ComfyUI-rogala

Custom node pack for ComfyUI.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/rogala/ComfyUI-rogala

You can also install my mods via ComfyUI-Manager—just search for “rogala”

Restart ComfyUI. All nodes appear under the rogala menu.

Project structure

ComfyUI-rogala/
├── __init__.py                        # Entry point — registers nodes + API routes
├── pyproject.toml                     # Package metadata
├── node_list.json                     # Node list for ComfyUI Registry
├── README.md                          # User-facing documentation
├── CONTRIBUTING.md                    # This file
├── LICENSE
├── fonts/
│   ├── DejaVuSans.ttf                 # Default font for all nodes
│   ├── DejaVuSans-Bold.ttf            # Bold variant
│   ├── DejaVuSansMono.ttf             # Monospaced — ideal for numeric overlays
│   └── DejaVuSansMono-Bold.ttf        # Monospaced bold variant
├── nodes/
│   ├── _template.py                   # Copy this to create a new node
│   ├── sampler_scheduler_iterator.py
│   ├── aligned_text_overlay_images.py
│   ├── aligned_text_overlay_video.py
│   ├── ltx_resolution_selector.py
│   ├── fmlf_ltx.py
│   ├── sampler_ltxv.py
│   └── advanced_style_selector.py
├── js/
│   ├── doc_popup.js                   # Shared ? popup — used by ALL nodes
│   ├── align_the_nodes.js             # Canvas alignment toolbar — always active
│   ├── sampler_scheduler_counter.js   # Node-specific JS for SamplerSchedulerIterator
│   └── advanced_style_selector.js     # DOM widget for AdvancedStyleSelector
├── config/
│   ├── sampler_scheduler.json         # Reference list (auto-generated, do not edit)
│   ├── sampler_scheduler_user.json    # User's personal list (edit this)
│   ├── styles.json                    # All styles — name, category, prompt, negative_prompt, thumbnail
│   └── favorites_styles.json          # Favorited styles (managed by API, do not edit manually)
├── thumbnails/
│   └── styles/                        # Base thumbnails — one .jpg per style entry
│       ├── anime.jpg
│       └── ...                        # ~1000+ files, named by style key
└── workflows/                         # Example workflows in ComfyUI JSON format
    ├── LTX-2.3-fmlf_ltx-sampler_ltxv-AlignedTextOverlayVideo.json
    ├── Z-Image-Turbo_AdvancedStyleSelector.json
    ├── Z-Image-Turbo_AlignedTextOverlayImages.json
    └── Z-Image-Turbo_SamplerSchedulerIterator.json

SmartAttentionDispatcher

• SmartAttentionDispatcher is a ComfyUI node that patches model attention with SageAttention kernels (SA2 / SA3) to speed up image and video generation.

• Automatically detects GPU architecture, installed libraries, and available kernels. Supports SA2 (fp16, fp8, fp8++, triton) and SA3 (Blackwell only), with a dynamic mode that switches kernels per diffusion step. Works with Flux, SD3.5, SDXL, ErnieImage, Qwen, Z-Image, Wan, LTX and other DiT-based models.

• SA2 — Turing (RTX 20xx) and newer. Best results on Ada (RTX 40xx) and older hardware where PyTorch SDPA is not yet fully optimized for the architecture.

• SA3 — Blackwell (RTX 50xx) only. Requires CUDA 12.8+, sageattn3 package, SM 100+.

• Tested on RTX 5060 Ti (Blackwell, 16GB) — results vary significantly by model size and type. Image models: +8–12%. Video models with large sequences: up to +25–50% (ltx-2.3-22b-distilled, Wan2.2).

• Node status panel shows active mode, GPU tier, kernel availability, and model architecture after each run.

• Recommended launch flags: --fast

• Do not use --use-sage-attention together with this node.

  

---

Advanced Style Selector

A custom ComfyUI node for visual style selection with a built-in thumbnail gallery. Applies one or more styles to positive/negative prompts and encodes them directly to CONDITIONING — no extra CLIPTextEncode node needed.

