NodeWise Studio
RUDownload for Linux

Editing a picture by words

<!-- lesson: edit-words -->

Lesson "edit-words", in the group Pictures: techniques of the lesson list. Steps: 13.

What you will be able to do

A model made for editing changes a picture as an instruction says: "make the sofa green", "remove the cup".

What you need

  • It goes on from Qwen-Image: words inside the picture. The card says so; it is a hint, not a lock.
  • Nodes: Load Diffusion Model, Load CLIP, Load VAE, ModelSamplingAuraFlow, Load Image, FluxKontextImageScale, Text Encode Qwen Image Edit Plus, VAE Encode, KSampler, VAE Decode, Save Image / Preview Image.
  • The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

The steps

The words of each step, as the lesson shows them. Node names are in bold.

  1. What an editing model does (reading)

    An editing model gets a picture and an instruction and changes only what the instruction is about. Unlike inpainting there is no mask: the model understands itself what to change.

    This lesson builds Qwen-Image-Edit 2511 the way ComfyUI's official template does, from an empty canvas: "File ▸ New project". The model is large: the bf16 file is about 40 GB, the fp8 one about 20 GB.

  2. Load the three parts

    Add Load Diffusion Model, Load CLIP and Load VAE.

    • In unet_name of Load Diffusion Model: the qwen_image_edit_2511 file, from models/diffusion_models.
    • In clip_name of Load CLIP: qwen_2.5_vl_7b, from models/text_encoders. Set type to qwen_image.
    • In vae_name of Load VAE: qwen_image_vae.

    The text reader of this model is itself a model that sees pictures: that is how it understands an instruction about a picture.

    If the file you need is not in the list, your ComfyUI does not have it. The app downloads nothing. The easiest way is to open, in ComfyUI itself, the official template of Qwen-Image-Edit: it shows which files are missing and where to put them. Then choose "Sync ▸ Sync now" in this app's menu, and the files appear in the lists.

  3. Set the noise shift

    Add a ModelSamplingAuraFlow node and pass the model through it: the output of Load Diffusion Model into its model input. Set shift to 3.1, as the template does.

    The number changes how much noise is left at each step; this family is set up with this node.

  4. Load the picture to edit

    Add Load Image and FluxKontextImageScale. Choose the picture in image of Load Image and connect its output to image of FluxKontextImageScale.

    FluxKontextImageScale brings the picture to about one megapixel with sides the model takes. A picture of another size would be cut or distorted.

  5. Write the instruction

    Add a Text Encode Qwen Image Edit Plus node and connect to it: clip from Load CLIP, vae from Load VAE, image1 from FluxKontextImageScale. Write the instruction in its prompt field, in English or Chinese: "Change the color of the sofa to dark green."

    The node shows the picture to the text reader and to the model together with the words.

  6. The second text

    Add a second Text Encode Qwen Image Edit Plus node with the same three links and leave its prompt empty.

    The sampler wants two prompts. An empty one with the same picture means "nothing in particular", as in the template.

  7. The starting latent

    Add an VAE Encode node: the output of FluxKontextImageScale into pixels, the output of Load VAE into vae.

    The sampler starts from your picture, so the result keeps its size.

  8. The sampler

    Add a KSampler node and connect: model from ModelSamplingAuraFlow, positive from Text Encode Qwen Image Edit Plus, negative from Text Encode Qwen Image Edit Plus, latent_image from VAE Encode. Set steps to 40 and cfg to 4, and leave denoise at 1.

    The picture is kept through the texts, which carry it; that is why denoise stays 1.

  9. Decode and save

    Add VAE Decode and Save Image: the output of KSampler into samples, the output of Load VAE into vae, the picture into images.

  10. How to write instructions (reading)

  • One change at a time: "replace the cup with a vase of tulips". Several changes at once often lose some.
  • Name what must stay: "keep the face and the pose unchanged".
  • Words on a sign go in quotes: change the sign to "OPEN".
  • image2 and image3 take more pictures: "put the jacket from image 2 on the man in image 1".
  1. Faster: four steps (reading)

The Lightning LoRA of this model, Qwen-Image-Edit-2511-Lightning-4steps, cuts the work to 4 steps. Put it into a LoraLoaderModelOnly node between Load Diffusion Model and ModelSamplingAuraFlow, then set steps to 4 and cfg to 1. The template has both ways behind one switch.

