Luma Text to Image

Luma Text to Image is the original Luma image node - the one that first put Luma's Photon models inside a ComfyUI graph as a cloud call. It's been quietly superseded by the newer UNI-1 nodes (Luma Image Node 2 in this same family), so think of this as the "classic" Luma generation path: simpler model list, but with a genuinely nice set of reference inputs that the newer node dropped. Partner node, ships with core, no install, billed per call against your Comfy account.

It's the one to reach for when your job is just text-to-image and you want the reference controls that matter: style, character, and up to four generic image references. Luma's image models (Photon 1 and the cheaper, faster Photon Flash) are solid generalists, and the style-image input in particular is a cheap way to get a consistent look without training anything.

How it works

Prompt and model go to Luma's generations/image endpoint through Comfy's proxy; the node polls the job and downloads the result as a regular IMAGE tensor. Reference handling is the differentiator:

• image_luma_ref - a LUMA_REF connection from the Luma Reference node, for up to four generic reference images. This is the "use these as visual references" channel.
• style_image - exactly one style reference image, whose look gets blended in.
• character_image - one or more images of a character, up to four, to keep an identity consistent across runs.

The inputs that matter

• prompt - the actual text prompt (min 3 chars, so "cat" works but "hi" doesn't).
• model - photon-1 or photon-flash-1. Flash is the budget fast one; Photon-1 is the quality pick. The price badge changes with the model.
• aspect_ratio - default 16:9, plus square, portrait, and ultrawide options.
• style_image_weight - 0 to 1, default 1. How hard the style image's look is applied. Only matters if a style image is connected, which the tooltip says plainly.
• seed - the family standard: controls re-running, not the result.

Output is IMAGE, ready to save or chain.

Where people get burned

• Reaching for this when you want UNI-1. If you're on a current ComfyUI, the UNI-1 nodes are the newer Luma model family. This node is still functional and still fine for Photon jobs, but it's the previous generation.
• The min-prompt gotcha. A one- or two-character prompt errors out. Give it a real sentence.
• Style weight when no style image is attached. The weight does nothing until you connect one - it's not a global "creativity" slider.
• Reference discipline. Style image takes only one image; character takes up to four; luma_ref up to four. Shove a batch into the wrong input and you'll get surprising results rather than an error.
• Cost. Photon-1 is roughly four times the price of Photon Flash per image. If you're iterating, start on Flash and save Photon-1 for the finals.