ComfyUI_WolfBypass

Bypass LoRA loader with block chopping and per-block strength for Flux/Chroma and similar models.

Security

Scriptable nodes execute arbitrary Python. Only run scripts and workflows from trusted sources. See Scriptable Block Merge for details.

---

Table of Contents

• Installation
• Node map
• Quick reference
• Block filter nodes
• Block-strength nodes
• Scriptable block merge
• Img/Txt separate weights
• Block strength inspection
• Block filter syntax
• Per-block strength
• Documentation
• Changelog
• Requirements

---

Installation

Place this folder in your ComfyUI custom_nodes directory (or clone/link). Restart ComfyUI.

All node names end with (🐺) in the UI.

---

Node map

Nodes are grouped by family: filter (text pattern), block-strength (widget or script), img/txt split (Flux-style), and inspection.

flowchart TB
  subgraph filter [Block Filter]
    A1[Load LoRA Bypass Block Filter]
    A2[Load LoRA Bypass Block Filter Model Only]
  end
  subgraph bypass_blocks [Block Strength Bypass]
    B1[Load LoRA Bypass Per-Block Strength]
    B2[Load LoRA Bypass Per-Block Strength Model Only]
  end
  subgraph merge_blocks [Block Strength Merge]
    M1[Load LoRA Merge Per-Block Strength]
    M2[Load LoRA Merge Per-Block Strength Model Only]
  end
  subgraph scriptable [Scriptable]
    S1[Load LoRA Merge Per-Block Scriptable]
    S2[Load LoRA Img/Txt Separate Scriptable]
  end
  subgraph imgtxt [Img/Txt Separate]
    I1[Load LoRA Bypass Img/Txt Separate]
    I2[Load LoRA Merge Img/Txt Separate]
  end
  subgraph inspect [Inspection]
    D1[Block Strength Inspector]
    D2[Block Strength Plot]
  end
  MODEL[MODEL] --> A1
  MODEL --> B1
  MODEL --> M1
  MODEL --> S1
  MODEL --> I1
  A1 --> MODEL
  B1 --> block_strengths_debug
  M1 --> block_strengths_debug
  S1 --> block_strengths_debug
  I1 --> block_strengths_debug
  block_strengths_debug --> D1
  block_strengths_debug --> D2

---

Quick reference

Class	Display name	File	Inputs	Outputs
LoraLoaderBypassWolf	Load LoRA (Bypass, Block Filter) (🐺)	nodes.py	model, clip, lora_name, strength_model, strength_clip; optional: block_filter, block_filter_mode, compile_forward	MODEL, CLIP
LoraLoaderBypassWolfModelOnly	Load LoRA (Bypass, Block Filter, Model Only) (🐺)	nodes.py	model, lora_name, strength_model; optional: block_filter, block_filter_mode, compile_forward	MODEL
LoraLoaderBypassWolfBlocks	Load LoRA (Bypass, Per-Block Strength) (🐺)	nodes_blocks.py	model, clip, lora_name, strength_model, strength_clip, block_strengths; optional: compile_forward	MODEL, CLIP, block_strengths_debug
LoraLoaderBypassWolfBlocksModelOnly	Load LoRA (Bypass, Per-Block Strength, Model Only) (🐺)	nodes_blocks.py	model, lora_name, strength_model, block_strengths; optional: compile_forward	MODEL, block_strengths_debug
LoraLoaderBypassWolfBlocksDoubleSplit	Load LoRA (Bypass, Img/Txt Separate Weights) (🐺)	nodes_blocks.py	same as Bypass Per-Block	MODEL, CLIP, block_strengths_debug
LoraLoaderBypassWolfBlocksModelOnlyDoubleSplit	Load LoRA (Bypass, Img/Txt Separate, Model Only) (🐺)	nodes_blocks.py	same as above, model-only	MODEL, block_strengths_debug
LoraLoaderWolfBlocks	Load LoRA (Merge, Per-Block Strength) (🐺)	nodes_blocks_merge.py	model, clip, lora_name, strength_model, strength_clip, block_strengths	MODEL, CLIP, block_strengths_debug
LoraLoaderWolfBlocksModelOnly	Load LoRA (Merge, Per-Block Strength, Model Only) (🐺)	nodes_blocks_merge.py	model, lora_name, strength_model, block_strengths	MODEL, block_strengths_debug
LoraLoaderWolfBlocksDoubleSplit	Load LoRA (Merge, Img/Txt Separate Weights) (🐺)	nodes_blocks_merge.py	same as Merge Per-Block	MODEL, CLIP, block_strengths_debug
LoraLoaderWolfBlocksModelOnlyDoubleSplit	Load LoRA (Merge, Img/Txt Separate, Model Only) (🐺)	nodes_blocks_merge.py	same, model-only	MODEL, block_strengths_debug
LoraLoaderWolfBlocksScriptable	Load LoRA (Merge, Per-Block, Scriptable) (🐺)	nodes_blocks_scriptable.py	model, lora_name, strength_model; optional: block_strengths, script	MODEL, block_strengths_debug
LoraLoaderWolfBlocksScriptableDoubleSplit	Load LoRA (Img/Txt Separate, Scriptable) (🐺)	nodes_blocks_scriptable.py	same + script for img/txt blocks	MODEL, block_strengths_debug
BlockStrengthInspector	Block Strength Inspector (🐺)	nodes_block_inspect.py	block_strengths_debug (required)	block_strengths_txt (STRING)
BlockStrengthPlot	Block Strength Plot (🐺)	nodes_block_inspect.py	block_strengths_debug (required); optional: width, height	image (IMAGE)

