This is a ControlNet focused workflow for SDXL, designed to be visually compact and simplified for ease of use.
This workflow utilizes CLIP L and CLIP G, which most people don't even realize is the actual way you're supposed to use SDXL. This is a dual positive prompt system. Technically you'd want to prompt each clip differently, but most people just use the same prompt for both CLIP models. CLIP L is generally better at handling keywords, while CLIP G is better at natural language, and prompting differently between them can potentially get better results.
This workflow utilizes Kohya Deep Shrink High Res Fix, an infinite LoRA loader, and is geared toward DMD2 acceleration and LCM sampling settings out of the box. Kohya High Res Fix should ideally be used at a block_number of 16-32, everything else can be left alone. It can be used at 1024x1024, but generally you want to use it when you generate at resolutions higher than that. I find using it at 1500x1500 to be a good bump up in image quality while maintaining coherence. I'd experiment with enabling/disabling Kohya or the CLIP models node if you're not getting the results you want, which you can do by selecting the node and pressing Ctrl+B.
Negative prompt is minimized by default. I think this is a good space saving solution, as most people either use no negative prompt or the same negative prompt across generations, and CFG 1 ignores negative prompt anyways. Clicking the box beside "Negative Prompt" will maximize the node and make it available for use if needed.
_________
Required models:
DMD2 LoRAs:
FP32: https://huggingface.co/tianweiy/DMD2/resolve/main/dmd2_sdxl_4step_lora.safetensors?download=true
SDXL ControlNet models:
https://civitai.com/models/136070/controlnetxl-cnxl
Recommended ControlNet models:
ControlNet Union: https://civitai.com/models/136070?modelVersionId=655749
ControlNet Union ProMax: https://civitai.com/models/136070?modelVersionId=655836