Nodes/ComfyUI-BrainDead/BD Lotus-2 Model Loader
ComfyUI Node

BD Lotus-2 Model Loader

Load Lotus-2 (FLUX.1-dev + LoRAs + LCM bridge) for depth or normal estimation. Diffusion-based, SOTA-quality monocular geometry prediction. FLUX.1-dev is loaded from HF cache (~24 GB). Lotus-2 weights (~1.5 GB depth or ~2.9 GB normal core+sharpener) auto-download to ComfyUI models/lotus2/{task}/ if missing. Reuse the loaded model across multiple Predict calls — first load is slow, subsequent same-task calls hit the in-memory cache.

By BizaNator·Created 7 months ago·Updated 16 days ago· 14
BD Lotus-2 Model Loader
    • model
    taskdepth
    dtypebf16
    deviceauto
    flux_repo_idblack-forest-labs/FLUX.1-dev
    lotus2_weights_dir
    cpu_offloadfalse
    Category🧠BrainDead/Depth

    Inputs (6)

    NameTypeDefaultDescription
    taskCOMBOdepthWhich Lotus-2 LoRA pair to load. depth = ~1.5 GB, normal = ~2.9 GB (different LoRA rank).
    dtypeCOMBObf16Inference precision. bf16 recommended on Blackwell/H100.
    deviceCOMBOauto3 options: auto, cuda, cpu
    flux_repo_idoptSTRINGblack-forest-labs/FLUX.1-devHF repo id for the FLUX.1-dev base. Loaded via HF cache (HF_HOME env). Override only if you have a fork.
    lotus2_weights_diroptSTRINGOverride directory containing the three Lotus-2 weight files. Default: <ComfyUI models dir>/lotus2/{task}/
    cpu_offloadoptBOOLEANfalseEnable diffusers model CPU offload. Each pipeline module (transformer / VAE / text encoders) lives on CPU and moves to GPU only when needed. Peak VRAM drops from ~35 GB to ~27 GB (largest single module + activations) at the cost of ~1.5x slower inference per call. Turn on when running Lotus-2 alongside other large models (Qwen-Image-Edit, FLUX gen, etc).

    Outputs (1)

    NameTypeDescription
    modelLOTUS2_MODEL