TextEncodeSystemEditPlusAdvanced

The "Plus" encoder's fixed three sockets are fine until they're not. TextEncodeSystemEditPlusAdvanced is the same idea with the throttle opened: the three named image1/2/3 inputs are replaced by an autogrowing image_inputs socket, so you can feed five references, or ten, and the node numbers them as it goes. Same chat-template encoding, same single CONDITIONING output - just no ceiling.

One thing to know before you get attached to the name: it's a deprecated alias. The pack's canonical node is the UC_-prefixed version, and ComfyUI may offer to swap this one out when it loads a workflow. The replacement is functionally identical, so take the offer.

What it does

clip, prompt, system_prompt, an autogrowing set of image inputs, and one CONDITIONING out. It's built for the Qwen3-VL-style encoder generation (Krea 2 and friends): the node wraps your text in the chat template - system prompt, then user description, then assistant turn - and passes your images through the VLM's semantic path at the chosen resolution.

The autogrow behavior is where "Advanced" earns its name. The tooltip spells out the contract: images are flattened in ascending socket order, and every image in a connected batch becomes the next sequential image input. That's a genuinely useful property if you've got a batch of frames or a directory of references - wire a batched image list into one socket and each item becomes its own numbered input rather than collapsing into a single collage. Pair that with the pack's numbered-placeholder prompt syntax and you can address specific images from the prompt.

The rest of the schema is the standard structural stack you'll see across this family:

• vlm_resolution - preset combo for the semantic image path: Fast (384) → XX-Large (1536), plus Original. Higher = more detail read from each reference, slower.
• vae_resolution - preset combo for the structural VAE path.
• ref_latent_mode - off / single / multi / parallel-single / parallel-multi. VAE-reference latent conditioning; the parallel-* modes keep it in a separate stream so it doesn't override the semantic meaning. Default off, and off is the right default for most workflows.
• vae_dimension_multiple - pixel alignment (default 8) for reference images before VAE encoding.
• vae - optional, needed only if you enable a reference-latent mode.

Output is one CONDITIONING into your sampler's positive input. No model patching, so it's cheap to iterate.

Install

Via ComfyUI Manager (search ComfyUI-UtilsCollection) or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI. Only real deps: opencv-python and typing-extensions.

When you'd reach for it

When your reference count is "as many as I feel like dragging in" and you want the encoder to keep up. The batch-flattening behavior makes it the better pick than the fixed-socket Plus node for anything video-frame or directory-driven. If you need spatial fusion configs or per-image <Picture N> anchors, keep climbing the family ladder to UC_AdvancedVisualConditioningEncode - but for "encode a growing pile of references," this is the one.