Qwen Image Edit 2511 ONE STUDIO (TJ)

Qwen Image Edit changed what "editing an image" means: you hand it a picture and a sentence - "make it night, add a jacket" - and it just does it, no mask, no adapter, no ControlNet. Qwen Image Edit 2511 ONE STUDIO (TJ) wraps the 2511 line of that model in the one-node UI of this pack, and adds the bells: text-to-image, img2img, multi-reference editing, inpaint/outpaint, faceswap, a camera-angle drag control, and SeedVR2 upscale.

The "edit with a sentence" model is genuinely a category shift - it's what took the work masks and adapters used to do. It's also structurally loose: it re-emits the whole frame, so pixels you didn't ask about move, and faces drift if you chain edits. That's a model trait, not a bug in this node, and worth knowing before you blame the UI.

How it works

Same engine as every node in this pack: the Python class is a thin shell that returns the last generated image as an IMAGE, and the JavaScript builds and queues the actual ComfyUI graph for the mode you picked - the Edit mode is built on the FluxKontext multi-reference workflow from KJNodes. So the pack-wide rule applies: generate with the node's ▶ Generate button, not ComfyUI's RUN. RUN just re-emits the most recent image through the output slot; click it before generating and you'll get a blank grey 64×64 IMAGE and a moment of panic.

The inputs and outputs that matter

One optional input: prompt_override, a multiline STRING prepended before whatever prompt you type in the panel - your hook for pushing prompts in from a Primitive (String) or TJ_NODE's Prompt Studio. One output: image (IMAGE). Output node, so you can chain it into a preview or save. Everything else is panel + Settings.

The modes to actually use: EDIT takes the original plus up to three reference images and a text prompt; PAINT does masked inpaint (denoise 0.7–0.9 is the README's recommendation) with outpainting included; FACESWAP needs the BFS LoRA and pre-fills a default prompt for you; ANGLE is the party trick - drag H (azimuth), V (elevation) and Z (zoom) and the model re-shoots the scene from that camera position, powered by a Multi-Angle LoRA you set in Settings.

Installing it

ComfyUI Manager (search ComfyUI-TJ_NODE_STUDIO_ONE) or:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

Restart, then install dependencies with the bundled install_requirements.sh / install_requirements.bat. This node specifically needs ComfyUI-KJNodes (the FluxKontext nodes it builds Edit mode from) and ComfyUI_FaceAnalysis plus insightface for faceswap.

Models, from the README's Comfy-Org links: the edit diffusion model (Qwen2.5-VL-7B-Image-Edit-bf16.safetensors) to models/diffusion_models/, the Qwen2.5-VL vision encoder to models/text_encoders/, the ae.safetensors VAE to models/vae/, and the LoRAs - BFS for faceswap (required for that mode), the Multi-Angle LoRA (required for ANGLE), and the optional Lightning LoRA if you want 4-step fast generation. Low VRAM? There's a GGUF line via unsloth, which needs ComfyUI-GGUF installed.

Where people get stuck

The classic failure here is the missing FluxKontext node error - that's KJNodes not being installed or being stale, so update it rather than hunting a phantom. If faceswap finds no face, check insightface actually installed (the README lists this exact symptom). If ANGLE does nothing, you forgot the Multi-Angle LoRA in Settings - ANGLE won't work without it.

And the one mental adjustment that'll save you grief: because this model edits by re-emitting the whole frame, "everything else moved" is normal on a big edit, especially if you didn't mask. The 2511 line is the newest open weight of Qwen-Image-Edit, which has meant the same thing all along - for a beginner it's the friendliest "describe the change and get it" experience in ComfyUI, just keep expectations calibrated about the parts of the photo you told it to leave alone.