javawock7618/comfy-ltx-2.3-workflows

LTX-2.3 Distilled GGUF & INT8 Workflows Collection

This repository provides a comprehensive collection of optimized ComfyUI workflows tailored for the lightricks/ltx-2.3 base model (specifically using Distilled GGUF and INT8 variants). These workflows leverage distilled, quantized capabilities to enable high-speed, high-quality video generation, upscaling, multi-subject synthesis, and audio generation directly within ComfyUI.

---

Workflow Files & Overview

The repository contains the following experimental and production-ready workflow files.
(Note: YYMM.x in the file names represents the release year, month, and version number, e.g., 2606.1 or 2607.1)

#	File Name	Workflow Target	Key Features & Quick Notes
1	LTX2.3_i2v-javanoYYMM.x.json	Image-to-Video (Long Gen)	Standard Image/Text-to-Video. Supports Single/Extend mode, Double-Frame mode for smooth motion, and prompt enhancement.
2	LTX2.3_int8_i2v-javanoYYMM.x.json	Image-to-Video (INT8 / Long Gen)	High-speed INT8 variant for long video workflows. Offers up to 1.6x speedup over GGUF when running on Ampere-generation GPUs.
3	LTX2.3_v2v-javanoYYMM.x.json	Video-to-Video	Modifies existing videos via Motion Track (OpenPose/Canny/Depth) or Inpaint Edit ([Add], [Remove], [Replace]) modes.
4	LTX2.3_int8_v2v-javanoYYMM.x.json	Video-to-Video (INT8)	High-speed INT8 variant for video-to-video processing. Delivers optimized generation speed on Ampere architecture GPUs.
5	LTX2.3_Director-javanoYYMM.x.json	Director Control (Multi-Key-Frame)	Keyframe-driven composition. Visually sequence multiple images, adjust timing, and control dynamic camera choreography.
6	LTX2.3_int8_Director-javanoYYMM.x.json	Director Control (INT8)	[NEW] High-speed INT8 variant for Director Control. Accelerates timeline-driven multi-keyframe synthesis and camera panning workflows.
7	LTX2.3_Detailer-javanoYYMM.x.json	Detailer / Upscaler	High-fidelity latent/spatial upscaling. Eliminates artifacts, improves facial consistency, and smooths noise.
8	LTX2.3_MSR-javanoYYMM.x.json	Multiple Subject Reference	Synthesizes videos by simultaneously referencing up to 5 individual elements (1-4 subjects + 1 background).
9	LTX2.3_int8_MSR-javanoYYMM.x.json	Multiple Subject Reference (INT8)	Highly optimized INT8 variant for multi-subject processing. Significantly reduces generation latency during complex multi-asset visual blending.
10	LTX2.3_MF-javanoYYMM.x.json	Multi Key-Frame Gen	Visually schedules multiple keyframes. Utilizes an optimized 2-pass generation and hybrid sampling pipeline for visual continuity.
11	LTX2.3_TTS-javanoYYMM.x.json	Text-to-Speech (Audio Only)	Bypasses video rendering completely to harness LTX-2.3's native acoustic capabilities for fast speech synthesis.

---

Detailed Workflow Breakdown

1 & 2. Image-to-Video / Long Video (i2v / int8_i2v)

Designed for extended video creation using fast Distilled GGUF or highly optimized INT8 processing based on the lightricks/ltx-2.3 architecture.

• INT8 Acceleration: The dedicated LTX2.3_int8_i2v-javanoYYMM.x.json workflow introduces native INT8 quantization support. On Ampere architecture GPUs (e.g., RTX 30-series, A100), this version can achieve up to a 1.6x generation speed improvement compared to standard GGUF variants.
• Core Modes: Seamlessly switch between a one-time video generation or iteratively extending the video sequence.
• Prompt Enhancement: Features three prompt modes—Ollama-driven enhancement, native LTX prompt enhancement, and plain text. If you do not use Ollama, you can safely bypass or delete the Ollama SubGraph node.
• Stability Features: Includes a toggleable preview during sampling, audio-driven syncing, and Double-Frame mode to eliminate facial/structural distortion during high-motion scenes.

