Qwen Image Edit Prompt MXD (Triple)

Qwen-Image-Edit is the current community default for "take this picture, change this thing" - you hand it an image plus a sentence and the edit happens, no mask, no ControlNet, no IP-Adapter. This node is the Maxed Out packaging of that workflow, and "Triple" is the differentiator: it accepts up to three reference images in a single encode pass.

The name is doing real work here. Where the standard Qwen-Image-Edit node takes one reference, this one wires image1, image2, and image3 all into the same conditioning - useful for the classic character-consistency move of feeding front, side, and back references of a subject so the edit model knows who it's working with before it starts changing things.

How it works

It's a single node that does the whole conditioning pipeline in one go. Each supplied image gets two treatments: scaled to a ~384px-area for the vision model (so the LLM can "see" it), and, when a vae is provided, scaled to a ~1024px area and VAE-encoded into reference latents. Those get attached to the conditioning via reference_latents, and the text becomes a Qwen-style prompt with <|vision_start|><|image_pad|><|vision_end|> markers inserted for each picture. The output latent is an SD3-style empty latent (16 channels, 1/8 resolution) sized to match the first reference image - so you don't need a separate empty-latent node.

Inputs and outputs

• clip - required, the Qwen-Image-Edit CLIP.
• prompt - required, your edit instruction, multiline.
• batch_size - 1 to 4096, default 1; sets the latent batch.
• vae, image1, image2, image3 - all optional. No images = a straight text-to-latent encode.

Outputs: CONDITIONING (with the reference latents attached, if any) and LATENT - both wire straight into your sampler.

Installing it

Part of the ComfyUI-MaxedOut pack - but read this one carefully: the Qwen nodes only register when comfy_api is available, which current ComfyUI ships with. If the node doesn't appear in your search, that's the culprit.

cd ComfyUI\custom_nodes
git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut.git

Restart ComfyUI, and make sure the Qwen-Image-Edit model files are where your workflow expects them - this node encodes; it doesn't download the model.

Where people get burned

The reference images matter more than the prompt, so garbage references produce garbage consistency - that's a Qwen-Image-Edit property, not this node's fault. And the honest limit from the KB: the model re-emits the whole frame, so pixels you didn't ask about can drift, and faces can degrade on a chain of edits. Use it for single, well-scoped edits with clean references and it's genuinely great.