Load Eva Clip (PuLID Flux)

One of three loaders you wire into a PuLID-Flux workflow, and by far the most boring - in a good way. It loads the EVA-CLIP vision encoder that PuLID uses to read your reference image, and it does it with zero inputs. You drop the node down, connect its one output, done. No dropdown to fumble, no file to point at.

Here's where it fits. Getting a face into Flux with PuLID takes two different readings of your reference photo. InsightFace (the face-analysis loader) produces the ArcFace identity embedding - the geometric fingerprint of the face. EVA-CLIP encodes the broader visual content of the image. PuLID's adapter uses both together, which is part of how it holds a likeness while still respecting your prompt. This node supplies the EVA-CLIP half. Specifically it loads EVA02-CLIP-L-14-336, the encoder the PuLID-Flux model was trained against.

Inputs and output

There's nothing to configure. The node has no inputs at all - it just loads the encoder when the graph runs. Its single output is EVA_CLIP, which plugs into the eva_clip slot on ApplyPulidFlux. That's the entire job.

That simplicity is probably why this page gets found more often than the rest of the pack - people wiring up a PuLID graph hit "what's this EVA_CLIP thing and where do I get it," and the answer is: this node, and it handles itself.

Installing it - and the auto-download

Comes with the ComfyUI_PuLID_Flux_ll pack. Manager search ComfyUI_PuLID_Flux_ll, or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/lldacing/ComfyUI_PuLID_Flux_ll.git
cd ComfyUI_PuLID_Flux_ll
pip install -r requirements.txt
pip install facenet-pytorch --no-deps
# restart ComfyUI

The nice part: the EVA-CLIP weights auto-download on first use. The pack fetches EVA02_CLIP_L_336_psz14_s6B.pt into ComfyUI/models/clip/ for you. (Heads up if you're following an older tutorial - a January 2025 update moved this file's expected location to models/clip/, so guides that point somewhere else are stale.) You can pre-download it manually from the QuanSun/EVA-CLIP HuggingFace repo and drop it there yourself if you'd rather not wait on the download.

Standard pack requirements apply: ComfyUI ≥ 0.3.7, and disable any old balazik PuLID-Flux nodes first since this pack reuses the ApplyPulidFlux name.

Common issues & troubleshooting

First run hangs or errors on a download. The auto-download is pulling ~1.7GB from HuggingFace. If your box can't reach HF, or the download got interrupted, it'll fail. Fix: download EVA02_CLIP_L_336_psz14_s6B.pt by hand from QuanSun/EVA-CLIP and place it in ComfyUI/models/clip/, then restart.

It downloaded but ApplyPulidFlux still complains. Make sure the file actually landed in models/clip/ and not some cache dir, especially if you're on a version straddling that January 2025 path change. Move it into models/clip/ and restart.

The node has no widgets and you think it's broken. It's not - no inputs is by design. If the output socket is there and typed EVA_CLIP, it's working. Just connect it to ApplyPulidFlux.

That's genuinely all there is to it. This is the one node in the PuLID chain you'll never have to think about twice.