JoyCaptionAlpha1Online(NYJY)

When JoyCaption Alpha One landed, the training community's verdict was quick and unusual for this space: it was the best open captioning model for training data - fast, and able to run in as little as 8.5 GB VRAM. The catch was the setup. This node removes the setup entirely: it calls the official joy-caption Alpha One space over the web, so you get that captioning quality with zero model downloads and nearly zero VRAM. Feed it an image, pick a caption style, and get a caption back. It's the same trick the pack uses elsewhere - rent the GPU by the request - pointed at one of the genuinely best captioners in the community.

How it works

Under the hood the node speaks the Gradio protocol to the fancyfeast-joy-caption-alpha-one.hf.space endpoint: it uploads your image to the Space, joins its processing queue, and polls until the caption is done, then returns the text. No local models, no weights, no inference on your machine. Because it's a shared public Space, it's free - and subject to that Space's daily quotas and busy periods.

The inputs that matter

• image - the IMAGE to caption.
• caption_type - the big one: descriptive (default; rich natural-language description), training_prompt (formatted like a training caption, great for feeding back into dataset building), or rng-tags (random tag-style output).
• caption_tone - formal or informal. "Formal" is the clean, caption-standard tone; "informal" gives you a more chatty, casual read.
• caption_length - from any (default) through very short/short/medium-length/long/very long, plus exact word counts like 20, 30, 40, up to the hundreds. Set 30 and you get roughly a 30-word caption - the single most useful control for tuning caption density for training.

Output: captions, a STRING.

Installing

Standard pack install: ComfyUI Manager → Install via Git URL → https://github.com/aidenli/ComfyUI_NYJY, restart. The fake_useragent and requests deps come with the pack. Nothing to download, nothing to configure.

Common issues & honest caveats

It's a shared public Space. Free means finite: when the Space hits its quota or a queue backs up, you get slow responses or nothing. The author's docs warn the online JoyCaption options have daily caps (on the order of a couple dozen requests). If you need captioning at scale, the local JoyCaption node in this same pack is the upgrade path - same captioning family, no quotas, at the cost of ~10 GB VRAM and a big download.

Your image leaves your machine. It's uploaded to Hugging Face's Space; keep that in mind for anything sensitive. For the job this node is built for - quickly captioning a few images, sampling caption styles for a dataset, or just trying the captioner before committing to local - it's hard to beat. And when you find a style you like, you can hand the same settings to a local node and scale up without losing the recipe.