Load unCLIP Checkpoint

Every once in a while you load a workflow and run into a loader that outputs four things where you expected three. Load unCLIP Checkpoint is that node. It's a checkpoint loader for the Stable Diffusion 2.1 unCLIP family - the models Stability AI built on OpenAI's unCLIP/DALL-E 2 idea - and its extra output is a CLIP vision model. If you've never heard of unCLIP, that's not a knowledge gap; that's the point. This is a loader for a model family the community largely forgot.

What unCLIP was

The unCLIP approach is image-conditioned generation: instead of (or alongside) a text prompt, you feed the model a reference image, and it generates something related to it. Stability's SD 2.1 unCLIP checkpoints (sd21-unclip-h and sd21-unclip-l, the suffixes being which CLIP vision model they pair with) let you do variations and image-to-image-ish work where the visual content of a source image steers generation. It was a 2022/2023-era idea, interesting, and then it just... faded - eclipsed by img2img, IP-Adapter, and everything that came after. The node stayed in core because the format is still loadable and somebody's old workflow still references it.

How the loader works

One input, ckpt_name, listing checkpoints from models/checkpoints/ like any other checkpoint loader. The difference is what it returns - four outputs:

• MODEL - the diffusion model
• CLIP - the text encoder
• VAE - as usual
• CLIP_VISION - the extra one: the checkpoint carries its own vision encoder, and the loader surfaces it so you don't need a separate Load CLIP Vision node

That CLIP vision output is the whole point. It goes into CLIP Vision Encode to produce a CLIP_VISION_OUTPUT, which then feeds unCLIPConditioning - the node that merges the image-derived signal into your text conditioning with a strength slider (default 1.0) and a noise_augmentation knob that controls how much the reference image's details survive.

The traps

First, this loader expects an unCLIP checkpoint. A normal SD1.5 or SDXL checkpoint loaded through it won't give you a meaningful CLIP vision output - that extra output only exists when the checkpoint actually bundles a vision encoder. If your CLIP_VISION wire comes up empty, the answer isn't a bug in your graph, it's the model.

Second, don't expect modern quality. The SD 2.1 unCLIP models predate basically everything you use now, and "image-conditioned variation" has far better modern homes. This node is for historical workflows and format exploration, not new creative pipelines. But it's a genuinely interesting piece of lineage - the ancestor of the idea that "feed it a picture, not just words" - and if you stumble on it in an old workflow, now you know what that fourth wire is for.