Nodes/ComfyUI/Load Checkpoint With Config (DEPRECATED)
ComfyUI Node Runs on cloud

Load Checkpoint With Config (DEPRECATED)

The old checkpoint loader that needed a config file — and why it died

By Comfy-Org·Created 4 years ago·Updated about 3 hours ago· 129,870
Load Checkpoint With Config (DEPRECATED)
    • MODEL
    • CLIP
    • VAE
    config_name
    ckpt_name

    If you've loaded a model in ComfyUI more than twice, you've used the descendant of this node without knowing it existed. Load Checkpoint With Config is the original checkpoint loader - the one that required you to also pick a YAML config file describing the model's architecture - and it's now marked DEPRECATED in the source. It still works. You just shouldn't be the one who finds out the hard way why it was replaced.

    What it does

    A checkpoint is the whole network in one multi-gigabyte file: model, CLIP, and VAE bundled together. The "With Config" part is the historical catch. Early Stable Diffusion checkpoints didn't carry enough metadata for ComfyUI to know what architecture they were, so you had to tell it by selecting a config file - those YAML files in models/configs/ like v1-inference.yaml or v2-inference.yaml. Get the config wrong and the load goes sideways.

    The node's dropdown literally shows the 11 classic configs, one for each flavor of SD1/SD2 - base, fp16, clip-skip-2, inpainting variants, and the v2 line. It's a fossil record of the 2022–2023 era when that manual step was necessary.

    Inputs and outputs

    Two required inputs:

    • config_name - the architecture YAML, selected from models/configs/
    • ckpt_name - the checkpoint file from models/checkpoints/

    Three outputs, which should look familiar because every modern loader still produces them:

    • MODEL - the denoising model, wired to the sampler
    • CLIP - the text encoder, wired to CLIP Text Encode
    • VAE - wired to VAE Decode to turn latents into pixels

    Why it's deprecated

    Load Checkpoint (Simple) - the CheckpointLoaderSimple you already know - does the same job with one input instead of two. It calls load_checkpoint_guess_config, which inspects the checkpoint's weight keys and infers the architecture automatically. No config file, no matching dance, no "why does my model look like static" moments. The two-node dance this node requires is precisely the failure mode guess_config eliminated, and the source flags this class with DEPRECATED = True.

    Should you ever use it?

    Honestly, almost never. The only real argument for it is a legacy checkpoint whose architecture the guesser can't identify - an edge case that shrinks every year. If you're following an ancient workflow that names this node, swap in Load Checkpoint (Simple) and see if it loads cleanly; it usually will. If the workflow truly needs the manual config, it'll still be here, quietly deprecated and unbothered. That's the whole personality of this node: an ancestor, still running, politely waiting for you to upgrade.

    Categorymodel/loaders

    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