Flux.2 [max] Image

Flux.2 [max] is BFL's highest-tier API model for the Flux.2 generation, and Flux.2 [max] Image is the dedicated ComfyUI node for it - the same backend the newer combined Flux.2 Image node reaches through its model dropdown, but with the settings laid out flat in front of you instead of hidden behind a DynamicCombo. It landed in December 2025, a month after the [pro] node, when BFL split the tiers into separate nodes before the 2026 consolidation folded them back together.

Standard partner-node preamble, and I'll keep it short because you've heard it if you've touched any BFL node: no local model, no VRAM, no models/ download. Your prompt and images go to Comfy's proxy at api.comfy.org, BFL's servers run the model, and credits come off your Comfy account. No bring-your-own-key. The price badge on the node is the current per-megapixel cost and it updates as BFL reprices, so read it at run time rather than trusting any number you saw in a blog post.

The reason you'd pick this node over the combined Flux.2 Image is explicitness. width and height are right there (256–2048, step 32), which matters because Flux.2's headline feature is large output - it'll happily render 4MP, and the credits scale with megapixels so you want to see the number you're committing to. The optional images input takes up to 9 reference images for native multi-reference editing: drop in a subject, a pose, a lighting reference, and prompt the combination. prompt_upsampling defaults to true here (it's the opposite default from the older Flux.1 nodes), so if you've written a careful prompt, flip it off - it rewrites yours for "more creativity," and on a model this capable you usually don't want that.

seed is best-effort determinism, same caveat as the rest of the BFL family: fixed seeds won't give you bit-identical images, they just tell the API "treat this as a rerun." Don't build a workflow that depends on exact reproducibility.

Output is one IMAGE tensor, already decoded - save it or feed it into an upscaler or composite downstream; there's no VAE step because there's no local decode. One thing to remember: the node validates at the source that you can't pass more than the supported reference count, so keep your reference stack under the limit.

Where it fits: [max] is the tier you reach for when [pro] isn't quite good enough and you're making a hero asset - product shots, campaign art, anything where the render is the deliverable. It's also the strongest multi-reference editor in the BFL family, so if your workflow is "three references + a sentence," this node does it in one call. The bill is real, but you can see exactly what you're spending before you commit, which is more than most APIs give you.