FLUX 2 Dev (BudgetPixel)

FLUX 2 was the moment Black Forest Labs jumped a generation: a 32B rectified-flow transformer with a Mistral-3 vision-language encoder, capable of both generation and multi-reference editing in one checkpoint. The community consensus is that it's the most capable open-weights image model around - and also that running it locally is a hardware slog that wants 18–24 GB of VRAM even quantized. This node is the back door: FLUX 2 Dev quality, hosted, with no weight download and no GPU requirement, at 25 credits per generation. That's cheaper than the pack's own FLUX 1.1 Pro, which is the tell for how far the API pricing has come.

The headline feature is multi-reference composition. Where FLUX 1.1 Pro takes one optional image, FLUX 2 Dev accepts up to four reference images for editing and composition - feed it a person, a garment, a scene, a style reference, and a prompt that ties them together, and the vision-language encoder does the spatial reasoning that CLIP-era models just couldn't. This is the node's reason to exist.

How it works

Same machinery as every BudgetPixel node: prompt plus up to four images get packaged into a POST /images/flux-2-dev call, the node polls /images/{id} until succeeded, and the result downloads back as an IMAGE batch. The multi-reference bit is where the pack does something clever: image batches you wire in are mapped onto the reference slots automatically, so a batch of four images becomes four references without you juggling anything. Small images go inline as data URIs; anything over the size cap is pushed through POST /v1/uploads.

The inputs that matter

• prompt - required, multiline. For multi-reference work, describe the relationship: "the person in image 1 wearing the coat from image 2, standing in the scene from image 3."
• reference_images - optional, up to 4. Wire IMAGE connections here (or a batch) and the node becomes an editing/composition engine instead of text-to-image.
• aspect_ratio - 7 options (1:1, 4:3, 3:4, 9:16, 16:9, 2:3, 3:2).
• num_images - 1–4, default 1. Multiplies credits per call.
• seed - default -1 (random).

Output is images as an IMAGE batch.

Install

Manager (search "BudgetPixel") or:

cd ComfyUI/custom_nodes
git clone https://github.com/budgetpixel-ai/comfyui-budgetpixel.git
pip install -r comfyui-budgetpixel/requirements.txt

Restart, then set BUDGETPIXEL_API_KEY or a budgetpixel.json ({"api_key": "bpx_live_..."}) in the pack folder. No model files, requests is the only dependency.

Where people get burned

Two things to keep in mind. First, this is a hosted FLUX 2 Dev - the API build runs BFL's moderation, and the open-weights version's censorship reputation is well documented (the modidex FLUX 2 panel goes on about the "lobotomized" model-card complaints). If you need the uncensored capability, the local build is the only door, and it's a hardware door. Second, multi-reference quality depends on the references, not just the prompt: conflicting or low-res reference images produce mush, and the same 25 credits get you nothing useful. Feed references that share a consistent look and let the VLM do its job. And remember force_rerun when you re-test - the cache won't spend your credits for you.