---

Block filter nodes

• Load LoRA (Bypass, Block Filter) (🐺)
  Applies LoRA in bypass mode with optional block filter: a single text pattern (e.g. double_blocks.0,double_blocks.1 or double_blocks.*) and include/exclude mode.

• Load LoRA (Bypass, Block Filter, Model Only) (🐺)
  Same as above, model-only (no CLIP).

flowchart LR
  MODEL[MODEL] --> Bypass[Bypass + block_filter]
  CLIP[CLIP] --> Bypass
  Bypass --> MODEL
  Bypass --> CLIP

See Block filter syntax below for block_filter and block_filter_mode.

---

Block-strength nodes (Bypass / Merge)

• Load LoRA (Bypass, Per-Block Strength) (🐺) and Model Only
  Bypass LoRA with per-block strength. The node scans the LoRA and shows a grid (D0, D1, …, S0, S1, …). Drag cells to set strength; stored as JSON.

• Load LoRA (Merge, Per-Block Strength) (🐺) and Model Only
  Same widget and block order, but applies LoRA via standard merge (weights modified as in core Load LoRA). Per-block strengths scale the merge; CLIP uses a single strength.

When you switch LoRA, the block list updates automatically. Order: double_blocks then single_blocks (see Block strength format and API).

---

Scriptable block merge (model only)

• Load LoRA (Merge, Per-Block, Scriptable) (🐺)
  Model-only. You provide a Python script that receives blocks, block_strengths, strength_model, lora_name, and must set output_strengths. Only the script’s output is applied; block_strengths input only seeds the script (and is used as fallback if the script fails).

• Security: This node runs user Python. Do not load untrusted workflows or scripts.

Full contract, injected variables, and examples: Scriptable Block Merge. Strength profiles and copy-paste recipes: Custom Merge Recipes.

---

Img/Txt separate weights (Flux-family)

For Flux-style models, each double block has an image branch (img_attn, img_mlp) and text branch (txt_attn, txt_mlp). The “Img/Txt Separate” nodes expose two strengths per double block: one for the image branch, one for the text branch. The widget shows D0i, D0t, D1i, D1t, … then S0, S1, … (i = img, t = txt).

• Load LoRA (Bypass, Img/Txt Separate Weights) (🐺) and Model Only
  Bypass with separate img/txt strengths per double block.

• Load LoRA (Merge, Img/Txt Separate Weights) (🐺) and Model Only
  Standard merge with separate img/txt strengths.

• Load LoRA (Img/Txt Separate, Scriptable) (🐺)
  Script receives blocks like double_blocks.0.img, double_blocks.0.txt, … and sets output_strengths. Can also set output_merge_function to change how the scaled delta is applied (see SCRIPTABLE_BLOCK_MERGE.md).

Use these when you want different LoRA strength on image vs text branches (e.g. stronger on img, weaker on txt).

---

Block strength inspection (debug / viz)

• Block Strength Inspector (🐺)
  Prints block strength values (index, block name, value). Requires block_strengths_debug from any Load LoRA block node. Output: STRING.

