fal.ai Vision

Somewhere in every serious image pipeline there's a "what is this?" step - describe a reference frame, check an outfit, catalogue a scene. StudioDeep's answer is the fal.ai Vision node, which takes an image from your graph and sends it to a hosted vision model for a text answer. It's the paid-cloud sibling of the pack's LLM-based nodes: instead of loading a vision model locally, you rent one for the duration of the call.

How it works

The node grabs the first image in the batch (image[0]), converts the ComfyUI tensor to a PNG data URL, and POSTs it to fal's openrouter/router/vision endpoint with your API key in an Authorization: Key ... header. A 120-second timeout keeps runaway requests from hanging your queue forever. On the way back you get two things: the model's response text, and a usage string - formatted JSON with the token counts, which is genuinely handy if you're watching your fal bill.

The inputs that matter

• fal_backend - the FAL_BACKEND output from fal.ai Backend. This is the key slot; without it the node can't authenticate.
• image - any IMAGE. Only the first frame is sent, so feed a single image, not a video batch.
• model - six hosted models to pick from: google/gemini-2.5-flash, anthropic/claude-sonnet-4.6 and 4.5, openai/gpt-4o, qwen/qwen3-vl-235b-a22b-instruct, and x-ai/grok-4-fast. Default is Gemini Flash, which is the usual sweet spot of fast/cheap for description work.
• prompt - what to ask. Defaults to "Describe this image in detail." Change it and you get directed analysis instead.

The optional ones: system_prompt (context before the question), reasoning (toggle - real chain-of-thought, slower and pricier), temperature (0–2, default 1.0), and max_tokens where 0 = no limit, per the author's own tooltip.

Outputs

response (STRING) is the text answer; usage (STRING) is JSON you can preview in a text node or parse for token accounting. Wire response into a String Passthrough to eyeball it, or straight into a prompt field.

Installing

ComfyUI Manager → search "StudioDeep", or:

cd ComfyUI/custom_nodes
git clone https://github.com/studiodeep-ai/comfyui-studiodeep StudioDeep
pip install -r StudioDeep/requirements.txt

Restart, then grab a fal.ai key from fal.ai/dashboard and drop it into fal.ai Backend. The pack pulls in requests as part of requirements.txt, so you're covered.

Troubleshooting

• HTTP 401 / 403 - your key is wrong or has no balance. Check the dashboard; fal is pay-as-you-go, and an empty wallet kills the call with an HTTP error that includes the body.
• Slow responses - you're hitting a 120s timeout ceiling on a huge image or a slow model (Grok-4-fast tends to think). If you're consistently timing out, shrink the image upstream or switch to Gemini Flash.
• "Describe this image" answers you don't want - that's the prompt field, not a bug. If you need structured output, be explicit: "Return JSON with shot, lighting, and subject."

One honest caveat: this is the node you reach for when you want to pay for vision. If you're running a local model happily, this adds latency and cents for no gain.