Qwen-Image-Edit-2511

Qwen-Image-Edit-2511 is the model that won instruction editing - the December 2025 revision the community standardized on because it fixed image drift, got character consistency right, and folded popular LoRAs into its base weights. It is also a 20B model that eats your VRAM and your patience. This node runs that exact open model on Runware's hardware so you get the editor's behavior without meeting its hardware requirements. If you've been eyeing the GGUF downloads and doing the math on whether your card can hold it, this is the version that skips the math.

Here's the workflow shift this family caused: editing stopped needing a mask, a ControlNet map, or an inpainting pass. The model takes an image plus a plain-language instruction - "change her dress to blue," "remove the watermark," "rotate this object" - and returns the edited frame, re-emitted whole. The flip side of that convenience is documented by its own users: because it re-renders everything, unedited regions come back close but not identical, and errors compound across long edit chains. Keep one input multiple of 112 to avoid the old geometric offset, and for surgical changes the mask-based inpainting route is still the guarantee if this feels too loose.

Inputs are honest about what they do. referenceImages is required and accepts one image or several (multi-image editing is where 2511 shines - swap a character, combine a person and a product). positivePrompt is the instruction, in a sentence. The extra controls this node exposes that the pure T2I Qwen nodes don't are the real reason to read this page: steps (30 default), a scheduler dropdown, CFGScale_value, and settings.trueCFGScale_value, which is the true-CFG knob for prompt adherence at the cost of quality. There's also a lora socket, one of the pack's typed builder sockets - stack the Runware LoRA builder into it to load a fine-tune on the fly. advancedFeatures.watermark.* and the safety toggles do what they say. The output is image.

Install and key

One pack, one setup: search Runware in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/Runware/ComfyUI-Runware && pip install -r ComfyUI-Runware/requirements.txt, then restart. Add your Runware API key in Settings → Runware API key, via RUNWARE_API_KEY, or runware auth login. No checkpoints download; the heavy model lives on Runware's side, which is exactly the point of the node.

When to reach for it

Run this when you want the Qwen-Edit behavior - face and identity preservation that still beats most rivals, real text-in-image editing - but don't have the hardware, or don't want to manage a 20B quantized install for occasional edits. It's per-call billed, so it suits editing jobs that are worth it rather than draft spam. One honest caveat: since 2511 has open weights, the "why not run it locally" question is legitimate, and the answer is usually "my card says no" or "I want it inside this workflow without the setup." Your image leaves the machine and passes through Runware's moderation either way; if that's a problem for what you're editing, stay local.