• Block Strength Plot (🐺)
  Plots block strengths as a bar chart (dark theme, PIL). Requires block_strengths_debug; optional width / height. Output: IMAGE.

flowchart LR
  LoadNode[Load LoRA block node] --> debug[block_strengths_debug]
  debug --> Inspector[Block Strength Inspector]
  debug --> Plot[Block Strength Plot]
  Inspector --> STRING[STRING]
  Plot --> IMAGE[IMAGE]

Connect block_strengths_debug from any Bypass/Merge/Scriptable block node to Inspector or Plot.

---

Block filter (original nodes)

• block_filter: Comma-separated names or patterns, e.g. double_blocks.0, single_blocks.0-5, double_blocks.*.
• block_filter_mode: include = only these blocks get the LoRA; exclude = all blocks except these.

---

Per-block strength (Blocks nodes)

• block_strengths: JSON array of numbers, one per block (e.g. [1, 0.5, 1, ...]). Filled by the widget or wired from another node.
• Widget: D / S with indices; on Img/Txt nodes: D0i, D0t, … Drag a cell up/down (0–2).

---

Configuration

Edit config.yaml in this folder (or set env vars) to change behavior:

• disable_scriptable_nodes (default: false): If true, scriptable block merge nodes (those that run user Python) are not registered. Use in locked-down environments to disable remote code execution.
• Env override: COMFYUI_WOLFBYPASS_DISABLE_SCRIPTABLE_NODES=1 forces scriptable nodes off.

flowchart LR
  subgraph sources [Config sources]
    YAML[config.yaml]
    ENV[COMFYUI_WOLFBYPASS_DISABLE_SCRIPTABLE_NODES]
  end
  subgraph init [__init__.py]
    load[_load_config]
    reg[Register nodes]
  end
  YAML --> load
  ENV --> load
  load --> reg
  reg -->|disable_scriptable_nodes: true| no_script[Scriptable nodes not registered]
  reg -->|false| yes_script[Scriptable nodes registered]

---

Build and debug

• Minified scripts: By default the UI loads minified JS. To regenerate: npm install then npm run minify (writes web/scripts/*.min.js_).
• Debug (unminified): In the browser console set localStorage.setItem("ComfyUI_WolfBypass_debug_scripts", "1") then reload the page to load unminified scripts.
• CodeMirror bundle: After editing build/codemirror_entry.mjs, run npm run build:codemirror from this extension; it writes web/scripts/codemirror_bundle.js_. Script sources for the loader live in web/scripts/*.js_; run npm run minify after editing them.
• Editor resize grip: Disabled by default (autoresize mode). Opt in only when needed: localStorage.setItem("ComfyUI_WolfBypass_enable_editor_resize_grip", "1") then reload.

Script and bundle load flow:

sequenceDiagram
  participant ComfyUI
  participant register as register_wolfbypass.js
  participant scripts as /extensions/.../scripts/
  participant Editor as python_script_editor
  participant Block as block_strength_editor

  ComfyUI->>register: load (only .js in web/)
  register->>scripts: GET codemirror_bundle.js
  scripts-->>register: codemirror_bundle.js_
  register->>scripts: GET python_script_editor[.min].js
  scripts-->>register: script
  register->>scripts: GET block_strength_editor[.min].js
  scripts-->>register: script
  Note over register: window.WolfCodeMirror set before Editor runs
  register->>Editor: execute
  register->>Block: execute

---

Documentation

• CHANGELOG — version history and session summaries.
• Block strength format and API — block order, JSON format, scan API.
• Scriptable Block Merge — merge mode, contract, injected variables, examples.
• Custom Merge Recipes — strength profiles: sandwich, ramp, Gaussian, img/txt, inverse, curves, copy-paste.
• Python script editor (CodeMirror) — build and bundle.
• Image ops / shaders (WolfSigmas) — advanced PyTorch shader design (caustics, torch light, filter operators).

---

Changelog

See CHANGELOG.md for version history and session summaries (e.g. CodeMirror, minify, config, script loading, editor resize).

---

Requirements

• ComfyUI with bypass LoRA support (core or patched).
• safetensors for LoRA scanning (block list from file keys only).
• Block Strength Plot: uses PIL; no matplotlib required (placeholder if dependencies missing).

License

Same as your ComfyUI / project.