Nodes/ERPK Collection/Qwen Image Max Edit
ComfyUI Node

Qwen Image Max Edit

Six reference images and one instruction, $0.07 per pass

By eRepublik-Labs·Created 11 months ago·Updated 21 days ago· 1
Qwen Image Max Edit
  • client
  • image
prompt
images
width1024
height1024
seed-1

This is the edit half of Qwen Image Max, and its superpower is the input list: up to six reference images in one edit call. Most edit models top out at one or three references - six means you can feed a character sheet, a style sample, a location, and a prop, then ask for a single coherent image combining them. It's the "give me the kitchen from image 1, the chef from image 2, the lighting from image 3" workflow, and the premium tier is the one that holds all of that together.

It runs through the WaveSpeed AI API via the ERPK Collection, the community pack wrapping the service for ComfyUI. Billed at $0.07 per edit, no local compute.

How it works

You provide images - comma-separated URLs or file paths, up to six - plus a text instruction, and the node sends the bundle to WaveSpeed's Qwen Image Max edit endpoint. The model reads all references and returns a single edited image as a ComfyUI image IMAGE tensor. Note the format of the input: this node takes URLs or paths as a string, not an IMAGE tensor socket, so you'll typically source references from somewhere web-accessible (or a local path the API can read).

Because Max is a full 20B model, it treats the text instruction as the boss: "swap the car in image 1 for the car in image 3, keep the lighting from image 4" is the kind of multi-hop edit it actually handles.

The inputs that matter

  • images - your reference stack, comma-separated. Required, up to six. Order matters if your instruction refers to "image 1", "image 2", so keep a stable order.
  • prompt - the modification. The more explicit about which reference does what, the better the result.
  • width / height - 256 to 1536, step 8, default 1024×1024. Output canvas; editing doesn't upscale your references.
  • seed - real seed for reproducibility, -1 for random.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

Or ComfyUI Manager → search erpkERPK Custom Nodes. WaveSpeed key with credits via right-click canvas > ERPK Settings (Settings > ERPK > API Keys).

Common issues

  • References not loading - URLs must be reachable by WaveSpeed's servers, and paths must be valid. A local C:\... path from another machine won't resolve.
  • The model ignores one reference - six is a lot of context; the instruction has to explicitly name each image's role. Vague phrasing gets vague borrowing.
  • Small output size - 1536px ceiling again. For bigger edits, the Seedream edit nodes in this pack go to 4096.
  • Cost - $0.07 per edit. With six references and a complex instruction, don't burn credits on shots you haven't storyboarded.

Where this earns its keep: character/consistency-heavy work where a single reference isn't enough. If you only need one image changed, the cheaper single-reference Qwen edit nodes are the more sensible default - Max Edit is for when the brief is genuinely multi-input.

CategoryERPK/WaveSpeedAI

Inputs (6)

NameTypeDefaultDescription
promptSTRINGText description of the desired image modifications
imagesSTRINGReference images to edit. Maximum of 6 images (comma-separated URLs or paths)
clientoptWAVESPEED_AI_API_CLIENTWaveSpeed API client (optional if API key is configured in Settings)
widthoptINT1024256–1536Image width (256 to 1536)
heightoptINT1024256–1536Image height (256 to 1536)
seedoptINT-1-1–2147483647Random seed for reproducibility (-1 for random)

Outputs (1)

NameTypeDescription
imageIMAGE