| Node | Description | |---|---| | ⚡ Winnougan LTX Resolution & Frame Calculator | Solves LTX's strict dimension rules — pick a preset or enter custom dimensions, wire outputs directly to your latent node | | 🔥 Winnougan LTX Sigma Schedule | Preset sigma schedules for LTX-2.3 Stage 1 and Stage 2 — no more copy-pasting cryptic number strings | | 🔥 Winnougan LTX NAG Guidance | Normalized Attention Guidance for LTX-2.3 distilled models — restores effective negative prompting | | 🔥 Winnougan LTX Save Conditioning | Save Gemma text encodings to disk as .safetensors — encode once, reuse forever | | 🔥 Winnougan LTX Load Conditioning | Load saved conditioning without running the text encoder — faster iterations, no VRAM wasted on Gemma | | Winnougan LTX Image Resize | Resize images to match LTX-valid dimensions (÷32) | | Winnougan LTX Latent Preview | Preview latent frames inline during generation | | Winnougan LTX Sound Notification | Play a sound when generation finishes — never alt-tab check again |

🚀 Installation

Option 1 — ComfyUI Manager (Recommended)

Search for ComfyUI_WLTX_nodes in ComfyUI Manager and install directly.

Option 2 — Manual Install

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/ComfyUI_WLTX_nodes.git

Restart ComfyUI. All nodes will appear under the Winnougan LTX category.

🖥️ Requirements

ComfyUI
LTX-2.3 model (distilled recommended)
Python 3.10+

⚠️ LTX-2.3 Rules You Need to Know

Before diving into the nodes, these are the constraints LTX enforces:

Width and height must be divisible by 32. Standard resolutions like 1080 and 720 are NOT valid — LTX uses 1088 and 736 instead.
Frame count must follow the formula 8n+1 (valid counts: 9, 17, 25, 33, 41, 49 … 97 …). Any other frame count will fail or produce garbage.
The distilled model doesn't respond to standard CFG negative prompting. Use the NAG Guidance node instead.

The WLTX nodes solve all three of these automatically.

📖 Node Reference & Tutorial

All nodes live under Winnougan LTX in the ComfyUI node menu.

⚡ Resolution & Frame Calculator

The first node to add to any LTX workflow. It handles all the math so you don't have to.

Inputs:

| Input | Description | |---|---| | preset | Resolution presets — all pre-snapped to ÷32. Includes landscape, portrait, and square. Use Custom to enter your own. | | target_width / target_height | Only used when preset is Custom | | fps | Playback framerate. LTX native is 24fps. Use 30fps for social media exports. | | duration_seconds | Video length in seconds — automatically rounded UP to nearest valid LTX frame count (8n+1) |

Outputs:

| Output | Wire to | |---|---| | width | EmptyLTXVLatentVideo width | | height | EmptyLTXVLatentVideo height | | draft_width / draft_height | Half-resolution inputs for Stage 1 draft pass | | frames | EmptyLTXVLatentVideo length (LTX 8n+1 snapped) | | raw_frames | PromptRelay max_frames or frame counters (plain seconds × fps) | | fps / fps_int | Pass to your sampler or video save node | | actual_seconds | Real duration after frame snapping | | info | Show Text node — displays a full wiring summary |

Available Presets:

LTX 512p (768×512) — fast draft
LTX 720p (1280×736) — standard quality
LTX 1080p (1920×1088) — high quality
LTX 1440p (2560×1440), LTX 4K (3840×2160)
Portrait variants for all sizes
Square 512 and Square 1024

Tip: Wire the info output into a Show Text node to see a full summary of what to connect where. It also warns you if any values were snapped.

🔥 Sigma Schedule

Replaces the ManualSigmas node with a human-readable preset dropdown. No more hunting for the right sigma values.

Inputs:

| Input | Description | |---|---| | preset | Select from the preset list or use Custom | | custom_sigmas | Comma-separated values. Only used when preset is Custom. |

Presets:

| Preset | Sigmas | Use for | |---|---|---| | Stage 1 — Distilled 4-step (fast) | 0.85, 0.725, 0.421875, 0.0 | Main generation, fast | | Stage 1 — Distilled 8-step (quality) | 1.0, 0.99375 … 0.421875, 0.0 | Main generation, more detail | | Stage 2 — Upscaler 3-step | 0.421875, 0.2109375, 0.0 | Refinement after spatial upscaler | | Stage 2 — Upscaler 4-step | 0.525, 0.421875, 0.2109375, 0.0 | Stronger upscaler refinement |

Outputs: sigmas → wire to SamplerCustomAdvanced, steps → wire to your sampler step count, schedule_info → Show Text

Tip: Use Stage 1 for your main generation pass and Stage 2 after your spatial upscaler node for the refinement pass.

