FaceCam (FFMPEGA)

Camera movement is the thing Wan 2.2 will not give you from a text prompt. Ask for "the camera orbits left around her face" and you get a cut, or a face that quietly turns into someone else's. FaceCam fixes that by not asking the model nicely - it draws the camera path onto the frame and hands it over as conditioning.

What it actually is

FaceCam is a partial fine-tune of Wan 2.2 I2V-A14B, trained on portrait video for camera control (CVPR 2026, by the FaceCam authors; this node wraps it). "Partial" is the important word: the shipped checkpoints only carry self-attention and patch-embedding weights, about 402 keys. The node swaps those into a base Wan 2.2 model at runtime. So you don't load FaceCam instead of Wan - you load your normal Wan 2.2 A14B high/low pair and the FaceCam weights ride along on top.

It keeps Wan 2.2's two-pass shape - high-noise model for motion and composition, low-noise model for detail. In this node the split is explicit: high_model_ratio (default 0.2) is the fraction of steps run on the HIGH model, which carries the camera trajectory, and the rest go to the LOW model for refinement. model_low is optional - leave it unconnected and the HIGH model runs every step.

How the camera signal is built

Camera pose is written as a MediaPipe face mesh drawn on your input frame, VAE-encoded, and fed in as the camera condition. The 2.20.0 release changed how that mesh is produced and it's worth knowing which path you're on. mesh_source defaults to auto, which poses MediaPipe's canonical 468-vertex model with the camera matrices directly and projects it. The old detected path rendered a 3D proxy head along the camera path and ran a face detector over the renders - and MediaPipe loses the mesh past roughly 40–50° of yaw, which is precisely the pose an orbit produces. That's why conditioning used to go blank at the ends of a sweep. analytic forces the new path and errors if it can't calibrate. Leave it on auto.

Inputs you'll actually touch

• model_high (required) and model_low - your Wan 2.2 A14B pair, loaded however you normally load it; the docs assume GGUF via Load Diffusion Model. Plus vae (Wan 2.2 VAE) and clip (T5-XXL).
• images or video_path - the source frame. Just a frame or two is normal; FaceCam animates from it.
• camera_preset - eleven options. orbit_left / orbit_right swing the camera to ±45°, zoom_in pushes FOV 50→25°, zoom_out pulls back, look_up / look_down tilt the subject, dramatic_pan, subtle_drift (the most natural), dolly_zoom, random, and custom.
• width / height - defaults are 480×704, and that's not a lazy default. FaceCam is trained on portrait video; landscape is off-distribution and camera control measurably weakens.
• num_frames - 81 out of the box, stepping in 4s.
• blockswap_blocks - 0 disabled and recommended at 24 GB+; 8–16 for 12–16 GB cards, 20+ for 8–12 GB.

Set camera_preset to custom and the manual sliders take over: start_azimuth/end_azimuth (±90°, negative = camera left), start_elevation/end_elevation (±60°, positive = looking down), and start_fov/end_fov (10–60, lower is tighter). Outputs are images, video_path, and a latent you can hand to another sampler or a second FaceCam pass.

Install

ComfyUI Manager → search ComfyUI-FFMPEGA → Install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/AEmotionStudio/ComfyUI-FFMPEGA.git
pip install -r ComfyUI-FFMPEGA/requirements.txt

Restart ComfyUI. The pack wants FFMPEG on PATH for everything it does. On first FaceCam run the node auto-downloads the FaceCam bf16 checkpoints - roughly 16.8 GB (high + low) into ComfyUI/models/diffusion_models/, plus the landmark task bundle and gaussians file - from AEmotionStudio/facecam-wan2.2-14b-bf16. Your base Wan 2.2 A14B pair is separate and still yours to source. Flip allow_model_downloads off if you'd rather place them by hand; the node then fails with a clear message instead of quietly fetching.

Where people get burned

VRAM. FaceCam doubles the temporal dimension, so it needs noticeably more memory than a plain Wan 2.2 run at the same resolution. If you're on a 12–16 GB card, set blockswap_blocks before you start, not after the OOM.

Sweeps past ~45°. The presets deliberately stop at 45° because that's where the face goes into profile and the landmark tracker loses the mesh. custom will happily let you punch in 90° and hand you mush - the tooltip says so, and it means it.

Patience. A plain Wan 2.2 720p 81-frame clip already runs about 25 minutes on a 5090 at full quality without speed LoRAs. FaceCam is heavier on top of that. Budget the run before you queue five of them.