Face Fix - Store Enhanced Anchors

The Face Fix pipeline has a hand-off problem. The anchor frames come out of one model, get enhanced by a different model, and then have to reach a third model in the exact right order. Somewhere in there the anchors have to be stored - written to disk in a deterministic sequence - so nothing depends on a fragile in-memory connection across what are effectively separate pipeline stages. VRGDGFaceFixStoreAnchors is that save point. Per the author's description, it "saves the enhanced anchor batch in deterministic order for LTX conditioning."

How it works. You feed it the decoded IMAGE output from Z-Image - the batch of enhanced anchor faces - plus the face_fix_context from the Load Anchors (Meta Batch) node. The context is the key: it links these enhanced results back to the specific Face Fix job and its anchor ordering, so the saved files land in a known folder with a known sequence. The node writes them out and reports where they went, how many there are, and their indices. It's marked as an output node, so this is a genuine endpoint - the moment the enhanced anchors are "on disk and accounted for."

The tooltip carries the one constraint that matters: the number and order of images in enhanced_anchors must exactly match the anchors loaded by the Meta Batch node. Z-Image might return a different count or a shuffled order if you look at it sideways, and if it does, LTX will get identity anchors pointing at the wrong frames. Determinism is the whole point of the deterministic-order language.

The inputs. enhanced_anchors (the Z-Image decode output) and face_fix_context (from Load Anchors Meta Batch). Both required, both simple.

The outputs. enhanced_anchor_folder (where the files went - feed it to LTX's anchor conditioning), anchor_indices (which frames these correspond to), anchor_count (how many), and face_fix_context passed through so the chain continues.

Where it sits. This is the "enhanced anchors proven saved" milestone. In the workflow, it plugs into the execution barrier (Collect LTX Inputs) that refuses to let LTX run until both the cropped face video and the enhanced anchors exist. StoreAnchors is what makes the "enhanced anchors exist" claim true.

Installing. In the vrgamedev pack:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

or install via ComfyUI Manager. You need Z-Image Turbo upstream and the LTX 2.3 model set downstream, same as the rest of the Face Fix family.

Troubleshooting. The failure mode to watch is a count mismatch: if anchor_count here disagrees with what the Meta Batch loader produced, don't run LTX - fix the anchor selection upstream (usually an anchor_interval change in Prepare) and re-run. Also, this node writes to disk, so if the enhanced_anchor_folder output looks wrong, check where the pack is configured to write outputs; a stale folder from a previous job can quietly hold old anchors. When in doubt, clear the anchor folder and let it regenerate.