Nodes/ComfyUI-InstantTensorLoaders/Instant Load Checkpoint With Config (DEPRECATED)
ComfyUI Node

Instant Load Checkpoint With Config (DEPRECATED)

The deprecated node you can safely skip

By redstonewhite·Created 4 months ago·Updated 4 months ago· 0
Instant Load Checkpoint With Config (DEPRECATED)
    • MODEL
    • CLIP
    • VAE
    config_name
    ckpt_name

    This is the one node in the pack you should probably never use. Instant Load Checkpoint With Config mirrors ComfyUI's old CheckpointLoader - the one that asks you to pick a YAML config file alongside your checkpoint - and the author has marked it DEPRECATED right in the display name. ComfyUI itself abandoned config-picking years ago once guess-config loaders started detecting the architecture straight from the file's header. This node keeps the legacy plumbing only for compatibility with old workflows.

    Why it still exists

    The config file isn't decoration. For SD 2.x-era checkpoints it encodes v-prediction, and the config can carry a layer_idx (clip-skip) that the node re-applies after loading. If you have a workflow from the 2023 era that loads a v2 checkpoint with an explicit config, this is the faithful InstantTensor port of that flow.

    Inputs and outputs

    • config_name - the YAML config from ComfyUI/models/configs. The dropdown shows the classic v1/v2 set (v1-inference.yaml, v1-inference_fp16.yaml, v2-inference-v.yaml, the inpainting variants, and a few clip-skip configs).
    • ckpt_name - the checkpoint file.

    Outputs are MODEL, CLIP, and VAE - the same trio every checkpoint loader hands you.

    The honest take

    Unless you're running SD 2.x models from a workflow that predates guess-config, skip this and use Instant Load Checkpoint (InstantCheckpointLoaderSimple) instead. That node does the same job without you choosing a config, and for .safetensors it uses the pack's streaming path cleanly. This deprecated variant only takes the streaming path for safetensors too - but on top of that it reads the config and patches v-prediction and clip-skip, which is more moving parts for a scenario that mostly stopped mattering.

    A couple of pack-level notes so you're not surprised: this is a CUDA-only pack built on the instanttensor library, and non-safetensors files fall through to ComfyUI's normal loaders. If you load a .ckpt here you're getting the stock loader's behavior and the config handling, not much Instant magic. And because ComfyUI treats it as deprecated, it shows up flagged in the node list and modern template workflows won't reference it.

    How to install

    ComfyUI Manager → search ComfyUI-InstantTensorLoaders → install (accept the instanttensor dependency). Or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/redstonewhite/ComfyUI-InstantTensorLoaders
    

    Restart ComfyUI. It appears under advanced/loaders/instanttensor. My advice: install the pack for the other nodes, but build new workflows around the guess-config checkpoint loader, not this one.

    Categoryadvanced/loaders/instanttensor

    Inputs (2)

    NameTypeDefaultDescription
    config_nameCOMBO11 options: anything_v3.yaml, v1-inference.yaml, v1-inference_clip_skip_2.yaml, v1-inference_clip_skip_2_fp16.yaml, v1-inference_fp16.yaml, v1-inpainting-inference.yaml, +5
    ckpt_nameCOMBO0 options:

    Outputs (3)

    NameTypeDescription
    MODELMODEL
    CLIPCLIP
    VAEVAE