Flux.2 Image

Flux.2 is the big 2025 model whose open weights are a 32B monster most cards can't run comfortably, and whose API tiers - [pro] and [max] - are where the sane version of the model lives. Flux.2 Image is the ComfyUI node that calls those tiers, and it's the one that landed in May 2026 as the DynamicCombo-upgraded version of the family: one node, a model dropdown, and the settings that matter appear or vanish depending on which tier you pick.

Let's nail the shared partner-node ground once, since every BFL node in core works identically. Nothing runs locally: no checkpoint in models/, no VRAM, no sampler. Your prompt (and any reference images) go to Comfy's proxy at api.comfy.org, get forwarded to Black Forest Labs, and the credits come off your Comfy account balance. No bring-your-own-key, ever - that's a standing community complaint with a GitHub discussion you can vote on. The node's price badge reflects which tier you've selected, and it changes with BFL's pricing, so treat it as current-today.

The interesting part is the model input, which is a DynamicCombo. Pick Flux.2 [pro] or Flux.2 [max] and the node reconfigures itself: both expose width and height (256–2048, step 32) plus an images slot for up to 8 reference images. That reference support is the big deal - Flux.2 does image-to-image and multi-reference editing natively, so you can drop a few shots in and prompt "put this lighting on that subject" without any separate edit node. Pro and Max differ in quality and price (Max is the premium tier), not in interface.

Output is a single IMAGE tensor, already decoded pixels - no VAE decode, straight into Save Image or any downstream node. One grounded limit: the node refuses more than 8 reference images at the Python level, so don't try to sneak a 9th in via a batch.

Practical notes. The default resolution of 1024×768 is fine for most things, but Flux.2's calling card is 4MP output - crank width/height toward the top of the range and let the API do the heavy lifting if your use case is print or high-detail work; the credits scale with output megapixels, and the badge shows you that. And while seed is there, treat it as "make the node rerun" rather than "reproduce this exact image" - BFL's determinism is best-effort, not contractual. If you're comparing renders for a specific look, keep everything else fixed and accept that tiny variations are part of the API life.

For anyone who wants Flux.2 quality but can't stomach the 32B open weights, this is genuinely the frictionless path: it's the same model family, on someone else's GPU, at a per-image price you can see before you click.