🎛️ Style Mixer

Style Mixer is the pack's answer to a boring-sounding problem that eats real time: you have a reference photo with the right mood, a style preset with the right look, and a base prompt - and you need one prompt that carries all three without reading like a cut-and-paste crime scene. It takes up to four reference images and three text styles, lets you weigh each one, and hands back a styled_prompt you can feed straight into a CLIP Text Encode.

The clever bit is that it has two modes, and the default doesn't touch the network at all.

Two fusion modes

Weighted Stack (Fast) is deterministic string composition - no LLM call, no config needed, runs instantly. It merges the base prompt with whatever each image contributes (analyzed through the focus you set) and the style presets, scaled by your weights. This is the mode to live in for iteration: it's free, reproducible, and instant, so you can drag weights around a dozen times without burning tokens or waiting on a queue.

Smart LLM Fusion (Gen-Mix) sends that whole stack to the vision model and asks for a coherent rewrite - one fluent scene instead of stacked clauses. It needs config from the Provider Loader, and it's the mode to switch to the moment the stacked output reads like disconnected phrases rather than one image. The tell is simple: if the weighted result makes sense as a sentence, keep the fast path; if it reads like a shopping list, fuse it.

The inputs that matter

• base_prompt - the prompt the styles are applied to. The one field you'll always touch.
• image_1..image_4 - reference images, each with an img_weight (defaults taper 0.8 → 0.2) and an img_focus (Style & Texture, Color & Lighting, Subject & Composition, Mood & Atmosphere). The focus tells the model what to pull out of each reference - which is how you keep one image contributing color while another contributes mood.
• style_1..style_3 - presets from the pack's 409-entry catalog (photo, art, and both NSFW libraries), each with its own weight.
• fusion_mode - the big switch above.

Two outputs: styled_prompt (what you wire into the encoder) and style_overlay (just the style layer, if you want to stack it under another prompt elsewhere).

Where people trip

The most common mistake is loading up all four image slots and three styles on the first run. Everything at full weight is how you get mush. The defaults are tapered for a reason - the pack expects one dominant reference (0.8), with the others as seasoning. Start with one image and one style, check the output, then add. Also note the images aren't blended pixel-wise; this is a text node. It describes what the references look like and folds that description into the prompt. If you want literal pixel blending, this isn't the node.

LLM fusion failing is a provider issue, not a mixer issue - it falls back to the weighted stack, so you get the deterministic result rather than an error, which is a nice touch. If your LLM fusion returns something incoherent, the chained-reasoning lesson applies: the LLM is rewriting, not inventing, so feed it a well-formed weighted stack first.

Install and providers are the pack's standard: Manager search FiL_Design_ImageMind (or git clone + pip install + restart), and Ollama works keyless for the fusion mode. Fast mode needs none of it.