FLUX.1 [dev]

FLUX.1 [dev] is the model that ended the SDXL era and made "just type a sentence" prompting the norm - 12 billion parameters, natural-language prompts, hands that finally work, and in-image text that reads correctly. The catch has always been the hardware: a full FLUX stack needs serious VRAM, and the community spent 2024 squeezing it onto consumer cards with fp8, NF4, and GGUF. This node is the cheat code: it runs FLUX.1 [dev] on Runware's machines, so you get the quality without the 24GB tax, your GPU stays free for other work, and there's nothing to download.

One honest caveat before you commit: FLUX.1 [dev] is non-commercial (the Apache-2.0 tier is schnell, and it's the hard-to-finetune one). If you're doing paid work, that matters. For everything else, dev is the one people actually reach for.

What you set

Required: positivePrompt, width, and height - all straightforward, dims from 128 to 2048 in steps of 64. FLUX doesn't want tiny or weird resolutions; 1024ish is its happy place.

The interesting part is guidance. FLUX is a rectified-flow model, so the classic CFG math works differently. The node gives you two dials:

• CFGScale - off by default, and off is usually right. FLUX uses distilled guidance; cranking this up often just burns detail.
• trueCFGScale - the real classifier-free guidance knob, for when you genuinely need stronger prompt adherence. Higher helps you obey, but pushes quality down. Start at default and only touch it when the image ignores you.

Everything else is standard for the pack: negativePrompt (ignored when CFGScale ≤ 1, so don't be confused when it seems to do nothing), seed, steps (FLUX defaults are few - this is a distilled model, more steps isn't automatically better), a 76-choice scheduler dropdown you can ignore, and numberResults for up to 20 variations.

The genuinely FLUX-specific goodies: layerDiffuse, which switches on transparent-image (RGBA) generation - one of the few places you get real cutouts without post-processing - plus sockets for pulid (face consistency), controlNet, ipAdapters, and lora, all fed by the pack's builder nodes.

The single output is image (IMAGE), which wires straight into Preview Image or Save Image.

How it works

Every node in this pack is the same shape: it builds a request with the model AIR and taskType, sends it over REST through the Runware SDK, and downloads the result back as a native IMAGE tensor. The heavy lifting - loading the 12B model, the T5 encoder, the denoising - all happens in the cloud. Your node is basically a very well-dressed API call, with the model's real parameters as widgets instead of JSON. Each run prints its cost on the title bar, which for FLUX is usually fractions of a cent.

Installing

ComfyUI Manager: search Runware, install, restart. Or:

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

No model files - dependencies are just runware-sdk, pillow, and soundfile. You do need a Runware API key: Settings → Runware API key, or RUNWARE_API_KEY, or runware auth login.

Troubleshooting

The missing-key error is the classic first stumble: set the key before you blame the model. If outputs look soft, remember FLUX is distilled - check that you haven't cranked CFGScale into the double digits. And if you're here for transparent PNGs, layerDiffuse only works when it's actually toggled on.