Magiiic.org

Project Category: ComfyUI

  • Oli Prompt Tools

    A collection of practical ComfyUI utility nodes designed to reduce workflow complexity. Each node solves a specific recurring problem — avoiding VRAM crashes, picking prompts deterministically, filtering incompatible LoRAs, or labelling model pipelines. No external dependencies beyond the standard ComfyUI environment.

    Nodes

    Node Category Purpose
    Mega Lora Loader Oli/loaders Stackable multi-LoRA loader with automatic compatibility filtering
    Prompt Line Pick Oli/prompt Seed-driven line picker — uniform distribution, fully independent instances
    Video Frame Limit Oli/utils Caps video duration to avoid VRAM out-of-memory crashes
    Model Info Oli/utils Returns architecture details of any connected model
    Node Label Oli/utils Reads and passes through the upstream node’s title

    Installation

    Via ComfyUI Manager (recommended): search for Oli Prompt Tools and click Install.

    Manually:

    cd ComfyUI/custom_nodes
git clone https://github.com/magicoli/oli-prompt-tools

    Restart ComfyUI. No additional dependencies required.

    Mega Lora Loader (Oli)

    Stackable multi-LoRA loader inspired by rgthree’s Power Lora Loader. Key additions: stackable, and each LoRA’s safetensors header is checked against the connected model’s key map before loading — incompatible LoRAs are silently skipped rather than producing errors or corrupted output.

    • Add as many LoRA rows as needed with the ➕ Add LoRA button.
    • Each row has an on/off toggle, a LoRA selector, and a strength slider.
    • Incompatible LoRAs are highlighted in the UI (orange = incompatible, grey = disabled).
    • Connect the enable input to a boolean condition to bypass all LoRAs at once — useful with switch or router nodes.
    • Two usage modes: connect model/clip to apply LoRAs immediately, or leave them unconnected to only build a LORA_STACK without applying it. In stack-only mode the node acts as a LoRA definition block — chain several together, then feed the final stack to a downstream loader that has model/clip connected. This lets you define LoRAs once and apply them to multiple models at different points in the workflow.

    Inputs

    Input Type Default Description
    lora_stack LORA_STACK — Incoming stack from another loader (optional)
    model MODEL — Model to apply LoRAs to (optional)
    clip CLIP — CLIP encoder to apply LoRAs to (optional)
    (lora rows) — — Added dynamically via the ➕ button
    enable BOOLEAN true When false, passes model/clip/lora_stack through unchanged

    Outputs

    Output Type Description
    lora_stack LORA_STACK Full stack including upstream entries
    MODEL MODEL Model with all compatible LoRAs applied
    CLIP CLIP CLIP with all compatible LoRAs applied

    Prompt Line Pick (Oli)

    Seed-driven reimplementation of the easy promptLine concept. Replaces start_index + max_rows with a single seed input. The selected line is determined by sha256(seed:node_id) % len(lines).

    • Uniform distribution — every line has equal probability regardless of list length.
    • Full independence — multiple instances in the same workflow each pick independently, even with the same seed, because the node ID acts as a channel discriminator.
    • No prime modulo management — the SHA-256 hash eliminates correlation between lists of similar length.
    • Preserves the COMBO output for use with COMBO-typed inputs.

    Inputs

    Input Type Default Description
    prompt STRING — One item per line
    seed INT 0 Workflow seed — share with KSampler for reproducible pairs
    remove_empty_lines BOOLEAN true Strip blank lines before picking

    Outputs: STRING, COMBO

    Video Frame Limit (Oli)

    Caps video generation duration to avoid VRAM out-of-memory crashes. The frame budget is derived from transformer peak memory first principles rather than empirical constants:

    bytes_per_latent_frame = TENSOR_COPIES × (width÷8) × (height÷8) × hidden_dim × 2
max_frames             = total_vram × safety_margin ÷ bytes_per_latent_frame

    Where TENSOR_COPIES = 5 (Q, K, V, attention output, residual activations) and hidden_dim is auto-detected from the connected model. Uses total VRAM rather than free VRAM — ComfyUI offloads weights layer-by-layer, so peak activation memory scales with total VRAM, not the remainder after model loading.

    The node displays detected VRAM, model name, hidden dim, requested and capped frames directly on the canvas after each execution — making it usable as a standalone config panel for the whole generation.

    Inputs

    Input Type Default Description
    width INT 832 Generation width in pixels
    height INT 480 Generation height in pixels
    fps FLOAT 16 Frames per second
    duration FLOAT 10 Requested duration in seconds
    safety_margin FLOAT 0.95 Fraction of total VRAM to budget (0.95 = 5% headroom)
    model MODEL — Optional — enables hidden_dim auto-detection

    Outputs

    Output Type Description
    width INT Pass-through
    height INT Pass-through
    frames INT Capped frame count
    fps FLOAT Pass-through
    duration FLOAT Actual duration after capping

    Model Info (Oli)

    Returns architecture details of any connected model (MODEL, CLIP, VAE, or any other type). Also reads the upstream node’s title. Useful for debugging model pipelines, inspecting what arrived on a bus, or driving conditional logic based on model class.

    Inputs

    Input Type Description
    model * Any model type (optional)

    Outputs

    Output Type Description
    class_name STRING Python class name of the model (e.g. WAN21, Flux)
    dim INT Hidden dimension auto-detected from the model
    label STRING Title of the upstream node that produced the model

    Node Label (Oli)

    Passes any value through and outputs the title of the upstream node at a configurable traversal depth. Useful for labelling outputs, routing between branches, or building self-documenting workflows where node titles carry semantic meaning.

    Inputs

    Input Type Default Description
    depth INT 1 Hops to travel upstream (1 = direct parent, 2 = grandparent…)
    node * — Any value — passed through unchanged (optional)

    Outputs

    Output Type Description
    node * Pass-through of the input value
    label STRING Title of the node depth hops upstream

    Examples

    The examples/ directory contains minimal starter workflows for each node. Open them in ComfyUI and wire them into your own workflow.

    File Node demonstrated
    examples/mega_lora_loader.json Mega Lora Loader
    examples/prompt_line_pick.json Prompt Line Pick
    examples/video_frame_limit.json Video Frame Limit
    examples/model_info.json Model Info
    examples/node_label.json Node Label

    License

    GNU Affero General Public License v3.0