Load Style Model

Load Style Model loads a specific, somewhat forgotten category of model from ComfyUI/models/style_models and hands you a STYLE_MODEL wire. It's the loader behind SDXL's reference-style feature - the "give me an image in the style of this picture" trick that predates IP-Adapter and does a narrower job than it. If you've never seen this node in a real workflow, that's partly because its niche shrank, but it's also genuinely useful when you know what it's for.

What a style model actually is

The file it loads is not a diffusion model, not a text encoder, and not a full ControlNet. It's a small adapter trained to pull style - color, texture, compositional feel - out of a reference image and inject it into generation. The classic SDXL implementation works through the reference-style path: encode the reference image, project its style features, and fuse them into the conditioning so the sampler reproduces the look without copying the content. Structure (where things go) was always ControlNet's job; appearance (how it looks) is this adapter's. That pairing - style model for appearance + ControlNet for structure - was a real 2024 power move, and the style-model half is the one that survived less visibly.

How it works

One input, style_model_name, listing files in models/style_models. The output STYLE_MODEL goes into the Style Model Apply node, which takes the style model, your CLIP conditioning, and a reference image, and returns new conditioning for the sampler. Same shape as ControlNet Apply, one input short. You're wiring: Load Style Model → Style Model Apply → (conditioning into KSampler), with the reference image providing the look.

Honest 2026 context

The reason this node faded from view is the same architecture shift that ate IP-Adapter. The reference-image job mostly moved off the classic adapter path: on modern DiT bases, appearance transfer became the domain of edit models and in-context reference conditioning, and neither runs through this node. The community's verdict over the years has been consistent - style models were always the lighter half of the reference story, better at color and mood than at identity, and usually paired with something else. The "doesn't actually copy the subject" complaint was the recurring theme. If you want a specific person or object to look right, that's a LoRA or an identity-preservation workflow, not this.

Where it genuinely still earns its place: SDXL workflows that want a consistent painterly or photographic mood from a reference, without the weight or fragility of a ControlNet. It's fast, it's small, and it ships with core. And on the modern side, style-conditioning has a quieter second life: the SigLIP-based style/reference projectors that load through Load Model Patch are doing a version of this job for Flux, which is the lineage continuing under a new name.

Gotchas

• It's a style adapter, not an identity copier. Manage expectations: reference style, not reference subject.
• Models are rare. The style_models folder is one of the emptier default folders, because few people train or share these adapters anymore. If the dropdown is empty, that's the ecosystem talking, not a broken install.
• Needs a reference image and existing conditioning. The node has nothing to do alone - it only shapes conditioning that's already there.

It ships with ComfyUI core - no install. It's a niche node that answers a specific question, and if you're on SDXL chasing a look rather than a face, it's still a clean answer.