---

Features

• Visual gallery — browse styles as thumbnails with category filtering and search
• Up to 6 styles simultaneously — selected styles are merged into a single conditioning output
• Two modes — Manual (click to select) and Iterator (cycles through styles automatically, one per queue run)
• Favorites — click ⭐ on any thumbnail to save it to config/favorites_styles.json, appears as a separate category at the top
• name_timestamp — optional checkbox to append a timestamp to style_name output for unique filenames when connected to Save Image
• save_prompt — optional checkbox to save positive and negative prompts to a JSON file in output/prompts/ after each run
• Category colors — each category has a unique color across the full hue wheel
• Hover popup — shows full prompt and negative prompt text on thumbnail hover
• Live reload — reload styles from disk without restarting ComfyUI
• Model thumbnail presets — create a subfolder in thumbnails/ named after your model (e.g. FLUX_1, WAN_2_2) and place style thumbnails there. Select the preset from the dropdown in the node — missing thumbnails fall back to the base styles/ folder. Folder names: letters, digits and underscores only.
• Theme aware — follows ComfyUI light/dark theme via CSS variables
• Resizable — gallery height grows with the node when you drag it taller
• Negative conditioning control — toggle between encoded negative or ConditioningZeroOut (for Flux, SD3, etc.)
• Styles format — config/styles.json, each entry has name, category, prompt, negative_prompt, thumbnail. Use {prompt} in prompt field to insert user text at a specific position.
• Installation — Find the node under rogala/Prompting → Advanced Style Selector

My Styles

• My Styles — save your own styles directly from the node. Enter a name in the Style name field and click Save Style — the current positive and negative prompts are saved to config/my_styles.json as a new style in the My Styles category.
• If a style with the same name already exists it will be updated. The category appears first in the list automatically.
• To add a thumbnail place a .jpg file named after your style in thumbnails/my_style/ and click Reload Styles.
• After updating the version via ComfyUI Manager, the my_styles.json file and the ./thumbnails/my_style/ folder are not deleted or modified.

---

---

Inputs

Pin	Type	Description
clip	CLIP	Connect from Load Checkpoint
positive_text	STRING	Your positive prompt. Styles are applied on top.
negative_text	STRING	Your negative prompt. Hidden when use_negative is OFF.

Hidden widget inputs (controlled by the gallery UI)

Widget	Description
use_negative	ON: encode negative normally. OFF: output ConditioningZeroOut
mode	Manual or Iterator
style_1 … style_6	Active style slots (synced by the gallery)
iterator_categories	Comma-separated categories to iterate over (empty = all)
iterator_seed	Seed for iterator order (0 = alphabetical)
append_counter	Enable name_timestamp suffix on style_name output

---

Outputs

Pin	Type	Description
positive	CONDITIONING	Encoded positive prompt with all selected styles applied
negative	CONDITIONING	Encoded negative prompt, or ConditioningZeroOut
style_name	STRING	Active style names joined by - for use in file naming

---

Modes

Manual

Select up to 6 styles by clicking thumbnails in the gallery. Active styles appear in the strip at the top of the panel as thumbnails with their slot number. Click × to remove a style, or use the 🗑 button to clear all.

Iterator

Cycles through all styles in the selected categories automatically — one style per queue run. Stops after the last style and resets. Use Reload Styles after changing the style list to reset the iterator.

---

Gallery UI

Element	Description
🗑 (left of categories)	Clear selected category filters
Category pills	Click to filter gallery by category. Active categories are highlighted in their category color
⭐ Favorites	First category if non-empty. Click ⭐ on any thumbnail to add/remove
Active strip	Shows currently selected styles as thumbnails. Hover for prompt details
🗑 (in strip)	Clear all selected styles
Search	Filter by style name
Thumbnail grid	Click to select/deselect. Badge shows slot number (1–6)
use_negative	Toggle negative conditioning mode
name_timestamp	Toggle timestamp suffix on style_name output
Mode selector	Switch between Manual and Iterator
Reload Styles	Hot-reload config/styles.json without restarting ComfyUI

---

Script for Aligning Nodes

A lightweight frontend extension that adds a persistent toolbar at the bottom of the canvas for aligning, distributing, and resizing selected nodes.

Features: align left / right / top / bottom · distribute horizontally and vertically (gap-aware, not center-based) · match width to widest node (aligns to leftmost) · deselect all

The toolbar auto-scales based on screen resolution (1080p / 2K / 4K) and includes manual size control (5 levels, −2 to +2) so the buttons stay comfortable at any display density. Localized tooltips (EN, UK-UA, DE, FR, ES, IT, PT-BR, ZH, JA)

Independent implementation inspired by KayTool (kk8bit) and ComfyUI-NodeAligner (Tenney95). No shared code — similar idea, different approach.

---

LTX Video 2.3 — FMLFLTX + SamplerLTXV

A two-node pipeline for LTX Video 2.3 distilled models with spatial upscaling. FMLFLTX_2.3 prepares the latent space from up to 6 guide images. SamplerLTXV_2.3 runs a two-pass denoise — low resolution first, then upscale and refinement.

