System Edit Text Encode (Advanced)

The stock TextEncodeSystemEditPlus from ComfyUI core is how Qwen-Image-lineage edit models take "here's the image, here's the edit request." This node is that node with the limits removed: your own system prompt, up to sixteen autogrowing image sockets, batches flattened into sequential inputs, and full control over the VLM-versus-VAE reference paths. If you've hit the wall where a single reference image wasn't enough for the edit you had in mind, this is the encode you want.

How it works

The inputs tell the story:

• clip - the Qwen-Image-style edit encoder.
• prompt + system_prompt - the edit request and the role instruction. The system_prompt is the big upgrade over the plain core node: a persona like "you are a photo editor, keep composition identical" materially changes how the encoder frames the edit.
• image_inputs - autogrow image sockets, flattened in ascending order, and every image in a connected batch becomes the next sequential input. That's the multi-image and batch superpower: one node, many references, no fanning out by hand.
• vlm_resolution - how big the image gets for the semantic path (Fast 384 → XX-Large 1536, or Original). 384 is fast; 768+ is for detail-dependent edits.
• vae_resolution - the structural reference-latent resolution (Ultra 512 → Detailed 1536 / Original).
• ref_latent_mode - single/multi append reference latents to the conditioning; parallel-single/parallel-multi run them in a separate stream so the reference can't semantically override your edit.
• vae_dimension_multiple (8) - alignment multiple before VAE encoding.
• vae - optional, needed only for reference-latent modes.

Output: one CONDITIONING tensor. If the prompt contains image_input_N tokens, the flattened images map onto them in order - so you can reference "image_input_1" and "image_input_2" explicitly in your edit instruction.

Installing it

It's part of ComfyUI-UtilsCollection (author: silveroxides):

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

Restart ComfyUI, or install "ComfyUI-UtilsCollection" via Manager. The pack's requirements are just opencv-python and typing-extensions - encoding runs through core's CLIP machinery.

Where people get tripped up

The parallel-* modes are semantic switches, not speed options - use them when the model keeps redrawing the reference instead of applying your edit. And this is a memory-heavy node by construction: two image paths plus a big VLM encoder can OOM fast if you set both resolutions to "Large." Keep the VLM path at the resolution your edit actually needs. Also - if your workflow came from an older Core node of the same family, the pack keeps the legacy node ID alive, so your old workflows load; but the UC_ version is the one with the system prompt and the autogrow sockets. Use it.