Modelverse Flux Dev

Flux Dev is the open-weights model the community treated as the local-diffusion renaissance - right up until everyone remembered it wants a GPU with opinions. This node is Flux Dev through UCloud's Modelverse API, and it's the most "normal" node in the pack: it behaves like a familiar local text-to-image / img2img node, right down to width/height/seed/steps/CFG, except the sampling happens on somebody else's hardware. If you've ever downloaded Flux and quietly uninstalled it after watching your VRAM choke, this is the version you can actually use.

What it is

A full-featured Flux Dev wrapper. Text-to-image by default, and the moment you wire an image into the optional image input it becomes image-to-image - the strength input (default 0.8) then controls how much the source survives. For people coming from local ComfyUI, this node is the bridge: same mental model, zero model management.

The inputs that matter

• client - required wire from the Modelverse Client / Secret Client node.
• prompt - the description. Flux is good at following it; write real sentences.
• width / height - 512–1536, default 1024². Keep within the model's sweet spot; extreme aspect ratios beyond ~2:1 can get soft.
• strength - 0.01–1.0, default 0.8. Only matters in img2img mode (when image is connected).
• num_inference_steps - 28 default, up to 50. 28 is a sensible floor; more steps cost more, not necessarily better.
• guidance_scale - 3.5 default, 0–10. Flux likes lower CFG than SD models - 3.5 is already on the gentle side.
• seed - -1 for random; set it to reproduce.
• num_images (1–4 per request) and num_requests (1–10) - batch knobs, and the bill multipliers.

Output is a single image (IMAGE tensor) - preview or save it with any standard node, no download step needed.

Installing and keying up

Standard for this pack. ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ucloud/comfyui-plugin

restart ComfyUI. Dependencies are light (requests, pillow, numpy, openai, GitPython) - and that's the appeal: the 24GB of Flux weights you'd otherwise manage never enter the picture. The only setup is a ModelVerse API key from the UCloud console, entered via the Modelverse Secret Client node (key in secrets.json, name only in the workflow) or a config.ini MODELVERSE_API_KEY=... line.

Common issues

• Output is nothing like local Flux - it's the same model family but a hosted build with its own defaults. Don't transplant local LoRA expectations; treat it as its own flavor.
• img2img ignores the image - strength at 0.8 is transform-heavy. For subtle edits you want it lower; if the image input isn't wired, you're in t2i mode regardless.
• "API key is not set" - empty client node. The eternal first check.
• Soft results at extreme ratios - Flux was trained mostly square-ish; pushing 512×1536 is asking a lot.
• Costs - every request and image is metered. num_requests × num_images is a real multiplier; iterate cheap, then scale.

It's the node to reach for when you want that Flux "just works" reliability without the hardware negotiation. No .safetensors pilgrimage required.