Load CSM Checkpoint

Load CSM Checkpoint is the least impressive node in this pack and also one of the most annoying to skip. It's a dropdown that points at your CSM-1B weights file and hands the TTS nodes an absolute path. That's it. No model loading happens here - despite the name, the heavy lifting is done later, lazily, inside CSM Text-to-Speech.

Why does a path need a node? Because the TTS nodes in this pack take model_path and tokenizer_path as plain strings. You could type C:\ComfyUI\models\sesame\ckpt.pt into a text box, but that's fragile, ugly, and breaks the moment you move your install. This loader keeps everything inside ComfyUI's model-folder convention: it lists files from ComfyUI/models/sesame/ and resolves the full path for you. Think of it as the pack's equivalent of a checkpoint loader in an image workflow - it just doesn't actually load anything into VRAM.

How it works

At workflow-build time it scans ComfyUI/models/sesame/ and gives you a dropdown of checkpoint files (it falls back to ckpt.pt if the folder is empty). When the graph runs, it resolves the selected file to its full path and outputs it as a string. The real model load - reading ckpt.pt, building the Llama backbone with torchtune, pulling the Mimi codec - happens on first call in the TTS node, and it's cached there.

The one input and one output

• checkpoint - dropdown of files in ComfyUI/models/sesame/. If you only have the one ckpt.pt, it's already selected.
• model_path - STRING, wired into the model_path input of either TTS node.

It pairs with Load CSM Tokenizer (same pack, lists tokenizer directories from ComfyUI/models/sesame_tokenizer/ and outputs the tokenizer_path). The TTS nodes need both wires.

Setup

Same install as everything else: clone thezveroboy/ComfyUI-CSM-Nodes into custom_nodes and pip install -r requirements.txt, or find "ComfyUI-CSM-Nodes" in ComfyUI Manager, then restart. The model file itself:

• Download ckpt.pt from sesame/csm-1b - it's the raw PyTorch checkpoint, a bit over a gigabyte.
• Create ComfyUI/models/sesame/ and drop it there. The node expects the file directly in that folder, not in a subdirectory.

Common issues

• File not showing in the dropdown. Wrong folder, or it's nested in a subfolder. It must sit directly in ComfyUI/models/sesame/, and you may need to reload the workflow after adding it.
• Russian error about the checkpoint. If you see something like Чекпоинт ... не найден, the path resolution failed - usually a renamed file or a missing folder.
• Renamed weights. The default is ckpt.pt; if you renamed it, pick it from the dropdown rather than relying on the fallback.