Comflowy Flux Dev Lora

If you've ever found a great Flux LoRA on Civitai and then stared at a 24GB download and a VRAM meter going red, this node is aimed squarely at you. Comflowy Flux Dev Lora generates images with the hosted Flux Dev model and loads a LoRA you point it at - but the LoRA lives server-side. You don't download the LoRA, you don't merge it, you don't hold it in VRAM. You give the node a download address, and Comflowy's service pulls it, applies it, and generates.

That's the trick that makes this node interesting: lora_path is just a URL. The README's framing is "fill in the download address of the Flux LoRA model into the node to load and use it." For Civitai, grab the direct download link for the safetensors file. This turns "test this LoRA style" into a five-second workflow instead of a disk-and-RAM negotiation.

The meter again: about 350 credits per image per the README. Plus, you're paying for the Flux Dev generation itself on top of the convenience. Fine for taste-testing LoRAs; expensive as a production pipeline.

How it works and the inputs

Standard Comflowy API pattern: prompt + settings POSTed to app.comflowy.com/api/open/v0/flowy, result downloaded back as an IMAGE. The node-specific part is that lora_path (a multiline string, required) is sent up to the service, which fetches the LoRA and applies it at lora_scale before sampling.

The settings that matter:

• prompt (required, multiline) - normal Flux prompting.
• lora_path (required) - direct URL to a Flux LoRA safetensors file.
• lora_scale - 0 to 4, default 1.0. Standard LoRA weight; 0.7–0.9 is a common sweet spot for style LoRAs, 1.0+ for strong character/identity ones.
• num_inference_steps - default 28, range 1–50. Flux Dev is happy around 20–30; more steps don't help much past that.
• guidance_scale - default 3.5, range 1–20. Lower is looser/more creative; Flux Dev generally likes low-ish guidance.
• image_size - the usual presets plus custom (with height/width 256–2048 used only then).
• output_format - jpeg or png. JPEG if you're iterating fast, PNG if the output is final.
• seed, num_images (1–4), safety_tolerance (1–6).

Output: a single IMAGE.

Installing and the failure modes

Install Comflowy's Custom Nodes via Manager (search "Comflowy") or git clone https://github.com/6174/comflowy-nodes into custom_nodes, restart, set your key with Comflowy Set API Key. Only requests as a dependency; there's genuinely nothing to download locally - that's the node's whole personality.

The gotchas are all about that lora_path field:

• The URL must be a direct file link. A Civitai model page URL won't work; you need the actual .safetensors download link (the ?type=model&format=SafeTensor one).
• Wrong-base-model LoRAs silently underperform. Flux LoRA ecosystems are fragmenting across Flux dev / schnell / new versions. If your LoRA was trained on a different Flux variant, the style just won't land and it'll look like a weak prompt, not a crash.
• Server-side fetch can fail on big files. LoRAs over a couple GB may time out. If the generation errors after a long wait, that's usually the fetch, not your settings - check the console.
• Network + key issues as with every node in this pack; the README flags connectivity as the top cause of failures.

If you already run Flux Dev locally with a full LoRA library, you don't need this node - your disk is faster and your credits stay in your pocket. Where it earns its keep is the machine without VRAM headroom, or the "I want to try this one LoRA someone just dropped" moment without committing to a download.