Luma Image to Image

Luma is mostly famous for video - Dream Machine, Ray, that whole family - but its image line, Photon, is quietly solid, and the Luma Image to Image node is the simplest possible way to use it: an image, a prompt, and one dial that decides how much of the original survives.

It's a stripped-down img2img node, and that's its appeal. No mask, no denoise stack, no ControlNet plumbing. You set a weight, you write what you want, you get a modified image back. When you want to restyle a photo without rebuilding a graph around it, that's the whole job.

What it actually is

A core partner node: ships with ComfyUI, no install, cloud-backed through Comfy's proxy to Luma's API. Sign in with your ComfyUI account, spend credits per call, and Photon runs on Luma's servers. Synchronous submit-and-poll - your queue waits, then a finished IMAGE tensor drops into your graph.

The inputs that matter

• image - the source you're modifying.
• prompt - what changes. "Make it golden hour," "turn the cat into a lion," "1980s VHS still" - img2img territory.
• image_weight (0–0.98, default 0.1) - the one dial that matters, and the tooltip explains it backwards from intuition: the closer to 1.0, the less the image gets modified. A weight near 0.1 (the default) is a loose edit - Photon takes the structure and runs with your prompt. Crank it toward 0.9 and you're doing a subtle nudge, keeping the photo's identity. This is your denoise control by another name.
• model - two options: photon-1 (the full flagship, better quality, pricier) and photon-flash-1 (the fast/cheap tier for iterating).
• seed - with the usual cloud-model asterisk: the tooltip is honest that results are nondeterministic regardless of seed; the seed mainly controls whether the node re-runs.

Output is a single IMAGE, ready for Save or further processing.

Where it fits

Think of it as the "one-button restyle" rung of a workflow. The niche where this beats local img2img is turnaround: you don't have a good img2img checkpoint for the style you want, or you don't want to download one for a single shot. Point this at a photo, describe the new look, and you're done - including styles no local model you own covers.

It's also a fine first pass before heavier work: modify in the cloud, then take the result into your local graph for detail passes.

Gotchas

• image_weight is the whole game, and people get burned treating the default as a "strength" slider where high = strong. It's inverted: high = less change. Start at the default 0.1, and only push toward 0.8+ when you specifically want the source preserved.
• Photon-flash is the iterating tier, photon-1 the commit tier - they cost differently, and the price badge reflects it.
• Nondeterminism: don't chase an exact duplicate. If a run lands close, lock the seed and rerun for another take rather than expecting pixel-identical repeats.
• Standard partner-node rules: Comfy account with credits, localhost-only login, needs internet.

If you've ever wanted "just img2img it, without the graph," this is the node. One dial, one prompt, done.