RunPod Flux Dev

Flux Dev is the model you'd run locally if you had the VRAM and didn't mind babysitting a T5-XXL text encoder. This node is the version for the rest of us: it's a thin wrapper around RunPod's public black-forest-labs-flux-1-dev endpoint. Paste in your RunPod API key, type a prompt, and a server you'll never meet turns it into a normal ComfyUI IMAGE tensor that plugs into Save Image, an upscaler, or anything downstream like it was made on your own card. It's the "no local option" move from the API-node playbook, just aimed at a model that does have open weights.

How it works

Under the hood it's embarrassingly simple. The node POSTs your prompt and settings to https://api.runpod.ai/v2/black-forest-labs-flux-1-dev/run, grabs a job ID, then polls the status endpoint once a second until the job lands. You'll see the whole trip in your ComfyUI console:

[RunPod] Submitting job to black-forest-labs-flux-1-dev...
[RunPod] Status: IN_QUEUE (elapsed: 0s)
[RunPod] Status: IN_PROGRESS (elapsed: 15s)
[RunPod] Job completed successfully

The returned image comes back base64 and is decoded into an IMAGE tensor. No weights, no VRAM, no model files - the only thing this node runs locally is the HTTP client.

The inputs that matter

Most of the fields are Flux Dev's normal knobs, with a couple of caveats:

• api_key - your RunPod key. Every node in this pack wants it; wire one Primitive to several nodes instead of pasting it repeatedly.
• prompt - the actual generation text.
• width / height - 256 to 1536, stepped by 64, default 1024. Flux is happy up around two megapixels, so don't feel locked to 1024².
• num_inference_steps - default 28, which is the sensible Flux Dev range.
• guidance - here's the take: the default is 7.5, which is hot. Flux Dev is guidance-distilled and the community settled on ~3.5 (and lower for realism); 7.5 is where images start looking overcooked. Drop it to 3–4 first.
• negative_prompt - the node sends it, but Flux Dev's guidance distillation means it does very little. Don't lean on it.
• seed - -1 means random. Set a real number if you want reproducibility, which you will the moment you get a keeper.
• image_format - png or jpeg. JPEG is cheaper on the wire but lossy; png is the default for a reason.

What comes out

A single image output (IMAGE). Wire it to Preview Image or Save Image - that's the entire pipeline. If you're comparing against local Flux, throw a VHS_VideoCombine or your usual save stack on the other side and treat the API node like any other generator.

Installing it

One install covers all 27 nodes in this pack (the rest are Flux Kontext, Qwen, Seedream, WAN, Sora, plus your own serverless workers). Grab it in ComfyUI Manager by searching ComfyUI_RunpodNodes or RunPod, or:

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

Then restart ComfyUI. There's no requirements.txt in the repo - the only real dependencies (requests, Pillow, torch) ship with ComfyUI already, so this is one of the rare packs that works immediately after a restart. You'll need a RunPod API key from runpod.io.

Gotchas

• It "hangs" - it isn't. Public endpoints queue; 30–120 seconds is normal. Watch the console for the polling log before you give up.
• Cost - per-call billing, roughly $0.02–0.03 per megapixel per the README. A 1024² render is about one megapixel, so it's cheap for experiments and not free.
• "Job failed" - almost always a bad key or empty credits. Check both.
• Your data and your key leave the machine - that's the mechanism, not a bug. This is a small third-party pack holding a credential and phoning home by design, so skim the repo before you trust it. Prefer it for the convenience it genuinely is, not as a default for a model you could run yourself.

For a model with open weights this is a convenience purchase, not a capability unlock. But when you're out of VRAM, it's a surprisingly clean one.