9abed.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/d6178593b5834f9fa487bca389ce5d83/7e4e8f9d2b6dcf0d19c1ab693869ad848ec0dd1b339071c357fa23e80ab9abed.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style="width: 100%;"/>ComfyUI recently introduced a sampling method called Euler A++, based on CFG++ technology. Finally, the SDXL model's long-standing deficiencies in saturation, contrast, have been addressed while optimizing scenarios with low CFG values that previously resulted in blurry and overexposed images.It perfectly supports models with low CFG parameters like HYPER, Lightning, LCM, Turbo, compensating for losses in image quality. Compatible with CADS \ UNIT \ ACG \ MIST XL series acceleration models (Hyper), not only does it provide better inference sampling results within 8-10 steps, but also further enhances image quality.This workflow for CFG++ sampler image generation configuration aims to inspire more applications supporting new technologies and sustain the prosperity of the SDXL ecosystem! <img src="https://liblibai-online.liblib.cloud/img/d6178593b5834f9fa487bca389ce5d83/4e4aaef76f6bd6d0ba5716da442dd0a7b12395c9d79d3957eaebe63959586b74.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" alt="" data-href="https://liblibai-online.liblib.cloud/img/d6178593b5834f9fa487bca389ce5d83/4e4aaef76f6bd6d0ba5716da442dd0a7b12395c9d79d3957eaebe63959586b74.png?x-oss-process=image/resize,w_1290,m_lfit/format,webp" style=""/> The CFG++ workflow configuration is equally user-friendly, avoiding cryptic methods (AiARTiST has always believed that practical value lies in technology that can be diffused).Make sure to update ComfyUI to versions after June 26, 2024 ————————————————————————————————————————

CADS3XL-CFG++-workflow.json

Files with SHA256: 90f9ed20f49959ad09b691a62ba5807bca5a8d9749e53448e51703a180c5ef2a