unCLIPConditioning

Before IP-Adapter and reference-latent tricks existed, this was the way to get a diffusion model to care about a picture: you ran the image through a CLIP vision encoder and injected its embedding straight into the model's ADM conditioning. That's what unCLIPConditioning does, and it's one of the oldest nodes in this whole group - it's been in ComfyUI core (nodes.py) since the SD 2.x days. It's not the hot new thing, but it still works, it's dead simple, and it's the right tool for the narrow slice of models that still speak the unCLIP dialect.

The mechanism: you feed it existing conditioning (from a normal text encode) plus a clip_vision_output (from a CLIPVisionEncode of your guide image). The node appends an unclip_conditioning entry to the conditioning - carrying the vision output, a strength, and a noise_augmentation value - and passes it through. If strength is 0, it short-circuits and returns the conditioning untouched, which makes it easy to toggle in and out of a graph. noise_augmentation adds noise to the image embedding before injection; it's a weird but real knob that was meant to make the model treat the image as a rough suggestion rather than gospel - bump it for looser interpretation.

Inputs: conditioning, clip_vision_output, strength (−10 to 10, default 1), noise_augmentation (0–1, default 0). Output: one CONDITIONING.

Where this actually shows up in 2026: SD 2.x-unCLIP checkpoints (there's a whole unCLIP variant family that conditions on image embeddings natively), and occasionally in old "reimagine" workflows built around the SDXL unCLIP models. If you're running one of those, you need the matching CLIP vision model from the loader - this node doesn't load anything itself. It ships with core, so no install.

The honest take: unless you're on an unCLIP checkpoint, you probably shouldn't reach for this. The modern equivalent - IP-Adapter or reference-latent conditioning - is easier to control and works with every model. But knowing what it is saves you from a confusing afternoon if a template from 2023 shows up with this node in it. And the one genuine failure mode to know: the clip_vision_output has to come from the right vision encoder. The SD2.x-unCLIP models pair with their specific CLIP vision, and using a generic one gives you conditioning that influences nothing - you'll get a perfectly normal image that ignores your guide entirely.