Style Conditioner (Mikey)

The README calls this one "chonky," and honestly, fair - it's got eight inputs and five outputs, which is a lot of sockets for a node whose whole job is: blend a style into your prompt. But the mechanism underneath is simple once you see it, and it's built specifically for SDXL's base+refiner two-stage pipeline that this pack's sampler nodes run on.

What it actually does

You've already encoded your real prompt into conditioning - positive and negative, for both the base model and the refiner. This node takes that conditioning and averages it against conditioning built from a chosen style, rather than concatenating the style text onto your prompt and re-encoding from scratch. strength controls the blend ratio: 0.5 means your prompt's conditioning and the style's conditioning get equal weight; push it toward 1 and the style dominates, toward 0 and your original prompt wins out. That's a meaningfully different effect than just appending <style:cinematic> text to your prompt string - you're mixing two already-encoded conditioning tensors, not asking the CLIP encoder to reconcile two ideas at once.

The style list itself is the same one baked into this pack's "Prompt with Style" node - dozens of options built off Stability's SAI style set (SAI-Photographic, SAI-Anime, SAI-Cinematic, and so on) plus a batch of community-flavored extras like wes-anderson and vaporwave. If that's not enough, the README documents a user_styles.json file you can drop in ComfyUI's root folder to add your own.

Inputs and outputs that matter

• style - pick from roughly 85 built-in options (plus none).
• strength - 0 to 1, default 0.5, in steps of 0.1.
• positive_cond_base / negative_cond_base / positive_cond_refiner / negative_cond_refiner - your already-encoded conditioning, one pair per stage.
• base_clip / refiner_clip - the CLIP models used to encode the style text itself.
• use_seed + seed - flip use_seed to true and the seed picks a style for you instead of you choosing manually, which is a nice way to get variety across a batch without hand-curating a style list.

Out come base_pos_cond, base_neg_cond, refiner_pos_cond, refiner_neg_cond - feed these straight into your base and refiner KSamplers in place of your original conditioning - plus style_str, a plain string telling you which style actually got used. Wire that into your filename or metadata node so a randomized style run doesn't leave you guessing what produced which image.

Installing it

Through ComfyUI Manager (search "Mikey Nodes"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/bash-j/mikey_nodes

Restart ComfyUI. No model downloads needed for this node specifically - the style prompts are text baked into the pack, not external files, unless you're adding your own via user_styles.json.

Where it trips people up

This node is built for a base+refiner SDXL setup - if you're only running a base model with no refiner stage, you'll have two of the four conditioning inputs with nothing sensible to plug in. Use Style Conditioner Base Only instead; it's the same idea stripped down to a single stage.

Since this blends conditioning rather than text, a strength near 1.0 can produce oddly generic output if your style choice is thematically miles from your prompt - the average of two very different concepts isn't usually either concept, it's a blur between them. Start around 0.3-0.5 and push higher only once you've confirmed the direction is right.