🔥 NAG Guidance

Normalized Attention Guidance for LTX-2.3 distilled models. The distilled model doesn't work with standard CFG negative prompting — NAG fixes that by patching the attention mechanism instead.

Inputs:

| Input | Description | |---|---| | model | Your loaded LTX model | | preset | Strength preset. Balanced is the right choice for most generations. | | nag_scale | How hard NAG pushes away from the negative prompt. Only used when preset is Custom. | | nag_alpha | Blend between guided and original attention (0–1). Only used when preset is Custom. | | nag_tau | Clipping threshold — lower = more aggressive. Only used when preset is Custom. | | negative_conditioning | Wire your negative CONDITIONING here | | inplace | In-place tensor ops to save a tiny bit of VRAM. Leave off unless extremely tight on VRAM. |

Presets:

| Preset | Scale | Alpha | Tau | When to use | |---|---|---|---|---| | Subtle | 6.0 | 0.15 | 3.5 | Gentle steering | | Balanced (recommended) | 11.0 | 0.25 | 2.5 | Most generations | | Strong | 16.0 | 0.35 | 2.0 | Aggressive negative prompting | | Maximum | 25.0 | 0.50 | 1.5 | Very strong avoidance |

Outputs: model (patched — wire to sampler), nag_info → Show Text

How to wire it:

[LTX Model] → NAG Guidance (model in)
[Negative Conditioning] → NAG Guidance (negative_conditioning)
NAG Guidance (model out) → Sampler

Note: NAG patches video AND audio cross-attention blocks when using LTX-2.3 with audio support. The nag_info output confirms whether audio was patched.

🔥 Save & Load Conditioning

Two nodes that work together to cache Gemma text encodings to disk. Encoding a prompt with Gemma takes time and VRAM on every run. If you're iterating on seeds, resolution, or samplers while keeping the same prompt — you're paying that cost unnecessarily.

Save once. Load in milliseconds on every subsequent run. No Gemma loaded, no VRAM used.

Save Conditioning

| Input | Description | |---|---| | conditioning | Wire your CONDITIONING here — it passes through unchanged | | filename | Name for the saved file (no extension needed) | | dtype | bfloat16 recommended. float16 for slightly smaller files. |

Saves to: ComfyUI/models/embeddings/<filename>.safetensors

The node UI shows the save path, file size, and timestamp after saving.

Load Conditioning

| Input | Description | |---|---| | file_name | Dropdown of all .safetensors files in your embeddings folder | | device | cpu = load to RAM (safe), gpu = load direct to VRAM (faster) |

Tip: Save your positive and negative conditioning separately (e.g. my_scene_positive.safetensors and my_scene_negative.safetensors). You can then skip the text encoder entirely on all subsequent runs and just use the Load nodes.

🔔 Sound Notification

Plays a sound file when your generation finishes. Drop it at the end of your workflow and never stare at the progress bar again.

Connect any audio file path (.mp3, .wav, .ogg, .flac, .aac, .m4a) to the node. The file is served through ComfyUI's server and plays in your browser tab when execution reaches the node.

🔗 Minimal LTX-2.3 Workflow

⚡ Resolution & Frame Calculator
     ↓ width, height, frames
EmptyLTXVLatentVideo
     ↓ LATENT

🔥 Sigma Schedule (Stage 1 — 4-step)
     ↓ sigmas, steps

[LTX Model] → 🔥 NAG Guidance → patched model
                ↑ negative conditioning

[Text Encoder] → positive conditioning
              → negative conditioning → NAG Guidance

SamplerCustomAdvanced
     ← patched model
     ← positive conditioning
     ← LATENT
     ← sigmas
     ↓ LATENT

VAE Decode → Video Combine → Save Video

🔔 Sound Notification (wire at the end)

With conditioning cache (faster iterations):

First run:
[Text Encoder] → 🔥 Save Conditioning → conditioning → sampler

Subsequent runs:
🔥 Load Conditioning → conditioning → sampler   ← no Gemma needed

❤️ Support

If these nodes save you time, consider supporting on Patreon — exclusive workflows, nodes, and LLM setups drop there first.

📄 License

MIT