Text-to-Image Workflows, ergonomic, 'ready to use' graphs that let you focus on creating - not hunting scattered knobs.
Each preset includes the key nodes for its model family, a Dual-CFG implementation, and a custom LoRA hooks system for CLIP/UNet paths.

Presets:

SD-Master
SD3-Master
FLUX-Master

workflow pic

Swap CFG - Switch CFG from value A/B at a chosen point in the run.
Fast LoRA - Toggle LoRA on the CLIP path (fast) or Conditioning/UNet (classic), while keeping synchronized LoRA strength on both paths.
Low VRAM friendly - Optionally load models to CPU to free VRAM for sampling.
Comfort sliders - Safe default ranges; adjust step/min/max via the context menu.
Mini tips - Small hints for the most important nodes.
Lazy Prompt - Prompt panel with searchable history of your past prompts.

After installing PG Nodes, the workflows are available under Templates → PG-Nodes.

workflow2 pic

Lazy Prompt

Lazy Prompt pic

Save → Recall → Iterate The core idea of Lazy Prompt is comfort: write a prompt once, and bring it back instantly next time. The node keeps a compact on‑disk prompt history (positive & negative) and lets you quickly save, search, and re‑use past prompts like a mini prompt library. Optional add-ons (lens, time of day, lighting, color temperature).

Saves every run to history and lets you load any previous entry with a couple of clicks.
Inline randomizer {red|green|blue} for quick variations.
Optional helpers: camera/lens phrases, time‑of‑day, lighting direction, white balance temperature.

Details

Default history file: custom_nodes\prompt_history.json.
Default cap: max_entries = 500.
Tip: right‑click the node (⋯) to open the context menu and tweak the history file path or the maximum number of stored entries.

Lazy Prompt (ext)

Starting with version 1.5.0, I added “Lazy Prompt (ext)”, which includes outputs for interoperability with other nodes:

Lazy Prompt ext pic

Lazy Prompt maps

Default config file: PG_Lazy_Prompt_Config.json.
Typical location: inside this repo folder (when installed via Manager/Manual).
You can change the config file path anytime from the node’s context menu.
Maps include examples like:
- POS_LIGHT_MAP, NEG_LIGHT_MAP - direction‑dependent phrases.
- LENS_CHOICES - lens list (includes none).
- TIME_MAP - time‑of‑day phrases.

History storage

Defaults at runtime: history_path = custom_nodes\prompt_history.json, max_entries = 500.
You can adjust history path and max entries from the node’s context menu.

Lora Bridge and Hooks

lora1 pic

Connection Patterns
LoRA Bridge Loader
Merge LoRA Clip
Merge LoRA Model
Merge LoRA
Create LoRA Clip Hook
Create LoRA Model Hook
Create LoRA Hook
Set Hook Keyframes
Set Clip Hooks
Set Model Hooks

Connection Patterns

“Old-school” (baked-in):
LoRA Bridge Loader → Merge LoRA Clip/Model/Both → downstream
Weights are merged into CLIP/UNet (destructive bake). Fast at runtime, fixed strength.
CLIP path (non-destructive):
LoRA Bridge Loader → Create LoRA Clip Hook → [optional: Set Hook Keyframes] → Set Clip Hooks
Hooks modify the text encoder live; keyframes let you change LoRA strength over time.
MODEL/Conditioning path (non-destructive):
LoRA Bridge Loader → Create LoRA Model Hook → [optional: Set Hook Keyframes] → Set Model Hooks
Hooks patch the UNet path; keyframes control the strength timeline.

LoRA Bridge Loader

Loads up to four LoRA files once and outputs a single LORA_BRIDGE bundle, ready for hook creation or merge nodes.

Merge LoRA Clip

Merges (bakes) the selected LoRA into CLIP. Destructive; no runtime toggling.

Merge LoRA Model

Merges (bakes) the selected LoRA into the MODEL/UNet. Destructive.

Merge LoRA

Bakes the selected LoRA into both MODEL and CLIP in one step.

Create LoRA Clip Hook

Builds a HOOKS group targeting CLIP from the chosen slot in LORA_BRIDGE, with its own strength control.

Create LoRA Model Hook

Builds a HOOKS group targeting the MODEL/UNet path, with independent strength.

Create LoRA Hook

Builds a HOOKS group that targets both CLIP and MODEL simultaneously, exposing separate strengths for each path.

