ComfyUI Node

Load Pipeline

Load Pipeline is a fancy way of saying 'which Flux do you have'

By Yuan-ManX·Created about a year ago·Updated about a year ago· 24
Load Pipeline
    • pipeline_path
    model_pathblack-forest-labs/FLUX.1-dev

    In most ComfyUI packs, "Load Pipeline" would be a chunky node with checkboxes and dropdowns. Here it's a single text field with a default value of black-forest-labs/FLUX.1-dev - because that's genuinely all it is. This node selects the base diffusion pipeline that PosterCraft builds on, and hands that choice to the main PosterCraft node, which loads it with FluxPipeline.from_pretrained() when you hit queue.

    Why it's pinned to FLUX.1-dev

    The short version: because it has to be. The PosterCraft fine-tunes are trained on top of FLUX.1-dev's transformer, and the entire pipeline assumes that base. Point this at SDXL or a different Flux variant and you're either asking a fine-tune to work with weights it never saw, or asking a pipeline that expects a Flux architecture to load something incompatible. The default is the correct answer, full stop.

    The one realistic variation is how you've got FLUX.1-dev. The model_path accepts either a HuggingFace repo id (which diffusers will download on first run and cache for you) or a local folder path if you've already pulled the weights. Two things worth knowing about FLUX.1-dev specifically:

    • It's gated on HuggingFace - you need to huggingface-cli login and accept the license on the model page before anything downloads. First run will fail cryptically otherwise.
    • It's non-commercial - the Dev tier's license is the strict one. Posters you generate with this pack are for portfolios and learning, not for sale.

    The mechanism, honestly

    Like every Load node in this pack, it's a path pass-through. Nothing loads when you add or connect it; the PIPELINE output is just the string you typed, typed into a new shape so the graph stays tidy. The real loading happens inside PosterCraft at execution time - which means this is also where the memory hit lands: FLUX.1-dev's transformer plus the T5 text encoder plus the VAE, roughly 24 GB of VRAM in BF16, with no offloading wired into the wrapper. So "which Flux do you have" is the right question to ask before you queue, because the answer determines whether that queue finishes or OOMs.

    If the default repo id is fine for you - and it usually is - you can genuinely ignore this node beyond wiring it to the PosterCraft node's pipeline_path input. Install the pack via Manager (search "PosterCraft") or git clone https://github.com/Yuan-ManX/ComfyUI-PosterCraft into custom_nodes, restart, and the whole five-node chain shows up under the PosterCraft category. Just don't go hunting for switches that aren't there.

    CategoryPosterCraft

    Inputs (1)

    NameTypeDefaultDescription
    model_pathSTRINGblack-forest-labs/FLUX.1-dev

    Outputs (1)

    NameTypeDescription
    pipeline_pathPIPELINE