Wan 2.7 (BudgetPixel)

Wan is Alibaba's video-and-image family, and the Wan 2.7 image model is the one that walks the line between "generate" and "actually plan what I meant." This node is the official BudgetPixel wrapper around it - and note up front, the display name says "Wan 2.7" but there are two nodes with that name in the pack: this one (image, category BudgetPixel/Image) and a separate video node (BudgetPixel_wan-2.7-video). If you search the menu and grab the wrong one, the error you get will be confusing. Check the category before you click.

Why you'd reach for it

At 35 credits per generation it's the cheap door into the Wan image line, and it brings two things the others mostly don't: a thinking mode and multi-reference composition. Flip thinking_mode on and the model plans the prompt internally before generating - Alibaba's spin on giving the model time to reason about composition and prompt intent, which helps on complicated one-shot prompts where you'd normally fight the output for ten reruns. For reference work, up to nine images compose into a single scene. It's not the flagship (that's Pro), but for daily text-to-image it's the sensible default.

The inputs that matter

• prompt - required. The whole text-to-image story.
• thinking_mode - BOOLEAN, off by default. Text-to-image only - it's ignored the moment you connect reference images, so don't expect it to help edits.
• reference_images - up to 9 for editing/multi-reference composition. Omit for text-to-image; set aspect_ratio to match_input_image to keep an input's framing.
• size - 2K (default) or 1K.
• num_images - 1–4, billed per image.
• seed - -1 for random, a real value to reproduce.

Output is images (IMAGE) - Save Image and done.

How it works

The standard BudgetPixel pattern: IMAGE tensors become PNGs, sent inline or via the upload endpoint, a job is created against the hosted API, and the node polls until it's done before pulling the result back into the graph. Server-side generation means your GPU never wakes up. The async catch applies everywhere in this pack: interrupt the graph and the job keeps running and billing server-side.

Install and key setup

ComfyUI Manager → search BudgetPixel → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/budgetpixel-ai/comfyui-budgetpixel.git
pip install -r comfyui-budgetpixel/requirements.txt

Only requests as a dependency - no torch, no model files. API key from a paid plan via BUDGETPIXEL_API_KEY or a budgetpixel.json in the pack folder; keys never live in workflow JSON.

Where people get burned

The two-"Wan 2.7"-in-one-pack naming is the first trap - make sure you're on the image node under BudgetPixel/Image. thinking_mode silently doing nothing on reference images is the second (it's documented as text-to-image only, but it reads like it should apply everywhere). Then the pack standards: ComfyUI caches unchanged outputs, so flip force_rerun when you want a genuinely fresh - and fresh-billed - image; and free accounts hit api_access_not_enabled because the API comes with paid plans.