Inspired by the workflows of WhatDreamsCost and princepainter.

---

FMLFLTX_2.3

Loads up to 6 guide images and distributes them evenly across the video timeline. Outputs video and audio latents ready for SamplerLTXV_2.3.

Inputs

Pin	Default	Description
video_vae	—	Video VAE for image encoding.
audio_vae	—	Audio VAE for empty audio latent creation.
img_compression	18	JPEG pre-compression strength (0 = disabled).
width	768	Latent width — connect from LTX Resolution Selector.
height	512	Latent height — connect from LTX Resolution Selector.
length	97	Frame count — connect from LTX Resolution Selector.
fps	25.0	Frames per second — connect from LTX Resolution Selector.
batch_size	1	Batch size.
image_1 … image_6	—	Guide images (optional). Connect only the slots you need.
strength_1 … strength_6	1.0	Conditioning strength per image (1.0 = fully conditioned).

Frame placement

Images are placed at evenly-spaced positions across the full video duration:

Images connected	Placement
1	frame 0
2	0%, 50%
3	0%, 33%, 66%
4	0%, 25%, 50%, 75%
5	0%, 20%, 40%, 60%, 80%
6	0%, 17%, 33%, 50%, 67%, 83%

Recommended number of images by video duration (targeting ~8 sec per image):

Duration	Images	Interval
10–15 sec	2	~7 sec
20–25 sec	3	~8 sec
30–35 sec	4	~8 sec
40–50 sec	5–6	~8 sec
50–60 sec	6	~8–10 sec

Outputs

Pin	Description
latent	Combined video + audio NestedTensor (legacy).
video_latent	Video latent → connect to SamplerLTXV_2.3 video_latent.
audio_latent	Empty audio latent → connect to SamplerLTXV_2.3 audio_latent.

---

---

SamplerLTXV_2.3

Two-pass sampler for LTX Video 2.3 distilled models.

Pass 1 — full denoise at the input resolution (video + audio NestedTensor).
Pass 2 — spatial upscale, re-embed first/last frames at upscaled resolution, refinement denoise.

Inputs

Pin	Description
model	Diffusion model.
video_vae	Video VAE — used for image encoding and upscale statistics.
positive	Positive conditioning.
negative	Negative conditioning.
video_latent	Connect from FMLFLTX_2.3 video_latent.
audio_latent	Connect from FMLFLTX_2.3 audio_latent.
start_image	First frame re-embedded after upscale (optional).
end_image	Last frame re-embedded after upscale (optional).
upscale_model	Latent upscale model for pass 2 (optional).

Pass 1 parameters

Parameter	Default	Description
add_noise_1	enable	Add noise before pass 1.
noise_seed_1	0	Noise seed for pass 1.
steps_1	8	Denoising steps for pass 1.
cfg_1	1.0	CFG scale for pass 1.
sampler_name_1	euler	Sampler for pass 1.
scheduler_1	linear_quadratic	Scheduler for pass 1.
denoise_1	1.0	Denoise strength for pass 1.

Pass 2 parameters

Parameter	Default	Description
add_noise_2	enable	Add noise before pass 2.
noise_seed_2	0	Noise seed for pass 2.
steps_2	3	Denoising steps for pass 2 (refinement).
cfg_2	1.0	CFG scale for pass 2.
sampler_name_2	euler	Sampler for pass 2.
scheduler_2	linear_quadratic	Scheduler for pass 2.
denoise_2	0.4	Denoise strength for pass 2 — lower = less change to upscaled latent.

Outputs

Pin	Description
video_latent	Upscaled and refined video latent → connect to VAE Decode.
audio_latent	Audio latent from pass 1 → connect to LTXV Audio VAE Decode.

Recommended settings

Resolution target	steps_1	steps_2	denoise_2
up to 960×544	8	3–4	0.4
1280×720	8–12	4–6	0.4
1920×1080	12	6	0.4

---

LTX Resolution Selector

Selects the correct input resolution and frame count for LTX Video models. Supports Dev mode and Distilled upscale modes (x1.5 and x2).

---

Inputs

Pin	Default	Description
mode	Dev (no upscale)	Render mode — see table below.
dev_target	960x544	Target resolution for Dev mode (12 options).
upscale_target	1920x1080	Target resolution for Distilled modes (6 options).
fps	24.0	Frames per second (1–120). Passed through to output.
duration_sec	5	Clip duration in seconds (1–300).

mode options:

Value	Description
Dev (no upscale)	Standard LTX Dev sizes, rounded UP to nearest multiple of 32.
x1.5 Distilled	Returns input size needed to reach the target after x1.5 upscale.
x2 Distilled	Returns input size needed to reach the target after x2 upscale.

