Qwen Image Edit + Latent MXD

Qwen-Image-Edit changed the edit game: instead of "draw a mask, pick a ControlNet, hope the inpaint holds," you hand it a picture and a sentence - "turn the shirt red," "remove the person behind him" - and the change just happens. The KB's panel on it calls this the class of work that "stopped needing a mask, a face adapter, or a control map." Qwen Image Edit + Latent MXD is the pack's bundle node: it takes your prompt, encodes it, and outputs both the conditioning and a matching empty latent - replacing the usual three-node chain (text encode + image encode + empty latent) with one.

What it does

Required: clip (the Qwen-Image-Edit CLIP/checkpoint loader) and prompt (multiline, so write your instruction naturally). Optional: image and vae, plus batch_size (default 1).

• With image but no vae: the image is scaled to ~384px area and fed to the vision encoder - this is how the model sees the reference. The prompt is wrapped in the Qwen-Image-Edit system template ("describe the image, then explain how the instruction should change it") with the image token inserted.
• With image and vae: the image is also scaled to ~1MP, VAE-encoded, and attached to the conditioning as reference latents - this is the img2img-strength ingredient.
• No image at all: it's a plain text-to-image encode, defaults to 1024×1024.

Outputs are CONDITIONING and LATENT. That latent is the quiet sleight of hand: it's built in the SD3-style format (16 channels at 1/8 resolution) sized to match the image, so you don't need a separate empty-latent node and you don't need to keep sizes in sync manually. Wire the conditioning + latent straight into your sampler.

Where it sits

This is the "one less node to think about" play, but it's also a correctness play: the latent and conditioning are generated from the same resolution math, which kills a whole class of "why is my edit at the wrong size" bugs. The node lives in the pack's conditioning family and pairs with the pack's LoRA/prompt tooling for keeping edit runs organized.

The one real gotcha

This node is built on ComfyUI's comfy_api (the newer Nodes 2.0-style API). If it's not in your node menu, your ComfyUI is too old - update ComfyUI first, then reinstall/restart the pack. It also needs the actual Qwen-Image-Edit model loaded through a compatible checkpoint/CLIP node; the node itself downloads nothing.

Install

ComfyUI Manager → search Maxed Out → install → restart, or:

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

No requirements.txt. Find it under MXD/conditioning. If you've been fighting the manual Qwen wiring - or avoiding mask-free editing because the setup looked like a plate of spaghetti - this is the node that collapses the graph down to something you can read at a glance.