Nodes/ComfyCollectorNodes/LoRA Truncate Rank (CCN)
ComfyUI Node

LoRA Truncate Rank (CCN)

Fast rank reduction for SVD-extracted LoRAs by directly truncating weight matrices. Runs in seconds vs. minutes for full re-decomposition. NOT valid for trained LoRAs — their rank dimensions are not ordered by importance.

By valkymaera·Created 3 months ago·Updated 4 days ago· 1
LoRA Truncate Rank (CCN)
    • filepath
    lora_name
    new_rank32
    filename_prefixloras/CCN_truncated_lora
    output_dtypematch_original
    verbosetrue
    CategoryCCN/lora

    Inputs (5)

    NameTypeDefaultDescription
    lora_nameCOMBOSource LoRA file. Must be from SVD extraction, not training.
    new_rankINT321–4096Target rank. Must be less than the source LoRA's rank. Layers already at or below this rank are left unchanged.
    filename_prefixSTRINGloras/CCN_truncated_loraOutput path and filename prefix (relative to output directory).
    output_dtypeCOMBOmatch_originalData type for saved tensors. fp8_e4m3 halves size vs fp16 with minimal quality loss.
    verboseBOOLEANtruePrint per-layer rank changes to console.

    Outputs (1)

    NameTypeDescription
    filepathSTRING