TFD LoRA Stack

TFD LoRA Stack is a chained list-builder for the Temporal Face Detailer pack. It packs one or more LoRAs into a LORA_STACK that the detailer applies internally, only while sampling the face crops - as opposed to applying the LoRA to the whole model upstream, where it influences every pixel of every frame.

That distinction is the whole reason the node exists. Say you have a face LoRA that nails a character's features but would wreck your background style if it ran across the full frame. Patch it in before the detailer and it touches everything. Feed it here via lora_stack and the crop samples get the identity boost while the rest of the video never sees it.

How it works

You pick a LoRA from the dropdown (everything in ComfyUI/models/loras), set model_weight and clip_weight (both default 1.0), and the node appends it to a stack. The optional lora_stack input lets you chain - wire one node's LORA_STACK output into the next's lora_stack input to combine several LoRAs:

TFD LoRA Stack (face_lora_A) ──▶ TFD LoRA Stack (face_lora_B) ──▶ detailer's lora_stack

The resulting stack plugs into the lora_stack input on Temporal Face Detailer or Tracked Face Detail, where it's applied before each crop is sampled.

Inputs and output

• lora_name - dropdown of your installed LoRAs; None means "don't add one" (handy for a passthrough link in a chain).
• model_weight / clip_weight - per-LoRA strengths, -10 to 10.
• lora_stack (optional) - the previous node's stack, to chain.
• Output: LORA_STACK, wired into the detailer.

That's the whole node - it builds data, it doesn't sample anything.

When to use it versus a plain LoraLoader

The README supports both: apply LoRAs upstream with a normal LoraLoader (Load Checkpoint → Load LoRA → detailer), or use the stack for face LoRAs you only want on the crops. The stack is also the tidier option when you're chaining several - one node per LoRA, no spaghetti through a model. And since the detailer accepts a patched MODEL too (that's how you'd combine this with an IPAdapter FaceID patch for strong identity conditioning), you can have both: an upstream identity patch for the whole frame and a face-specific LoRA stack for the close-ups.

One practical note: if you only need a single LoRA and it's safe on the whole frame, a plain LoraLoader upstream is fewer moving parts. Reach for this node the moment you want the LoRA scoped to the face - that's precisely the case it's built for.