Set Hook Keyframes

Defines a two-point strength curve for a HOOKS group (percent of “model strength”).
Use to fade in/out or swap strengths at specific progress points.

Set Clip Hooks

Applies a HOOKS group to CLIP (append/merge). Non destructive; respects keyframes.

Set Model Hooks

Applies a HOOKS group to CONDITIONING/MODEL (append/merge). Non destructive; works with keyframed strengths.

Unified Loader

Unified Loader pic

One loader, more VRAM for sampling Unified Loader of Checkpoint, Diffusion UNet, VAE/TAESD, CLIP/CLIP-Vision, ControlNet, IP-Adapter, and Upscale models-with a simple CPU switch. Offloading parts of the stack to CPU often frees up valuable VRAM for the sampler (KSampler/DiT).

Checkpoint → CP_MODEL, CP_CLIP, CP_VAE
Diffusion UNet (incl. FP8 variants where available)
VAE and TAESD shortcuts: taesd, taesdxl, taesd3, taef1
CLIP (multi‑file), CLIP‑Vision
ControlNet, IP‑Adapter, Upscale models

Tip: The CPU switch is best‑effort. Depending on your setup and model family, not everything can be moved off-GPU, but when it works, the sampler gets more VRAM.

Swap CFG Guidance

Swap CFG pic

Set two CFG values and switch between them at a chosen progress (%). Early steps use CFG 1, then at the trigger point, sampling continues with CFG 2. Great for front-loading structure (higher CFG) and then relaxing detail (lower CFG),or the opposite.
Parameters: cfg1, cfg2, and swap_percent (0–100%).

Select Nodes

A compact suite of selectors for different model families. In the UniLoader/ subfolder you’ll also find variants compatible with the Unified Loader.
Available selectors: MODEL, CLIP, VAE, CLIP_VISION, CONTROL_NET, IPADAPTER, UPSCALE_MODEL, Sampler, and Scheduler.

Just Save Image

Just Save Image pic

Nodes Minimal nodes that save images in an IMAGE batch to ComfyUI’s output directory when a boolean toggle is ON. Saves without preview.

Just Save Image - output node (no outputs).
Just Save Image & Out - passthrough version that returns the input IMAGE.

Checkpoint Switch

Checkpoint Switch pic

A simple boolean controlled switch, between two (MODEL/CLIP/VAE) triplets.
Flip the input to route either set A or set B downstream.

XOR Toggle

A boolean switch, that flips two outputs in opposition.
When toggle = TRUE → Boolean 1 = TRUE, Boolean 2 = FALSE.
When toggle = FALSE → Boolean 1 = FALSE, Boolean 2 = TRUE.

Percent

Converts an INT input to a FLOAT by dividing by 100.

Installation

Via ComfyUI Manager

Open ComfyUI Manager.

Search for “PG Nodes”.

Install and restart ComfyUI.

Manual

Go to the repo Code → Download ZIP.

Extract the ZIP.

Rename the extracted folder to PG-nodes.

Move that folder to your ComfyUI custom nodes directory:

Windows (typical): C:\Users\<you>\Documents\ComfyUI\custom_nodes\PG-nodes

Or wherever your ComfyUI root lives → custom_nodes/PG-nodes

Restart ComfyUI.

In the node search, look for the PG category (e.g., Lazy Prompt, Unified Loader).

Versioning & Updates

[2.0.0]
 Added
- Workflows: SD-Master, SD3-Master, Flux-Master
- LoRA Hooks bundle: compact toolkit for LoRA on MODEL/CLIP
- Select Nodes: set of selector helper nodes used to pick model file names
- Checkpoint Switch: two‑way switch for (MODEL, CLIP, VAE) triplets.
- Swap CFG Guidance: two‑phase CFG guider (CFG1 → early, CFG2 → late).
- Percent: INT 0..100 → FLOAT 0..1 helper.
- XOR Toggle:
- Checkpoint Switch:

[1.5.5]
 Fix
- Unified Loader, DIFFUSION_MODEL output fix and Class name.

[1.5.0]
 Added
- Lazy Prompt (ext).

[1.0.0]
 Added
- Initial public release.
- Nodes: Lazy Prompt (+mini), Unified Loader (+mini), Just Save Image (+Out).
- Prompt history JSON + config file `PG_Lazy_Prompt_Config.json`.