ChromaPulidFluxInsightFaceLoader

Every PuLID workflow has a face-finding step, and in this pack it lives here. ChromaPulidFluxInsightFaceLoader loads InsightFace's antelopev2 model pack - the detector and ArcFace recognizer that look at your reference photo, find the face, and turn it into the identity vector PuLID conditions on. No face analysis, no likeness; the apply node (ChromapplyPulidFlux) would have nothing to work with.

It's a simple node. One input, one output:

• provider - CPU, CUDA, or ROCM. Pick CUDA on an NVIDIA card, ROCM if you're on AMD - and this is the rare pack where that's actually tested, since the author develops on an RX 6900 XT. CPU works but is slow, and it's handy when you're debugging a provider-level crash.
• FACEANALYSIS output - feeds the apply node alongside the PuLID model and EVA-CLIP outputs.

Under the hood it builds an InsightFace FaceAnalysis session rooted at ComfyUI/models/insightface/ and prepares it with a 640x640 detection size. The apply node then sweeps your image at multiple scales, grabs the largest face (a deliberate fix in this lineage - the original code picked the smallest), and extracts its ArcFace embedding.

The download that breaks everyone

The antelopev2 pack (~407 MB) is not auto-downloaded by the InsightFace library - that's a manual step on third-party file hosting, and it's the single most common reason a PuLID setup "doesn't work." Get it, unzip it, and place it so the folder structure is:

ComfyUI/models/insightface/models/antelopev2/

Miss this and the loader throws a model-not-found error the first time you run it. ComfyUI Manager's model installer can fetch it for you if you search "antelopev2"; otherwise grab it from the InsightFace model zoo.

The rest of the install is the pack itself:

cd ComfyUI/custom_nodes
git clone https://github.com/nazgut/ComfyUI-PuLID-Flux-Chroma
pip install -r requirements.txt

The insightface dependency is listed in requirements.txt. It used to be one of the worst installs in local generation - prebuilt wheels vanished for years and you had to compile against a C++ toolchain - but InsightFace 1.0 (May 2026) dropped the build requirement, so a fresh pip install should just work on a current setup.

The part nobody reads until it matters

InsightFace's code is MIT, but the antelopev2 weights are non-commercial. That means everything this node produces for you - and by extension the whole PuLID pipeline in this pack - is fine for personal projects and portfolios, and a licensing problem if you plan to sell the output. It's the same trap sitting under every face tool, from InstantID to ReActor, because they all share this one backbone. The pack's own license is MIT, which makes it easy to miss that the models underneath are the real terms. Read that before you build a business on it.

Practically: pick the right provider, get antelopev2 in the right folder, and this node quietly becomes the least interesting part of your workflow. That's the goal.