Prompt With SDXL (Mikey)

SDXL's biggest ergonomic quirk for newcomers is that it doesn't have one text encoder - it has two, OpenCLIP ViT-bigG and CLIP ViT-L, wired together as clip_g and clip_l. The second encoder is where a lot of SDXL's extra prompt control over composition and lighting comes from, but the standard way of using it (a separate CLIPTextEncodeSDXL node with its own width/height/target/crop fields) is genuinely confusing the first few times you meet it. Prompt With SDXL collapses that whole setup, plus an empty latent, into a single node.

What it actually consolidates

This one node replaces what would otherwise be a CLIPTextEncodeSDXL node plus an Empty Latent Image node plus some manual bookkeeping about which text goes to which encoder. You get four separate text boxes - positive_prompt, negative_prompt, positive_style, negative_style - a ratio picker (the same trained-SDXL-resolution list as Mikey's Empty Latent Ratio Select SDXL node), a batch size, a seed, and one dropdown that decides the whole routing question for you: output_option.

That dropdown is the node's real value. Its five choices spell out, in plain language, exactly how your prompt and style text get distributed across the two encoders - for example "prompt -> clip_g and style -> clip_l" sends your main prompt to the bigG encoder and your style text to the L encoder, while "prompt+style -> clip_g and prompt+style -> clip_l" sends the combined text to both. This is the exact routing decision the community otherwise has to learn by trial and error with the stock node - here it's presented as a menu instead of a mystery.

The outputs

Nine of them, and each one earns its place: samples (the empty LATENT, sized to your chosen ratio and batch), positive_prompt_text_g / negative_prompt_text_g / positive_style_text_l / negative_style_text_l (the actual text strings that got routed to each encoder - useful for logging into a Save Image With Prompt Data node's metadata fields), and width / height / refiner_width / refiner_height. That last pair exists because SDXL originally shipped with an optional refiner model for a second denoising pass; the community mostly stopped bothering with it once fine-tuned checkpoints got good enough on their own, but this node still gives you the dimensions you'd need if you're one of the people still running a refiner stage.

Installing

Part of the full Mikey Nodes pack - ComfyUI Manager, search "Mikey Nodes," install, restart, or cd ComfyUI/custom_nodes && git clone https://github.com/bash-j/mikey_nodes and restart manually. Nothing extra to download for this node specifically; the ratio list and encoder logic are baked into the pack's own Python code.

Where people trip up

• Picking the wrong output_option and getting muddy results. If your positive prompt and style text end up fighting each other visually, try a different routing option - sending them to separate encoders (rather than combining both into both) tends to give each one more distinct influence.
• Forgetting this replaces two nodes, not one. If you already have a separate Empty Latent Image node in your graph and add this on top, you'll end up with two latents and have to explicitly pick which one feeds your sampler - this node is meant to be the only latent source in that part of the graph.
• Blank style fields doing nothing. Leaving positive_style/negative_style at their literal placeholder text ("Positive Style" / "Negative Style") means that placeholder text gets encoded as-is - clear the fields or replace them with real content before you rely on style routing actually doing something.