Tracked Face Detail (SDXL)

Tracked Face Detail is the other half of the split path: it takes a FACE_TRACKS object plus the original frames and runs the actual SDXL sampling - every anti-flicker trick in the Temporal Face Detailer pack, applied to the tracked face crops. It's the exact same detailing engine as the all-in-one node, just without detection inside. You wire Face Detect + Track's face_tracks into this, and suddenly you can iterate on sampler settings without re-running tracking, because tracking only happened once.

How it fits in

The split workflow is: Load Video → Face Detect + Track → Tracked Face Detail → Video Combine. And yes, you feed the same original frames into Tracked Face Detail (image input) - the detailer needs the source pixels to crop and paste back against. Model/CLIP/VAE come from a normal SDXL checkpoint. The README calls this the "tunable" path, and it's right: tracking is the expensive, error-prone half to redo, so isolating it means you can burn as many sampler experiments as you like.

What it does to each track

For each track, the node builds a stabilized square crop (constant size, landmark-anchored center, temporally smoothed - the geometry was already computed during tracking), resamples it at guide_size (768 default, max_size 1024 caps it), VAE-encodes, and samples. The important part: with noise_mode: fixed_per_track (default), every frame in the track is sampled from the same seed and noise, so the sampler literally cannot diverge frame-to-frame. Then it optionally blends in latent space (latent_blend), decodes, color-matches each crop to its source, applies the flow-guided pixel blend (flow_strength), and feather-pastes only the face region back (mask_dilation / feather).

detail_mode is worth understanding: img2img (default) resamples the whole crop, with the mask only limiting paste-back; inpaint restricts sampling to the face mask in latent space. If you're seeing the background change around faces, that's img2img resampling the whole crop - inpaint is the tighter option.

Inputs that matter

Beyond image, face_tracks, and the checkpoint, the ones you'll actually live on:

• denoise (0.35 default) - main quality-vs-consistency lever; 0.3–0.45 is the sweet spot. Add denoise_max for adaptive behavior: steady frames stay at the base, high-motion frames ramp toward the max.
• noise_mode - keep fixed_per_track unless you're A/B testing against naive per-frame detailing.
• flow_strength / latent_blend - the anti-flicker dials. Raise latent_blend (try 0.3–0.5) instead of cranking denoise when you want more detail.
• reference_image + reference_strength - an optional identity anchor: a clean face photo the crops are nudged toward. Invaluable when a character slowly drifts off-identity over a long clip.
• track_prompts - per-track positive prompt overrides, one per line like 0: photo of johndoe person, detailed face. Get the IDs from the debug overlay.
• lora_stack - feed a TFD LoRA Stack here to apply face LoRAs only on the crops.
• chunk_size - lower it (default 4) if you OOM; detail_every 2–4 is the keyframe speed mode for long clips.

Outputs

image goes to Video Combine (the finished frames), and face_masks (MASK) is optional - preview it via a Mask To Image node or save it to see exactly what got re-rendered. There's no debug_overlay or face_tracks out on this node; those come from the tracking side.

Where people get burned

The usual first mistake is forgetting the image wire - the node silently needs the original frames alongside the tracks, and a graph that "looks right" just produces an error until you connect it. OOM on long clips is the second: drop chunk_size. And if you're chasing detail and the face starts flickering again, the fix is almost never more denoise - it's latent_blend, or going back upstream to check that the tracking itself is stable.