sum_stack_QwenEditPlus

Qwen-Image-Edit is what happens when a capable instruction editor ships with the right license: Alibaba's Apache 2.0 answer to Flux Kontext took over the category, because it could build an unrestricted LoRA ecosystem and Kontext couldn't. sum_stack_QwenEditPlus is the Apt_Preset controller for it - and it's one of the most complete nodes in the pack, because Qwen-Edit genuinely supports the full editing stack: up to three input images, a system prompt that steers the visual-language encoder, a union ControlNet input, and per-image strengths. This is where the pack's instruction-editing energy lives today.

The honest summary from the KB: Qwen-Edit absorbs jobs that used to need a mask plus a ControlNet plus an IP-Adapter. You hand it a picture and a sentence, and the change happens.

What it actually does

It takes the context's model and clip, then runs your images through Qwen-Image-Edit's dual-encoding pipeline - the VL encoder extracts semantic understanding (what the image contains), the VAE encoder handles visual appearance - under the control of system_prompt, and produces the conditioning that makes the model edit rather than generate from scratch. The key parameters, straight from the author's own description:

• vl_size - visual size fed to the VL encoder, default 384. Bigger = richer image features extracted. If edits keep missing fine details (text, small objects), this is the first knob.
• auto_resize - how input images fit the VL window: crop (center crop), pad (black padding), or stretch (forced). Default crop.
• system_prompt - the instruction to the vision encoder: describe the image, then explain how your text instruction should alter it. The default is a solid English prompt; replace it to steer behavior.
• image1/2/3 + image1_strength…3 - up to three reference images with per-image influence (0–10, default 1.0).
• prompt - your edit instruction.
• union_controlnet - optional union ControlNet token for structure-guided edits.
• latent_image / latent_mask - output geometry and an optional edit mask.

Outputs: context, model, positive, negative, latent, clip, vae - the full bundle for the sampler.

Wiring it in

sum_load_simple in QwenEdit mode (Unet + clip1) → sum_stack_QwenEditPlus with your images and instruction → sum_Ksampler.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git
cd ComfyUI-Apt_Preset
pip install -r requirements.txt   # Windows: double-click install.bat

Or ComfyUI Manager → ComfyUI-Apt_Preset. You'll also need the Qwen-Image-Edit weights in models/unet.

Common issues

The KB is upfront about the structural limitation: Qwen-Edit re-emits the whole frame - it returns a new image, not a patched one, so pixels you didn't ask to change move, and faces drift across a chain of edits. The 2026 workflow that fixes it bolts a mask back on around the edit (latent_mask), which this node supports directly - use it for anything where the untouched regions matter. Second, vl_size matters more than people expect: leave it at 384 and small text or fine patterns in your reference may vanish from the VL features. And unlike Kontext-era workflows, the license here is clean - Apache 2.0, no content terms - so the only limits are the model's, not the lawyers'.