IAMCCS Influencer Latent / VAE

The "Influencer" pair of nodes exists to let the IAMCCS AI Influencer applications (Studio, Identity, and the Krea graph) run on stock ComfyUI without depending on someone's VAE utility pack. IAMCCS_InfluencerLatent is the encode side: it takes up to two images, a mask, and a VAE, resizes everything to a common canvas, and hands you latent_1, latent_2, a mask, and - this is the output the apps actually consume - an empty_latent ready for sampling.

The one-line description in the source is honest: "Studio and Identity consume empty_latent (output index 3)." So the headline feature is the empty latent, which defaults to 16 channels (latent_type default) - the shape Krea-style and Flux-family editing graphs want - with 4_channels available as a compatibility mode for older workflows. Don't overthink the channel count; match it to the model you're sampling or you'll get shape errors at the sampler, not here.

The inputs that matter:

• width / height (1024 default, step 8) plus resize_to_input - the manual size, or "copy the size from image_1 / image_2 / mask / latent instead." The auto modes are the ones you'll use most: they let you build a graph where one connected image defines the canvas and you never touch the width boxes again.
• image_1 / image_2 + vae - the two reference images. Connect both and the node encodes both to their own latents; vae is required only if an image is actually connected (leave both empty and it just emits empties).
• mask + mask_channel (red/green/blue/alpha) + invert_mask - the mask image gets resized and one channel extracted into a proper ComfyUI MASK. Great for turning a reference image's alpha into an inpaint region.
• interpolation (lanczos default) and method (stretch or fill / crop) - resizing behavior. fill / crop is the one that keeps aspect and crops, which is what you want for faces; stretch only when aspect literally doesn't matter.

Outputs: latent_1, latent_2 (VAE-encoded, or empty tensors if no image), mask (single-channel, possibly inverted), empty_latent, and the resolved width/height as INTs (useful to pipe into other nodes that need the canvas size). Feed latent_1/latent_2 into whatever sampling graph follows, or skip them entirely and just use empty_latent - both patterns are legitimate.

Installation is pack-standard: search IAMCCS in ComfyUI Manager, or clone https://github.com/IAMCCS/IAMCCS-nodes.git into custom_nodes and restart. No models ship with it - the vae comes from your own model folders, and for the Krea side you want a Qwen-Image VAE (that's what Krea 2 uses; the KB notes its VAE is the Qwen-Image one).

The gotchas are subtle and mostly about expectations. resize_to_input honors the dimensions of the chosen input, not its contents, so a 4:5 image in a square canvas stays square-cropped unless you change method. And the two-image path encodes each image independently to its own latent - this node does not concatenate or blend references, it just gets them both ready. If you were expecting a fused identity latent, that blending happens downstream in the sampler's patch, not here.