Load Checkpoint Image Only (img2vid model)

Here's a loader that breaks your mental model on purpose. Load Checkpoint Image Only takes a video checkpoint and hands you three wires - but one of them isn't a text encoder. You get MODEL, CLIP_VISION, and VAE. No CLIP. For image-to-video models, that's exactly right: the starting image, not a prompt, is the primary driver of what gets generated, so the checkpoint ships a vision encoder instead of a text one. The name "Image Only" is the honest description - the model conditions on an image, and the loader only extracts what that needs.

How it works

One input, ckpt_name, from ComfyUI/models/checkpoints. The loader reads the checkpoint and pulls out the diffusion model, the CLIP vision encoder, and the VAE - the pieces an image-conditioned video workflow uses. It's the img2vid sibling of Load Checkpoint: same file source, same split-it-apart idea, different component set.

The wiring that makes it do something is the SVD_img2vid_Conditioning node. It takes the CLIP_VISION, your starting image, the VAE, and a stack of video settings - width, height, video frame count, motion bucket ID, fps, augmentation level - and produces the positive conditioning, negative conditioning, and latent that your sampler then denoises into a video clip. You won't find a CLIP Text Encode in this graph, because there's no text encoder to feed it.

Why it exists

This is the Stable Video Diffusion pattern, and it generalizes to the whole image-conditioned-video family. When a checkpoint is built to animate a given frame rather than imagine a scene from words, the conditioning path runs the image through the vision encoder, and a text encoder would be dead weight - hence the loader simply not loading one. If you're coming from the text-to-image world, the mental shift is: the image is your prompt. Which is also why the practical advice for img2vid is to choose your starting frame carefully - the model's "prompt adherence" is really "frame adherence."

Gotchas

• No text conditioning, no prompt. If you wire this into a normal CLIP Text Encode workflow expecting to prompt your way out of a bad start frame, you'll get a connection error. The graph around img2vid looks different on purpose.
• Frame count and resolution are the quality dials, not strength. The conditioning node's motion bucket and fps settings shape the output's motion character. Keep resolution near the model's native range - pushing past it is how you get warping.
• Not for text-to-video. A t2v checkpoint wants a normal loader with a text encoder. Loading it through this node would strip out the very component it needs.
• Same checkpoint folder, different intent. It reads from models/checkpoints, so it's easy to accidentally point it at a model that isn't image-conditioned. If the load succeeds but the downstream conditioning node complains, that's usually it.

It ships with ComfyUI core - no install. It's the loader you reach for when the video starts from a picture, and the mental model change - image is the prompt - is the whole game.