SwarmUI: generate a fictional character dataset with SDXL

This Spanish tutorial demonstrates fictional character image generation in SwarmUI with SDXL-derived checkpoints and shared ComfyUI model folders. It focuses on generating and selecting images for a character dataset, with a brief explanation of training. It does not demonstrate a complete LoRA training workflow. Video generation, speech and consistent animation are left for later tutorials.

For this local AI workflow, the speaker recommends Windows, an NVIDIA RTX 3060 or higher, and at least 32 GB of RAM. He says less RAM can work but will be slower, and suggests renting a cloud computer if suitable hardware is unavailable. ComfyUI, SwarmUI and Antigravity must already be installed; installation is covered elsewhere.

The walkthrough uses detailed character descriptions, an SDXL checkpoint and SwarmUI's Grid Generator to compare candidates across four seeds. It shows how to point SwarmUI at ComfyUI's model folder. The speaker initially selects Analog Madness, then switches checkpoints after an unwanted explicit result. He lowers previews to 512 × 512 for selection and recommends higher resolution for final images.

After choosing a face, the speaker generates variations in clothing, viewpoint and lighting. He recommends selecting at least 20 or 30 consistent images and warns that many outputs may be unusable. The final demonstration uses ControlNET depth preprocessing to follow a reference pose, with prompt changes and reduced control weight to address errors.