Checkpoint Save

This is the node at the end of a model-merging workflow - the one that writes the result to disk. You've merged two checkpoints, or baked a LoRA permanently into a model, or otherwise assembled a MODEL + CLIP + VAE you're happy with, and now you want a single reusable checkpoint file instead of rebuilding the graph every time. Wire those three into this node and it saves a .safetensors you can load like any other checkpoint.

It's the same function core ComfyUI's checkpoint-save provides; it lives in the art-venture pack's Model Merging drawer alongside the rest of its merge tooling.

How it works

A ComfyUI checkpoint is really three things bundled together: the diffusion model, the CLIP text encoder, and the VAE. This node takes those three as separate inputs and packs them back into one file. That's why it wants all three wired - a checkpoint missing its VAE or CLIP would be a broken checkpoint. It's an output node with no outputs of its own: it does its work by writing a file, so it sits at the terminal end of the graph, nothing hangs off it.

The inputs that matter

• model (MODEL), clip (CLIP), vae (VAE) - the three components of the checkpoint. Feed the merged/modified versions from the tail of your workflow.
• filename_prefix (STRING, default checkpoints/ComfyUI) - where and what to name it. The prefix can include a subfolder path, so checkpoints/my-merges/dreamblend-v1 drops it in a tidy folder. ComfyUI appends a counter and the extension.
• dtype (optional, float16 or float32) - the precision the weights are saved at. float16 is the sane default: half the file size, and for image models the quality difference is negligible. Pick float32 only if you have a specific reason to want full precision and don't mind a file twice the size.

Installing it

ComfyUI Manager → search comfyui-art-venture → install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

then restart ComfyUI. It's pre-installed on comfy.icu, though note that saved files live on the executor's disk - for cloud runs you'll want to route the output somewhere you can retrieve it.

Common issues

The pack-wide one: a red "missing" node means art-venture isn't installed. Manager → "Install Missing Custom Nodes."

Practical gotchas: checkpoints are big. A full SDXL save is several gigs, a Flux-class model much more, so watch your disk. All three inputs are required - if your merge chain dropped the VAE somewhere, grab one with a VAELoader before saving or the file won't be a usable standalone checkpoint. And a workflow reminder: this saves the model as it stands after your graph. If you baked a LoRA in at strength 0.8 upstream, that's what's frozen into the file - the LoRA is no longer separable. That permanence is the point (you get a fast, single-file checkpoint), but it's also why you keep your source models and merge recipe around in case you want to tweak the blend later.