Flux.2 [pro] Image

Flux 2 is the open-weights model with a hardware problem. It's a 32B rectified flow transformer plus a 24B vision-language encoder - roughly 56B parameters of "most capable image model on paper" that needs 18–24GB of VRAM even quantized. The community's verdict on launch day was basically "great, where do I put it."

The Flux.2 [pro] Image node is the answer that doesn't involve a GPU upgrade: the same BFL model family, hosted, behind one node. No weights, no VRAM, no quantization debate. Type a prompt, get an image - and because it's the API version of the family, it also edits from reference images.

What it actually is

A core partner node - ships with ComfyUI, no install, cloud-backed through Comfy's proxy to Black Forest Labs' API. Sign in with your ComfyUI account, spend credits, and the generation runs on BFL's servers. It's synchronous: submit, poll until the job's ready (the status model is explicit about moderation and failure states), and a normal IMAGE tensor comes back into your graph.

One honest caveat before the details: in current ComfyUI this node is marked legacy. The BFL family has since grown - Flux.2 [max] and friends extend it with more headroom - and the [pro] node is the older door. It still works fine; just know you're on the entry tier of a lineup that moved on.

The inputs that matter

• prompt - the generation, or the edit instruction if you attach references.
• width / height - 256 to 2048, stepped by 32, defaults 1024×768. Resolution drives the per-call price, so the size you pick is literally the price you pay.
• seed - for reproducible noise. The tooltip's honest that this seeds the noise, not the content.
• prompt_upsampling (on by default) - BFL's prompt-rewriting pass. Leave it on for open-ended creative prompts; turn it off when you have a precise, deliberate prompt and you don't want the API paraphrasing it.
• images (optional, up to 9) - reference images. This is how the API version edits: feed one or more, describe the change in the prompt, and it regenerates with your references as context. It's the hosted sibling of what Flux 2 does with reference images at 4MP output locally - except you're not paying the VRAM bill.

Output is a single IMAGE. Save it, or pipe it into local detail work.

Where it fits

The obvious use is "I want Flux 2's quality and I don't have the hardware" - which, realistically, is most people, since the open version's hardware demands are exactly why it lost the local default to Z-Image in the first place. The less obvious one: the reference-image editing, which is a big capability you get without running anything.

Gotchas

• It costs credits and the price scales with output resolution - the node's badge shows the estimate before you queue. Big outputs, multiple references: expect a visible bill.
• It's the [pro] tier. If you need more output budget, the [max] node is the upgrade path, and the legacy marking is Comfy's way of nudging you there.
• Prompt upsampling surprises people: leave it on and your carefully-worded prompt may come back changed. For precise edit instructions, turn it off.
• Up to 9 reference images, no more - the node validates and errors over that.
• Standard partner-node rules: Comfy account + credits, localhost-only login, internet required.

If you want Flux 2's output without funding a 24GB VRAM purchase, this is the cheapest on-ramp. Just don't build your whole workflow on the [pro] tier without checking whether [max] fits your budget first.