LLS Qwen Image Edit

Qwen-Image-Edit is the open instruction editor that absorbed a huge chunk of image work: hand it a picture and a sentence - "change her dress to blue," "remove the watermark" - and it does the thing, no mask, no ControlNet. LLS Qwen Image Edit is a compressed wrapper for running it in ComfyUI, one node instead of the usual multi-node assembly. It loads the Qwen companion resources internally, supports the official multi-image edit conditioning (up to three input images), and even offers turbo/lightning LoRA hooks.

The basics are simple: pick the model in model_name (from your diffusion_models folder), connect the image you want to edit, type prompt, hit run. image2 and image3 are optional - that's the multi-image conditioning, for person+person or person+scene edits. An optional model input lets you chain an external (often quantized) MODEL instead of the node loading its own.

The sampling knobs

Same Qwen-family controls as its txt2img sibling: cfg (default 4), steps, sampler_name, scheduler, shift (3.1), plus cfg_norm_strength and - the one that only exists here - reference_latents_method, with offset, index, uxo/uno, and index_timestep_zero options. This controls how the reference image latents are attached to the conditioning, and the default index_timestep_zero is the sensible starting point. enable_turbo_mode + turbo_lora_name / turbo_strength apply a step-reduction LoRA when you want speed.

The honest warning

The KB's Qwen-Image-Edit panel puts the model's core weakness in one sentence: it re-emits the whole frame, so pixels nobody asked about come back "close but not identical," and the drift compounds across a chain of edits. The pack's Pro Image Edit chain and the community's crop-and-stitch pattern exist precisely because of this. If you need the rest of the image untouched, mask the edit region and composite - this node alone won't protect the unmasked pixels. There's also the resolution quirk: Qwen-Image-Edit output is known to be slightly misaligned or blurred if the input isn't a clean multiple of the model's preferred resolution.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Gin3601/LLS-node
# restart ComfyUI

Or ComfyUI Manager → search "LLS-node". No pip deps, but the model is the heavy lift - 20B at bf16 won't fit a consumer card, so community GGUFs and fp8 are the practical route. The dropdown shows "(no qwen image edit models found)" until a Qwen edit checkpoint lands in ComfyUI/models/diffusion_models/.

Gotchas

It needs a ComfyUI build with official Qwen image support (comfy_extras.nodes_qwen) - an old build errors on import. And remember the model file must be the edit variant, not the base Qwen-Image text-to-image model; the node's discovery filters by filename, so name your files so the "edit" is in the name.