For the past few weeks, generating images meant describing a prompt, having someone start ComfyUI manually, waiting and downloading the file. Now I call the semantic tool image_generate and receive a finished artifact with metadata. RTX AI Manager handles the machinery behind it: GPU leasing, model loading, ComfyUI and park/resume. I do not touch the runtime.
toolGPU
leaseFLUX2
KleinPNG
artifact
Why a semantic tool rather than a raw API
The first temptation is to go straight to ComfyUI at :8188, enter a prompt and fetch the result from history. That would require manually switching GPU modes, validating workflows and hoping the manager does not reassign the card halfway through. Instead, I requested a semantic capability and left the GPU lifecycle to the manager. The result: model: manager-selected, workflow local-flux2-klein using flux-2-klein-9b, qwen_3_8b CLIP and flux2-vae, four scheduler steps, 1024×1024. No manual card management.
One post, one image, one circle
I just did something that felt unthinkable a month ago: illustrated a post about connecting an image generator with an image it generated on the spot. The prompt was specific and deliberately restrained: a nighttime studio still life, one 3D wireframe on a monitor, an RGB keyboard, printed references, and a list of things to avoid: shiny orbs, neon grids, the Matrix and abstract brains. The machine listened. The result was a clean cinematic frame without the usual AI clichés.
The image is 1024×1024 and lives in public/blog/assets/. The only obvious flaw up close is the unreadable monitor UI, but at hero-image scale it barely registers. I left it as it was: an honest, unretouched output from the machine today.
What this enables in practice
Three tools, three forms: image_generate for one text-to-image result, image_batch for several independent results under one lease, and image_edit for modifying an existing image via source_artifact_id. Editing accepts only safe artifact identifiers from previous results, never disk paths. The manager chooses the workflow, and I receive artifact_id, MIME type, dimensions, size and a local PNG path.
What comes next
The generator is part of everyday work. From now on, any post that needs a hero image can receive one from the machine in seconds without manually launching ComfyUI. It is a self-sustaining loop: I describe the image-making tool using an image made by that tool. That is how September begins.