Muse Image 1.0 (BudgetPixel)

Most image models draw what your prompt says and hope for the best. Meta's Muse Image 1.0 is different: before it renders, it can do things - ground your prompt with web and image search, and run code to lay out charts, tables and text. So a request like "a bar chart comparing the top five AI models by parameter count, with accurate numbers" isn't a coin flip where the model hallucinates a plausible-looking graph; it's a request the model can actually research and render correctly. That's a genuinely different job than text-to-image, and it's why Muse is the node in this pack for infographics, data visualization and typography-heavy work.

It's a hosted model, like everything in the BudgetPixel pack - your prompt and reference images go to the API, the agentic pipeline (search, code, render) runs server-side, and the ~2K image comes back. Flat 12 credits per image, no matter what the agent did, which is a good deal for the search-and-code stuff. There's no local path for Muse anyway; it's a closed model with no weights to download.

The inputs

• prompt (required) - describe the image, or the data task, or both. Muse's search grounding is where the "factual accuracy" claim lives.
• reference_images - up to four input images for editing or multi-reference composition. Omit for pure text-to-image.
• quality - high or low. High (default) refines over multiple passes and uses search grounding fully; low responds faster. Both bill the same 12 credits, so the choice is about speed, not cost.
• aspect_ratio - ten options including match_input_image; an edit follows the input's frame unless you force one.
• num_images - 1–4.
• output_format - webp (default), png, or jpeg.

Output is a single images socket carrying core IMAGE tensors - wire it into Preview Image or Save Image like any local generator.

Where it shines and where it doesn't

Reach for Muse when the text matters: infographics, signage, memes with legible words, product mockups with correct labels. That's the classic weak spot of most diffusion models, and the agentic preamble is a real fix. What it isn't is a fast iteration tool - low quality helps, but the search-and-code pipeline is inherently slower than a plain text-to-image call, and you're paying per image rather than per GPU-second. For rapid ideation, the pack's P-Image node at 10 credits is snappier. For "make me a correct chart," Muse is the one.

Install and issues

Standard BudgetPixel install - ComfyUI Manager (search "BudgetPixel") or:

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

Restart and find it under Add Node → BudgetPixel → Image. You need a BudgetPixel API key (paid plans include API access) via BUDGETPIXEL_API_KEY or a budgetpixel.json file - no key widget, by design.

The shared pack gotchas apply: a recent ComfyUI for the core types, and interrupting a run only abandons the local wait - the server-side job keeps running and still bills its 12 credits. Also worth knowing: whatever filters Meta's model has are baked in. There's no local checkpoint to patch, so anything the hosted Muse refuses stays refused. If your project needs output the vendor won't allow, this isn't the node for it.