A small image-processing block before captioning and sampling. It replaces manual Photoshop touch-ups.
SDXL is 'quite strict about the resolution you put inside'. Set the target resolution in the resize node rather than feeding the raw export.
A slight adjustment, tuned to the input. A SketchUp screenshot may need a different tweak from a Blender export (Flat, shadowless, high-contrast base render).
Upscale the input image, then resize it back to the working resolution before the sampling stage. He found this keeps as much detail as possible.
The prepared image feeds Florence-2 detailed captioning, the ControlNet preprocessors (depth-and-canny-controlnet-maps-from-input) and the image-to-image latent (juggernaut-xl-img2img-checkpoint-graph). See Blender to ComfyUI SDXL image-to-image render pipeline.