Fat Mex Image Edit Sampler

The old way of editing an image - paint a mask, hope the inpaint holds, re-roll forever - has mostly been replaced by telling a model what to change in a sentence. This node is that workflow, collapsed into one box: reference images in, an edit described in plain English, an edited image out.

It wraps the entire Qwen Image Edit stack - TextEncodeQwenImageEditPlus + EmptyQwenImageLayeredLatentImage + KSampler + VAEDecode - which is exactly why it's fiddly to build by hand and why you'd rather have it as a single node. Qwen-Image-Edit is the default open instruction editor of this era (Apache 2.0, no Kontext-style content policing), and it handles object removal, garment swaps, relighting, re-posing, and face edits from a sentence, no mask and no ControlNet.

How it works

Under the hood it replicates TextEncodeQwenImageEditPlus conditioning. Each reference image gets two jobs:

1. A small 384×384 version goes into the vision-language encoder, so the Qwen2.5-VL text encoder actually sees the picture and understands "image 1" when you say it.
2. A ~1MP reference latent is attached as identity conditioning, so the model has real pixel-level information about the subject, not just a fuzzy thumbnail.

It then builds the Qwen layered latent - a 5D tensor with layers + 1 layers, one per reference plus the output - samples, decodes, and keeps only the final layer (the intermediate ones are internal reconstruction passes).

The inputs that matter

• prompt - describe the edit and refer to your references as "image 1", "image 2", "image 3". That's the contract: image1 is "image 1" in the prompt.
• image1 / image2 / image3 - the references. One is often enough; three lets the model compose (subject + outfit + scene, for example).
• layers - the Qwen layered-latent depth, default 3, "usually = number of images." If you use two references, 2-3 layers is right.
• model, clip, vae - from the Fat Mex Model Loader, and here's the catch: load a Qwen Image Edit preset (2509 or 2511), not Klein.
• width / height - output size; steps, cfg, preset_hint behave like the other samplers (0 / auto = preset defaults).

Outputs: images (the edit) and latent.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/FatMex/ComfyUI-FatMex-Nodes.git FatMex-Nodes
pip install -r FatMex-Nodes/requirements.txt

Or ComfyUI Manager → search "Fat Mex Nodes" → Install → restart. You'll also need the Qwen Image Edit model files in your ComfyUI/models/ folders (the loader presets name them exactly).

Where people get burned

• Wrong model loaded. If you loaded a Klein or Z-Image preset, this node is sampling with an architecture that doesn't do layered Qwen editing. The model, clip, and vae inputs expect the Qwen Edit trio. This is the #1 cause of garbage output here.
• Prompt says "image 2" but no image2 is connected. Referencing a picture you didn't wire in just confuses the encoder. Keep prompt references and connected inputs in sync.
• It edits the whole canvas. Like all instruction editors, Qwen-Image-Edit takes the context and emits a new image - regions you didn't mention will drift slightly. For surgical, mask-locked changes you want the classic mask route; for "change the shirt / add sunglasses / different background," this is the right tool and the KB says most of the community moved to it.
• Negative prompt is real here. Unlike the pack's inpaint variant (which zeroes the negative for quality), this node actually encodes your negative text - but on Qwen Edit models it's a light touch, so don't rely on it for structural fixes.