  1. Other editing models (reading)

ComfyUI's templates show several families:

  • Flux.1 Kontext dev: the picture goes in as a reference latent;
  • Flux.2 Klein: small and fast, edits by one or several pictures;
  • Qwen Image 2.1: the newest of the Qwen family, with a text node of its own. Each has its own nodes; open its template in ComfyUI to see the graph.

The templates also have paid API nodes for editing. They cost money and send your picture and your words to a cloud service.

  1. The graph is ready (reading)

Save the graph with "File ▸ Save workflow for ComfyUI…", open the file in ComfyUI and run it there. This app runs nothing in ComfyUI itself.

What to try: the same instruction with two seed values; an instruction that names what must stay.

What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

  • Step 2, Load the three parts:
    • a Load Diffusion Model node.
    • a Load CLIP node.
    • a Load VAE node.
    • unet_name of Load Diffusion Model: a value is chosen.
    • clip_name of Load CLIP: a value is chosen.
    • type of Load CLIP is qwen_image.
    • vae_name of Load VAE: a value is chosen.
  • Step 3, Set the noise shift:
    • a ModelSamplingAuraFlow node.
    • MODEL of Load Diffusion Model into model of ModelSamplingAuraFlow (a LoraLoaderModelOnly in between is accepted).
    • shift of ModelSamplingAuraFlow is from 2 to 4; Do this for me sets 3.1.
  • Step 4, Load the picture to edit:
    • a Load Image node.
    • a FluxKontextImageScale node.
    • IMAGE of Load Image into image of FluxKontextImageScale.
    • image of Load Image: a value is chosen.
  • Step 5, Write the instruction:
    • a Text Encode Qwen Image Edit Plus node.
    • CLIP of Load CLIP into clip of Text Encode Qwen Image Edit Plus (prompt).
    • VAE of Load VAE into vae of Text Encode Qwen Image Edit Plus (prompt).
    • IMAGE of FluxKontextImageScale into image1 of Text Encode Qwen Image Edit Plus (prompt).
    • prompt of Text Encode Qwen Image Edit Plus (prompt) holds at least 5 characters.
  • Step 6, The second text:
    • a Text Encode Qwen Image Edit Plus node.
    • CLIP of Load CLIP into clip of Text Encode Qwen Image Edit Plus (negative prompt).
    • VAE of Load VAE into vae of Text Encode Qwen Image Edit Plus (negative prompt).
    • IMAGE of FluxKontextImageScale into image1 of Text Encode Qwen Image Edit Plus (negative prompt).
  • Step 7, The starting latent:
    • a VAE Encode node.
    • IMAGE of FluxKontextImageScale into pixels of VAE Encode.
    • VAE of Load VAE into vae of VAE Encode.
  • Step 8, The sampler:
    • a KSampler node.
    • MODEL of ModelSamplingAuraFlow into model of KSampler (a CFGNorm or LoraLoaderModelOnly in between is accepted).
    • CONDITIONING of Text Encode Qwen Image Edit Plus (prompt) into positive of KSampler (a Apply ControlNet or InpaintModelConditioning or Conditioning (Combine) or ConditioningSetArea or ConditioningSetAreaPercentage or ConditioningSetMask or Apply Style Model or FluxGuidance or ConditioningZeroOut or FluxKontextMultiReferenceLatentMethod in between is accepted).
    • CONDITIONING of Text Encode Qwen Image Edit Plus (negative prompt) into negative of KSampler (a Apply ControlNet or InpaintModelConditioning or Conditioning (Combine) or ConditioningSetArea or ConditioningSetAreaPercentage or ConditioningSetMask or Apply Style Model or FluxGuidance or ConditioningZeroOut or FluxKontextMultiReferenceLatentMethod in between is accepted).
    • LATENT of VAE Encode into latent_image of KSampler.
    • steps of KSampler is from 4 to 50; Do this for me sets 40.
    • cfg of KSampler is from 1 to 5; Do this for me sets 4.
  • Step 9, Decode and save:
    • a VAE Decode node.
    • a Save Image or Preview Image node.
    • LATENT of KSampler into samples of VAE Decode.
    • VAE of Load VAE into vae of VAE Decode.
    • IMAGE of VAE Decode into images of Save Image or Preview Image.

See also

This page is imported from the application repository's wiki (docs/wiki/lessons/Edit-By-Words.md, revision b8e3f71). The same guide is built into the app: Help ▸ User guide.

To refresh it, run npm run import and then npm run build in the website folder.