FLUX.2 Dev · Edit

FLUX.2 Dev is the best open-weights image editor of its generation and almost nobody runs it. The reason isn't quality - it's arithmetic. A 32B rectified-flow transformer plus a 24B Mistral vision-language text encoder is roughly 53 GB before you've loaded anything comfortable, and the community verdict in the KB's telling was blunt: technically the strongest, practically rejected.

Which makes it a perfect candidate for a node. The complaint was never that FLUX.2 Dev was bad. It was that it was too expensive to run. At 3.47 credits an image - three and a half cents - the hardware argument evaporates.

What you're actually getting

Multi-reference editing, in one checkpoint. image_urls is an autogrow socket: connect an IMAGE and another socket appears, up to three sources. You describe what to change and what to stay the same, referring to them in order, and the model composes.

This is the capability people keep reaching for and rarely find: merging two people from separate photographs into one believable scene, holding a character's face across a reference sheet, putting a product into a room that was never shot with it. Flux 2 Dev was praised for exactly this - the KB quotes someone running it locally saying it "supports character sheets unlike qwen edit and has a very high level of understanding". Klein inherited the trick at a smaller size; Flux 2 Dev still does it better.

The order matters. Connect images in the order your prompt talks about them - first source is what "the woman on the left" means.

The fields

prompt - say what to change and what must stay the same. Both halves. The failure mode on multi-reference editing is drift: you ask for a new jacket and lose the face, because you never told it the face was load-bearing.

aspect_ratio defaults to auto, which keeps the shape of the first source image. You can override with 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9 or 9:21, and then resolution - 1k or 2k - sets the size tier. auto plus 1k is the sane default; choose an explicit ratio only when the output needs a different frame than the input.

seed starts fixed, so re-queueing is one charge not two.

That's the whole node. No mask, no strength, no steps, no guidance scale, and no loras input - this is the plain Flux 2 Dev, and LoRA support lives on the separate FLUX.2 Dev LoRA node. If you came here looking for a strength dial, there isn't one, because multi-reference editing is instruction-driven rather than img2img-driven.

Outputs: images and task id.

Under the hood

Generated from black-forest-labs/flux-2-dev/edit's published input schema. The node uploads each source image to a popnow:// file URI, quotes the run before credits are held, submits under an idempotency key so retries can't double-bill, polls, then downloads and decodes the results into an IMAGE batch. The final charge prints on the node when it's done.

3.47 credits per image, flat. Nothing scales with the number of references you attach, which makes a three-image composition the same price as a one-image restyle.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Or install PopNow from ComfyUI Manager and restart. Key from popnow.ai/api/console/keys under Settings > PopNow, or POPNOW_API_KEY in the environment, then restart once more so the catalog refreshes from your account. No downloads, no GPU, no 53 GB of weights - the entire point.

Where people get burned

Reference count. One to three images, and the tooltip's "in the order your prompt refers to them" is not a suggestion. Two references plus a prompt that says "the second photo" and "the first woman" interchangeably will give you a confidently wrong composite.

Second, expectations about identity. Flux 2 Dev is the best open model at this, not a biometric face-lock. If a face has to survive a hard change of angle or lighting, you're better off planning for a reroll or two than expecting one pass.

Third, remember this is a Dev-lineage model with a content filter attached at the provider. If a run comes back content_rejected, nothing was charged - argue with your prompt, not the node.