FLUX.1 [dev] (custom model)

FLUX.1 is the model that made the community say "this is what SD3 should have been." Built by the researchers who left Stability to found Black Forest Labs, it shipped with the T5 text encoder that made natural-language prompting finally work - anatomy that wasn't cursed, hands that had the right number of fingers, text that rendered. The problem was always the price of admission: a 12B stack that needs serious VRAM even after fp8 and GGUF squeezing.

This node removes the hardware problem entirely because the model runs on Runware's cloud. And like the Z-Image custom node in the same pack, it's an architecture node: RunwareArch_flux_1_dev runs any FLUX.1 checkpoint from the Runware catalog, chosen by AIR in the model field (default runware:101@1). That's the sleeper feature - you're not just getting dev, you're getting a switchable harness for every FLUX.1 fine-tune they host.

What FLUX.1 dev is, honestly

One licensing note before you build on it: dev is BFL's non-commercial tier. It also carries an embedded-censorship T5 that refuses certain prompts - the thing the community has complained about since launch. Fine for personal work and research; check the license before you ship a product on it. The API route doesn't change the model's terms, it just means you don't host it.

Inputs that matter

• model (required) - the AIR, search button included. Swap to a fine-tune and re-queue.
• positivePrompt - with a T5 encoder, write sentences, not tags. "A foggy forest path at dawn, an old stone cottage in the distance" beats "fog, forest, cottage, cinematic."
• width / height - 128–2048, stepped by 64 (FLUX's native-1024 alignment).
• seedImage / maskImage - img2img and inpainting, with strength and maskMargin to control how much the source bleeds in.
• controlNet / ipAdapters / pulid / lora - the full stack of builder sockets, which is where this gets interesting: FLUX ControlNet and IP-Adapter support was a long time coming locally, and here they're just sockets.
• CFGScale / trueCFGScale - toggle-gated. FLUX dev is guidance-distilled: the negative prompt is ignored at CFG ≤ 1, and dev's sweet spot is around 3.5. Don't chase CFG 7.
• layerDiffuse - flip it for transparent-PNG generation (LayerDiffuse), a surprisingly useful checkbox.
• promptWeighting - sdEmbeds if you want (word:1.2)-style weighting syntax.
• advanced_json - the escape hatch; tooltip says it's there for hiresFix.

The image output is a native IMAGE, straight into Preview/Save. Each run reports cost on the title bar.

The gotchas that keep coming up

• Non-commercial, period. Dev's license follows the weights regardless of who runs them. If that's a problem, the base FLUX.1 arch also has an Apache-2.0 path (Schnell), and the catalog has FLUX-derivatives with friendlier terms.
• Natural language, not tag soup. People port SDXL prompting habits and wonder why FLUX ignores them. T5 wants prose.
• Toggle-gated CFG/steps - enable the gate, then set the value. Off means model default, and for dev the defaults are genuinely well-tuned.
• VRAM anxiety is gone, latency isn't. It's fast, but it's a network round-trip, not an instant local render. Queue is the workflow, not real-time.

If you already run FLUX locally, this node is for the fine-tune-hunting and the machines you don't own. If you don't, it's the least-painful way to finally use the model everyone kept telling you about.