Nodes/ComfyUI_LC123_nodes/LC Model Optimizer (pipe) ⚡
ComfyUI Node

LC Model Optimizer (pipe) ⚡

This node analyzes your machine and loads the model, text encoder and VAE that suit YOUR card, with the speed-ups that actually work on it (Sage, Comfy Kitchen, etc.). Run the report with the button below, or from the LC123 settings. Missing a file? Pick a ⬇ entry and it downloads on the first run. Finetunes work too. Pick your own file and it still gets the speed-ups, plus a ⚠️ if the text encoder or VAE doesn't match. Any other model: set the base model to Custom and pick your own files (and the text encoder type). The speed-ups still apply; recommendations, estimates and checks show as unsupported. Current available models: Krea 2, Krea 2 (Raw), Qwen-Image 2.1, Z-Image Turbo, Flux.2 Klein 9B, Ideogram 4, Anima, SDXL, Illustrious, Pony, Custom.

By lonecatone23·Created 2 months ago·Updated about 5 hours ago· 26
LC Model Optimizer (pipe) ⚡
    • pipe
    • model
    • clip
    • vae
    • summary
    ◄base_modelAnima (2B)►
    ◄goalOptimal►
    ◄model_file★ Recommended►
    ◄text_encoder★ Recommended►
    ◄vae★ Recommended►
    ◄speed_upsAuto►
    ◄attentionauto►
    ◄fp16_accumulationauto►
    ◄step_cacheauto►
    ◄torch_compileoff►
    ◄qwen21_cacheauto►
    ◄clip_typeauto►
    CategoryLC123/optimizer

    Inputs (12)

    NameTypeDefaultDescription
    base_modelCOMBOAnima (2B)The model family. Picks the recommended files and the checks.
    goalCOMBOOptimalQuality = closest to bf16. Optimal = fast and near identical. Fast = fastest that still looks right.
    model_fileCOMBO★ Recommended59 options: ★ Recommended, From checkpoint, None, ⬇ Download: anima-base-v1.0.safetensors, ⬇ Download: anima-base-v1.0-fp8.safetensors, ⬇ Download: anima-base-v1.0-int8convrot.safetensors, +53
    text_encoderCOMBO★ Recommended20 options: ★ Recommended, From checkpoint, None, ⬇ Download: qwen_3_06b_base.safetensors, ⬇ Download: qwen_3_8b.safetensors, ⬇ Download: qwen_3_8b_fp8mixed.safetensors, +14
    vaeCOMBO★ Recommended7 options: ★ Recommended, From checkpoint, None, ⬇ Download: qwen_image_vae.safetensors, ⬇ Download: flux2-vae.safetensors, ⬇ Download: qwen_image_2.1_vae_bf16.safetensors, +1
    speed_upsCOMBOAutoAuto = from the System & Model Optimization Report. Manual = choose each one below.
    attentionCOMBOautoAttention backend. auto = the fastest one that passed the report's test.
    fp16_accumulationCOMBOautoFaster fp16 math with a small quality cost. auto = Fast goal only.
    step_cacheCOMBOautoEasyCache: skips steps that barely change. auto = Fast goal, 20+ steps.
    torch_compileCOMBOoffCompiles the model: the first run is slow, later runs are faster. Needs Triton.
    qwen21_cacheCOMBOautoQwen-Image 2.1 only (Qwen Image 2.1 Cache): where the KV cache lives and how it is stored. auto = lossless, int8 on the Fast goal or cards under 16 GB. int8 halves the cache at about bf16 accuracy, int4 quarters it but roughly doubles the per-step error. RAM costs little speed. off = slowest.
    clip_typeCOMBOautoCustom only: how the text encoder is loaded (CLIPLoader's type). auto = from the encoder's shape when it matches one known base model. Ignored for the other base models and for a checkpoint's own encoder.

    Outputs (5)

    NameTypeDescription
    pipeLC_PIPE—
    modelMODEL—
    clipCLIP—
    vaeVAE—
    summarySTRING—