Text Encode Qwen Image Edit Plus (Batch)

Qwen-Image-Edit is the open instruction editor that, in roughly a year, absorbed most of what masks and ControlNets used to do: you hand it a picture and a sentence, and the change happens. Text Encode Qwen Image Edit Plus (Batch) is the ComfyUI node that lets you do that to a whole batch of textures at once - hand it several reference images and one instruction, and it bakes all of them into a single conditioning. "Add rust to the metal," applied to eight tiles in one pass.

How it works

ComfyUI ships a built-in "Text Encode Qwen Image Edit Plus" node for single-image editing. This is a batch-aware wrapper with the same behavior, and the mechanism is worth understanding because it's where the power comes from:

• Every image in the images batch gets downscaled to ~384×384 and appended into the prompt as Picture 1: <vision tokens>, Picture 2: ..., and so on, followed by your instruction. The Qwen2.5-VL vision encoder reads those tokens, so the editor "sees" every reference image.
• If you also connect a vae, each image is additionally encoded at ~1024×1024 into reference latents, which get attached to the conditioning. That dual path - semantic (vision) plus appearance (latents) - is exactly how Qwen-Image-Edit is architected, and it's what keeps the edits grounded in the source pixels.

The output is a standard CONDITIONING that goes straight into a KSampler with the Qwen-Image-Edit checkpoint loaded. The clip input takes the model's text encoder.

The context you need

Two things shape how you use this. First, the model family: you need the Qwen-Image-Edit checkpoint and its CLIP loaded in ComfyUI - this node doesn't bring any weights, it just encodes. Second, the honest limit from the community: an edit model re-emits the whole frame, so pixels you didn't ask about move a little, and drift compounds across a chain of edits. For texture work that's often fine (you're editing the whole tile), but if you need untouched pixels, mask the result or fall back to crop-and-stitch - the same pack's inpainting nodes exist for exactly that.

Installing it

Part of amtarr/ComfyUI-TextureAlchemy. ComfyUI Manager → search "Texture Alchemy", or:

cd ComfyUI/custom_nodes
git clone https://github.com/amtarr/ComfyUI-TextureAlchemy

Restart, then find it under Conditioning. The node itself has no dependencies beyond ComfyUI's own Qwen support - but you're on the hook for the Qwen-Image-Edit model files, which are a few GB and downloaded separately.

Where people get burned

The batch input is an IMAGE batch, not a folder or a list - you need to assemble your textures into a batch (stack images) before connecting. And because each reference image costs a full vision-encode plus optional VAE encode, batches chew VRAM and time; a batch of eight 1024² images with the VAE connected will make your GPU feel it. Keep the VAE connected only when appearance grounding matters, and expect the pack's console banner on each run.