🔡 QwenEdit TextEncoder Plus

This is the multi-image version of the pack's QwenEdit TextEncoder. Where the plain node takes one reference image, this one takes up to three - image1, image2, image3 - which is what you need for anything Qwen-Image-Edit's newer revisions actually support: combining a person and a product shot, compositing a person into a scene, or blending two references into one edited output, in a single instruction rather than a chain of separate edits.

Why three images, and why it matters

Multi-image editing landed in the 2509 revision of Qwen-Image-Edit specifically - the model went from single-image instruction editing to accepting one to three inputs at once, with the community immediately using it for exactly the combinations you'd expect: person+person, person+product, person+scene. That's a meaningfully different capability from running three separate single-image edits back to back, because the model reasons about all the references together in one pass rather than compounding drift across several sequential edits. If you've hit the "unrelated pixels shift after a few chained edits" problem with the plain QwenEdit TextEncoder node, doing the composite in one Plus call instead of three chained single-image edits is the more reliable path.

The inputs and outputs that matter

• prompt - the instruction describing how the references combine ("put the person from image1 into the scene from image2").
• clip - the text/CLIP model input.
• image1, image2, image3 (all optional) - up to three reference images. Leave any unused ones unwired.
• img_size - target size cap (0–4096, default 0 meaning no resize).
• vae (optional) - needed for the latent output to reflect the actual reference images.
• translate - auto-translate the prompt from over a hundred source languages.
• seed - for reproducibility.

Four outputs, same shape as the plain node: conditioning, width, height, and latent - ready to feed straight into a KSampler without a separate Empty Latent Image step.

Installing it

ComfyUI Manager: search SDVN_Comfy_node, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node

Then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from your ComfyUI root, and restart. Same requirement as the plain QwenEdit node: you need the Qwen-Image-Edit checkpoint (the 2509 revision or newer for reliable multi-image support) loaded separately - this node builds the conditioning, it doesn't ship the weights.

Where people get burned

No mask input here, unlike the single-image node. Plus trades the mask slot for the two extra image slots. If you need to constrain exactly where an edit lands in a multi-reference composite, you're relying on the prompt wording to do that work rather than a mask - be explicit about which region each instruction applies to.

Older Qwen-Image-Edit checkpoints don't support multi-image reliably. Multi-image editing is a 2509-and-later capability. Point this node at an earlier revision of the model and the extra reference images may be effectively ignored or produce weaker results than the newer checkpoint would.

Which image is "first" matters to the model's framing. The instruction usually reads more naturally, and tends to compose more predictably, when your prompt's references (image1/image2/image3) match the order you describe them in the prompt text - mismatch the order and the model can attribute the wrong instruction to the wrong reference.