📂 加载本地GIF

GIFs are a surprisingly common source of motion reference - people animate stuff with them, and image-to-video models love a frame sequence to start from. ComfyUI's built-in loaders handle static images and videos, but an animated GIF lands awkwardly. This node, "📂 加载本地GIF", cracks the GIF open frame by frame and hands you the whole thing as an IMAGE batch, ready for a video model, an img2img pass, or per-frame processing.

How it works

Pillow does the heavy lifting. The node opens the file with Image.open(path), then iterates im.n_frames, seeking to each frame and converting it to RGBA. Every frame becomes a float32 tensor in [0,1], and they all get stacked into a single (N, H, W, C) batch - frame count first, then height, width, channels. Output is images (IMAGE).

Inputs and output

One input, one output:

• path (STRING) - where the GIF lives. Typed path, not a picker, resolved relative to where ComfyUI runs. Absolute paths avoid the classic "can't find it" dance.
• Output: images (IMAGE) - all frames in one batch.

The catch: RGBA

Here's the gotcha the README won't tell you. The code converts every frame to RGBA - four channels - because GIFs support transparency. That means the output batch is 4-channel, and a lot of ComfyUI nodes (VAE encode, most image ops) expect 3-channel RGB. You'll see channel-mismatch errors downstream and wonder what's wrong.

Two fixes, both inside this pack:

• Wire the output through CreateEmptyImageBatch, whose hidden talent is stripping alpha to plain RGB.
• Or just don't feed it into nodes that care - some processors tolerate RGBA.

Either way, know that "loads fine, errors downstream" with this node usually means channels.

Installing it

Standard pack install - ComfyUI Manager → search comfyUI_LLM, or:

cd ComfyUI/custom_nodes
git clone https://github.com/XieJunchen/comfyUI_LLM
# restart ComfyUI

Pillow and torch, both already in the pack's requirements.txt.

Other things worth knowing

• Frame count is your video length. A 30-frame GIF at 30 fps is one second of motion - check N before you feed it to a video model that expects a specific frame count.
• File paths, not directory paths. One GIF per node. For a folder full of them, you'd loop the node or use a batch loader instead.
• It's the local counterpart to LoadImgFromUrl - that one fetches a single remote image; this one loads a local animated sequence. Different jobs, same pack.

If a GIF is your source of truth for motion, this is the fastest way to turn it into something ComfyUI can actually work with. Just mind the alpha channel.