Outputs

Pin	Type	Description
width	INT	Input width in pixels (multiple of 32).
height	INT	Input height in pixels (multiple of 32).
length	INT	Frame count: 1 + 8 x round(fps x sec / 8).
fps	FLOAT	Passthrough of the fps input.

Example

To render a 5-second clip at 1920x1080 using x2 Distilled:

• mode = x2 Distilled
• upscale_target = 1920x1080 (Landscape)
• fps = 24, duration_sec = 5
• Node outputs: width=960, height=544, length=121, fps=24

---

Sampler Scheduler Iterator

Iterates over sampler x scheduler combinations one pair per execution. Outputs each pair to connected nodes (e.g. KSampler, Aligned Text Overlay Images). Node title updates automatically: Iterator: Step 3 / 12. Queue stops automatically after the last combination.

Note: Press Refresh before first use. This reads all installed samplers and schedulers from ComfyUI and writes the reference file. The counter resets to 0.

---

Inputs

This node has no inputs. Configure combinations in .\ComfyUI\custom_nodes\ComfyUI-rogala\config\sampler_scheduler_user.json.

Outputs

Pin	Type	Description
sampler_name	SAMPLER	Current sampler — connect to KSampler.
scheduler	SCHEDULER	Current scheduler — connect to KSampler.
external_text	STRING	`"sampler

Example

Edit config/sampler_scheduler_user.json to select combinations:

{
    "samplers": ["euler", "dpmpp_2m", "dpmpp_3m_sde"],
    "schedulers": ["karras", "simple"]
}

Total = samplers x schedulers (here: 3 x 2 = 6 runs). Copy names from config/sampler_scheduler.json (generated by Refresh). Unknown names are silently ignored.

---

Aligned Text Overlay Images

Renders a multi-line text block onto an image at a chosen corner before saving. Supports %NodeTitle.param% template tags resolved from the active ComfyUI prompt.

---

Inputs

Pin	Default	Description
image	—	Input image tensor.
text_template	see default	Template string with optional %Node.param% tags.
vertical	bottom	Vertical anchor: top or bottom.
horizontal	right	Horizontal anchor: left or right.
font_size	16	Font size in points (10–50).
text_color	white	Text colour.
bg_color	black	Background colour (none = transparent).
bg_opacity	150	Background opacity (50–250).
external_text	—	Optional string appended after the resolved template.

Outputs

Pin	Type	Description
image	IMAGE	Outputs an image with a text block superimposed on top.

Example

Default template pulls values directly from a KSampler node:

seed: %KSampler.seed% | steps: %KSampler.steps%
cfg: %KSampler.cfg% | %KSampler.sampler_name% | %KSampler.scheduler%

NodeTitle must match the title shown on the node in the graph. Numeric sampler / scheduler indices are decoded to names automatically. Connect external_text from SamplerSchedulerIterator to append the current "sampler | scheduler" pair to the overlay.

---

Aligned Text Overlay Video

Renders a multi-line text block onto every frame of a video tensor. Supports %NodeTitle.param% template tags resolved from the active ComfyUI prompt.

---

Inputs

Pin	Default	Description
images	—	Video tensor (B, H, W, C).
text_template	see default	Template string with optional %NodeTitle.param% tags.
vertical	bottom	Vertical anchor: top or bottom.
horizontal	right	Horizontal anchor: left or right.
font_size	16	Font size in points (10–50).
text_color	white	Text colour.
bg_color	black	Background colour (none = transparent).
bg_opacity	150	Background opacity (50–250).
first_frame_only	false	Apply overlay only to the first frame (fast preview).
external_text	—	Optional string appended after the resolved template.

Outputs

Pin	Type	Description
images	IMAGE	Outputs a video tensor with text overlay applied.

Example

Connect between VAE Decode and video output:

VAE Decode → AlignedTextOverlayVideo → VHS Video Combine

Default template pulls values directly from a KSampler node:

seed: %KSampler.seed% | steps: %KSampler.steps%
cfg: %KSampler.cfg% | %KSampler.sampler_name% | %KSampler.scheduler%

If you have multiple samplers, rename them in the graph (Right Click → Title) and reference explicitly:

steps: %Sampler_1.steps%

NodeTitle must match the title shown on the node in the graph. Numeric sampler / scheduler indices are decoded to names automatically.

Connect external_text from SamplerSchedulerIterator to append the current "sampler | scheduler" pair to the overlay.

Enable first_frame_only for fast preview (applies overlay only to frame 0).