Bernini Load Output Image

Bernini Run Single GPU doesn't output an image - it outputs a path to one, because under the hood it's launching a separate Python process that writes a file. For video that's fine, even preferable; video helper nodes want paths. But for text-to-image and image-to-image, you want the pixels inside the graph so Preview Image, Save Image, and everything downstream can use them. Bernini Load Output Image does exactly that conversion: path string in, IMAGE tensor out.

The one input

output_path - a string that comes straight from the runner's output_path output. The node accepts .png, .jpg, .jpeg, and .webp. It rejects anything else with a clear message, and it raises a proper "file does not exist" error if you queue it before the runner has finished - so make sure the runner runs first in the chain, not in parallel with this.

What comes out

image (IMAGE) - a standard ComfyUI image tensor, ready for Preview, Save, or further processing. Mechanically it's the same load path ComfyUI's own Load Image uses: PIL opens the file, EXIF-transpose, convert to RGB, normalize to 0–1, add the batch dimension. Nothing exotic, which is the point - you get pixels you can wire anywhere without adapting.

Where people get tripped up

The classic mistake is feeding the runner's .mp4 output into this node. The error is explicit - "only supports image files" - but the intended fix is to not use this node at all for video. Leave the path as a string and let a video-specific helper consume it. Bernini has no native audio anyway, so your video pipeline is usually routing to its own encoder or to an audio stage; a plain path is what those want.

Also worth knowing before you expect a preview: this node only turns files into tensors. If you want the runner to push a preview into the UI, that's the runner's own show_preview toggle, not this node's job.

Installing this node

One of the 16 nodes in the lzits Nodes pack (Lior Zitser, MIT). Install via ComfyUI Manager (search "lzits Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/lzitser23/lzits-nodes.git

Restart ComfyUI. No extra dependencies beyond PIL/numpy/torch, which ComfyUI already ships. The heavy lifting - cloning bytedance/Bernini and downloading ByteDance/Bernini-R-Diffusers - is upstream, and the Bernini Setup Commands node prints those commands.