FLUX-1.1-pro-TextToImage

FLUX-1.1-pro is one of those models you physically cannot download. Black Forest Labs ships it through its API and Microsoft's Azure OpenAI Images service - no weights, no GGUF, nothing to drop in your models folder. This node is the door from ComfyUI to that API: paste in an endpoint and key, type a prompt, and the image lands back in your graph as a normal IMAGE tensor, as if a local sampler had made it. It's one of four nodes in the azure-flux-comfyui pack by xpanceo-team (MIT-licensed, written by Boris Chernetsov), a deliberately minimal collection that does one thing - talk to Azure's hosted FLUX endpoints - and gets out of the way.

How it works

No model is loaded, no VAE, no sampler. Each run is a single HTTP POST to {your_endpoint}/openai/deployments/{deployment}/images/generations?api-version={api_version}, with your key sent in an Api-Key header (Azure's Images API convention, not the Authorization: Bearer you might expect). The JSON body carries your prompt, size, n, and output format. Azure runs FLUX-1.1-pro on its side, returns base64-encoded images, and the node decodes them into a batched float32 IMAGE tensor [N, H, W, C] in the usual 0–1 range. Output images wires straight into a Save Image or Preview Image node. That's the whole graph - no checkpoint loader.

The inputs that actually matter

• azure_endpoint - your Azure resource URL, like https://<your-resource>.services.ai.azure.com/. The code is lenient: it appends a missing trailing slash itself.
• api_key - your Azure OpenAI API key. It lives in plain text in the node, so keep workflows containing it out of shared exports.
• deployment - must match the deployment name you created in your Azure resource. Defaults to FLUX-1.1-pro; if you named it something else, this is where it bites.
• api_version - 2025-04-01-preview by default. Leave it unless Azure's error messages tell you otherwise.
• editing_prompt - the name is a lie; it's just your prompt. There's no editing on a text-to-image node; the field is shared across the pack.

The optional inputs are small: size (1024x1024 style - it must match a WIDTHxHEIGHT pattern or a regex rejects it before Azure ever sees it), n (1–8 images per call; each one is billed separately, so n=4 is four charges), and output_format (png, jpeg, or webp, default png).

Installing it

The install is shared across all four nodes in the pack. In ComfyUI Manager, search "azure-flux-comfyui" and install. Or by hand:

cd ComfyUI/custom_nodes/
git clone https://github.com/xpanceo-team/azure-flux-comfyui.git
cd azure-flux-comfyui
pip install -r requirements.txt

Then restart ComfyUI. The requirements are requests, Pillow, torch, and numpy - every one already ships with ComfyUI, so there are no model downloads and no GPU needed. The real setup cost is outside the node: you need an Azure OpenAI resource with a FLUX-1.1-pro deployment on it. That's an Azure portal and billing exercise, and it's the actual barrier to entry here.

Where people get burned

• 401/403 - wrong key, resource, or endpoint URL. The endpoint is your resource URL, not a model URL.
• 404/400 - the deployment name doesn't exist in your resource, or the API version doesn't support it. Confirm the deployment is actually deployed.
• Invalid size - must be exactly WIDTHxHEIGHT, e.g. 1024x1024.
• Timeouts - the node waits up to 180 seconds per call; drop n or use a smaller size.
• Trying to edit with it - 1.1-pro only does generations. If you want hosted image editing, that's the Kontext ImageToImage sibling from the same pack, not this node.

Last, the standard API-wrapper caveat: your prompt leaves the machine and goes to Microsoft, subject to its content filtering, and every run is metered. That's the right trade when the model can't run locally, which is exactly the situation with 1.1-pro. And at least this pack is small enough to audit - it's a single ~18KB Python file, so if you care where your key goes, you can read the whole thing in five minutes.