TextEncodeSystemEditPlus

Some encoder nodes chase maximum flexibility; this one chases maximum simplicity. TextEncodeSystemEditPlus is the entry-level "system prompt + a few reference images" encoder in silveroxides' UtilsCollection, and its whole personality is that it gives you exactly three fixed image sockets - image1, image2, image3 - plus a system prompt field, and nothing else to configure. For a beginner who just wants "encode my text and these three pictures," that's a feature, not a limitation.

The honest caveat first: it's a deprecated alias. The pack's canonical encoder is UC_-prefixed (and the newer ones use autogrowing image inputs instead of fixed sockets), and this class survives so old workflows keep loading. If ComfyUI asks to replace it, let it - but this article's walkthrough applies to the new node too, because the inputs you actually fill in are the same.

What it does

clip, prompt, system_prompt, three images, and out pops one CONDITIONING. The node builds the chat-style template your LLM-based encoder expects - system prompt first, user description after - then encodes text and images together through the VLM's semantic path. Which encoder it works with is the real question: this family targets the current generation of Qwen3-VL-style encoders (Krea 2 and friends), not SDXL-era CLIP. Plug in the wrong checkpoint family and you'll get garbage; plug in the right one and you get one-shot multi-image conditioning without hand-rolling template syntax.

The remaining knobs are the pack's standard structural stack:

• vlm_resolution - preset combo for the semantic image path: Fast (384), Balanced (512), Detailed (768), Large (1024), X-Large (1280), or Original. Default Fast; bump it when a reference's details matter.
• vae_resolution - preset combo for the structural VAE path (Fast 1024 → Original).
• ref_latent_mode - off, single, multi, parallel-single, parallel-multi. This controls whether reference images also get VAE-encoded into structural latents. Leave it off unless you specifically want that; parallel-* modes run them in a separate conditioning stream so they guide composition without overriding semantics.
• vae_dimension_multiple - pixel alignment for reference images before VAE encoding (default 8).
• vae - optional; only needed if you turn on a reference-latent mode.

The output CONDITIONING feeds the sampler's positive input. Nothing patches the model, so it's cheap to iterate.

Install

ComfyUI Manager → search ComfyUI-UtilsCollection → install, or:

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

Restart. Deps are opencv-python and typing-extensions; no model downloads for the node itself.

When you'd reach for it

When three fixed sockets are all you need and you don't want an autogrow widget sprawling across your graph. The moment you need a fourth image, or numbered <Picture N> anchors, or pixel-level formula blending, graduate to the Advanced variant - that's the exact ladder the pack is designed for. Think of this one as the training-wheels version that still does real work.