LTX Multiple Controls

LTX Multiple Controls is the node you use when one reference just isn't doing the job - when you need a guide video for motion, a mask for the replacement region, and a separate identity reference for who actually appears in it, all at once. It's built for the scail2v2-style LTX 2.3 identity/Edit-Anything LoRAs this pack supports, and it exists because of a nasty trap: chaining the pack's single-reference node (LTX Identity Transfer / LTXIdentityOverlapConditioning) in sequence does not combine the references. Each call overwrites the previous one's spec, so you're left with only the last thing you wired. This node takes all the slots and merges their RoPE reference specs in a single call, so they actually coexist.

How it works

Each slot - guide, mask, identity, identity_mask - gets its own source_id, phase_scale, layout, resize mode, and downscale factor. Nothing is hardcoded; the defaults just match the v13 training recipe the author tuned against: guide and mask at source_id 0 (no RoPE phase tag, because they line up frame-for-frame with the target), identity at source_id 2 (phase-tagged, because it has no positional correspondence with the target). If that paragraph reads like coordinates-speak, fine - the short version is "these defaults are what the checkpoint learned, change them only for a different recipe."

The one setting that makes or breaks identity transfer is auto_mask_guide, on by default. It paints guide_video black everywhere mask_video is white - an inpainting hole - so the model can't just copy the guide's own person into the replacement region and must pull identity from identity_image. That's how the training data was built, and the tooltip is blunt about it: this is the fix that made identity transfer actually work.

Inputs and outputs that matter

Required: model (after loading the LoRA), positive/negative, vae, latent - standard stuff. The optional slots are where the action is:

• guide_video - the motion/structure driving clip (an IMAGE batch, e.g. from GetVideoComponents). Unconnected = skip the slot.
• mask_video - per-frame replacement-region mask, same frame count as the guide. For the v13 recipe this must be a plain white-on-black binary silhouette; the tooltip explicitly warns not to use a colored mask - it leaks hue into generated clothing. If a recipe tells you to color masks, that's a different slot.
• identity_image - the appearance reference (face/character) that fills the hole. This is what you're usually after.
• identity_mask_image - a small flat color dot from LTX Color Mask (flat_marker style), not a silhouette. Unused by v13 - leave it unconnected unless your checkpoint is older.
• identity_temporal_offset_latents - default 0 (standard frame-zero placement). A negative value is the experimental knob for reducing the "reference briefly appears at the start of the clip" artifact.

Outputs are model, positive, negative, latent (straight into the sampler), plus a debug STRING that logs each active slot's spec. Turn on debug_log if the console is quiet.

Installing it

Same as every node in this pack - it's alisson-anjos/ComfyUI-BFSNodes:

cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes.git
cd ComfyUI-BFSNodes
pip install -r requirements.txt

Or ComfyUI Manager → search "BFSNodes". Restart after. requirements.txt drags in insightface, onnxruntime, librosa and opencv-python for the pack's face/audio nodes even though this node doesn't touch them - the insightface pin is the most common install snag, and it's unrelated to the LTX work.

Common issues

Identity ignored entirely → wrong layout or downscale_factor. These must match the checkpoint's training recipe - wrong values aren't a quality hit, they're coordinate positions the model never learned. The node's own debug log even reminds you: connect negative and run CFG 3–5, and skip LightX2V LoRAs. If you crank reference_guidance_scale above 1, know it costs an extra forward pass per step (that's ST-DRC-style reference-CFG, same units as sampler CFG).

Also temper expectations on hardware. These LTX 2.3 identity workflows are genuinely VRAM-hungry - people have reported OOMs even on a 4090 at 768px with the EditAnything LoRAs - so budget for fp8/GGUF weights and maybe a lower resolution before blaming the node.