SDXL
2026/03/17 updateDue to a comprehensive reconstruction of the training & merging strategy, v3.0 marks a significant leap in quality and fidelity compared to previous versions. The major improvements are as follows:Perfect Trinity Balance: v3.0 successfully finds the "Golden Intersection" between 2.5D, Realistic, and Anime styles. The output is exceptionally stable and detailed, mastering both Semi-realistic and Anime aesthetics. Whether for SFW, NSFW, or complex ControlNet workflows [see last three preview images], it maintains incredible image coherence and structural integrity.Structural Aesthetic Reinforcement: We have implemented targeted structural enhancements for makeup details and body proportions specifically for the 2.5D style. While preserving the signature "sharpness" of 2.5D art, characters now possess a more alluring mature charm and more aesthetically refined body curves.New Exclusive Trigger Tag: jwsfsgrl Unlock the model’s full potential with the new trigger tag: jwsfsgrl. Usage Tip: Pair this with tags like realistic or photorealistic to achieve a unique effect: Hyper-realistic lighting and textures overlaid on stylized anime proportions.IntroductionYesmix XL (illustrious) is based on illustrious XL and trained with around 16,000 anime images rated 'great' or higher by aesthetic evaluation model. It is now running smoothly and stably on ComfyUI. This model is dedicated to stably generating extremely high-quality 2.5D/anime/realistic SFW/NSFW images.UsageFor ComfyUI, simply download the uploaded node group (zipped .json file) and place it in the specified path to generate images of the same quality as the previews. All nodes in the node group are ComfyUI built-in nodes to ensure universal workflow compatibility.workflow path: /path of ComfyUI/ComfyUI/user/default/workflowsAvoid setting the weight of single prompt/tag above 1.5.It is recommended to use Euler or Euler A as the sampler, as images generated by these samplers are generally cleaner.Names of some well-known artists, such as sakimichan, liang xing, lexaiduer, and wlop, were used as tags in model training. Using these names for image generation may enhance the style of images.AcknowledgementOnomaAIResearch, for open-sourcing the Illustrious XL model.Kohya's Stable Diffusion training project and all contributors, for their technical support in training the SDXL model.DeepGHS Team, for building source mirrors of Danbooru, Gelbooru, and Sankaku datasets and for their contributions to intelligent image filtering and evaluation technologies.