3 & 4. Video-to-Video (v2v / int8_v2v)

Specialized in modifying existing video content or transferring motion from a reference source, now available in both standard GGUF and accelerated INT8 versions.

• INT8 Optimization: The LTX2.3_int8_v2v-javanoYYMM.x.json alternative brings the same high-efficiency Ampere-generation speedups to video translation and inpainting tasks.
• Motion Track Mode: Extract pose and structural data via Depth, Canny, or OpenPose (default) from a source video to animate a target starting image.
• Inpaint Edit Mode: Target specific elements to add, remove, or replace objects utilizing descriptive action tags (e.g., [Add], [Remove], [Replace]).

5, 6 & 10. Director Control & Multi Key-Frame Generation (Director / int8_Director / MF)

These workflows share a very similar core framework powered by WhatDreamsCost custom node suites (keyframeflf/flf), focusing on advanced image ordering and keyframe scheduling.

• INT8 Multi-Keyframe Speedup (New): The LTX2.3_int8_Director-javanoYYMM.x.json workflow provides a powerful option for heavy timeline setups, reducing computation times when sequencing multiple reference keyframes.
• Visual Timeline Control: These workflows allow you to visually specify image sequences, precise rendering order, and frame timing directly on a visual layout.
• Camera & Sampling: Tailor complex camera panning and transitions.

7. Detailer / Upscaler (Detailer)

A refinement pipeline dedicated to enhancing clarity and correcting visual issues in generated video outputs.

• Latent/Spatial Upscaling: High-fidelity detailing that enhances facial consistency and overall texture resolution.

8 & 9. Multiple Subject Reference (MSR / int8_MSR)

Generates coherent videos by synthesizing multiple distinct visual components simultaneously, offering a high-speed INT8 processing option.

• INT8 Multi-Asset Efficiency: Processing up to 5 image inputs simultaneously can be heavy on memory bandwidth; the LTX2.3_int8_MSR-javanoYYMM.x.json workflow reduces overhead to significantly speed up compilation and sampling times on supported GPUs.
• Multi-Subject Matrix: Reference up to 5 individual assets (1–4 distinct subjects and 1 background image) to create a single unified scene.
• Dynamic Prompt Pairing: Since explicit image indexing isn't natively supported, optimal blending is achieved by describing the unique visual characteristics of each reference image within the global text configuration.
• Voice Cloning Support: Features integrated ID-LoRA (id-lora-talkvid-3k, id-lora-celebvhq-3k) compatibility for specialized audio/voice cloning synchronized with the generation.

11. Text-to-Speech (TTS)

An experimental audio-only workflow focused strictly on leveraging LTX-2.3’s native acoustic engine.

• Pure Audio Processing: Completely bypasses the video rendering channels, drastically saving VRAM and generation time.
• Multilingual & Conditioning: Handles script and language definitions inside the main prompt. Features optional reference image loading to condition and clone the speaker's vocal characteristics.

---

Prerequisites & Installation

To ensure all nodes load correctly without errors, please update your environment to the latest versions:

1. ComfyUI Core: Update ComfyUI to v0.24.0 / Frontend 1.45.15 or newer.
2. Custom Nodes:
   • Update ComfyUI-KJNodes and ComfyUI-LTXVideo via the ComfyUI Manager.
   • Install specific keyframe/director custom node suites via ComfyUI Manager -> Install Missing Custom Nodes if any node appears red upon loading.
3. Required Models:
   • Base Model / Quantized Weights: Download the appropriate lightricks/ltx-2.3 Distilled GGUF checkpoints or the corresponding INT8 model weights depending on your selected workflow.
   • Upscaler: Download ltx-2.3-spatial-upscaler-x2-1.1 for workflows using the 2-pass detailer system.