ImageOnlyCheckpointSave

ImageOnlyCheckpointSave is the save-node for image-conditioned video checkpoints - the img2vid family like SVD (Stable Video Diffusion) and its cousins. It's the mirror of the "Load Checkpoint Image Only (img2vid model)" loader, and if you've ever merged or fine-tuned one of those models, it's how you make the result a shareable file.

The "ImageOnly" name is the whole story. A normal checkpoint packages a diffusion model with a text encoder (the CLIP), because normal generation is driven by prompts. An img2vid model doesn't take prompts the same way - it takes a start frame - so instead of a text CLIP it bundles a CLIP Vision encoder. That's the one meaningful difference between this node and Save Checkpoint: you hand it a CLIP_VISION where the other node expects a CLIP.

How it works

It's a subclass of CheckpointSave, which means the serialization path is the same - one .safetensors with the diffusion model, the conditioning encoder, and the VAE, plus metadata embedding the workflow. The clip_vision input just slots into the encoder slot where the text CLIP would normally go. Same filename counter behavior, same metadata, same output location, same everything else - the family DNA is intact:

ComfyUI/
└── output/
    └── checkpoints/
        └── ComfyUI_00001_.safetensors

find ~/ComfyUI/output -name "*_00001_.safetensors" | tail

The inputs that matter

• model (MODEL) - the diffusion model, e.g. the output of a merge or a fine-tuned img2vid model.
• clip_vision (CLIP_VISION) - the image encoder. This is the slot you'd otherwise feed a text CLIP into; the loader that reads the file back knows to expect it here.
• vae (VAE) - the VAE, usually matched to the model.
• filename_prefix (STRING, default checkpoints/ComfyUI) - where it saves.

It's an output node: nothing on a wire, just the file.

Where people get burned

Same output-vs-models trap as Save Checkpoint - the default lands in output/checkpoints/ while your loader looks in models/checkpoints/. The file is there; you just have to move it or change the prefix. And the same scope warning applies: this saves the model, it doesn't make it good. Merging two img2vid checkpoints and saving the result is instant, but the merge caveat from the whole family applies - blend compatible parents, test before committing a multi-GB write.

The one video-specific thing to know: img2vid checkpoints are pickier than text models about their conditioning encoder. If you save with a clip_vision that doesn't match what the model expects, the "same file" you load later can silently behave differently. Keep the pairing consistent - it's the video-model equivalent of the mismatched-VAE problem that the troubleshooting KB warns causes gray, washed output.

How you get it

Ships with ComfyUI core under model/merging (it lives in the video-model module but is categorized with the merge/save family) - no install. It arrived with ComfyUI's video support in 2024.