LTX 2.3 ComfyUI Setup Guide

1. Install ComfyUI

Clone the ComfyUI repository and install dependencies. Requires Python 3.10+ and a CUDA-capable GPU with 16GB+ VRAM.

git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt

2. Install ComfyUI-LTXVideo Nodes

Open ComfyUI Manager, search for "LTXVideo", and install the official Lightricks nodes. Or clone manually:

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

3. Download LTX 2.3 Model

Choose the right model for your VRAM. Place checkpoint files in ComfyUI/models/checkpoints/.

• 16GB, RTX 40xx+:ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors — v1.1 Distilled FP8 by Kijai (~25GB). Fastest, best quality for 16GB.
• 16GB + LoRA:ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors — Dev FP8 by Kijai (~25GB). Use this if you want to apply LoRA weights.
• 16GB, older GPU:ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors — MXFP8 variant for GPUs without standard FP8 support.
• 16/24GB, RTX 50xx:ltx-2.3-22b-dev-nvfp4.safetensors — Official NVFP4 dev model (21.7 GB). Native nvfp4 matmul on Blackwell. Falls back to slow paths on older GPUs.
• 24GB VRAM:ltx-2.3-22b-distilled-1.1.safetensors — Official v1.1 with sequential offloading enabled in ComfyUI settings.
• 32GB+ VRAM:ltx-2.3-22b-distilled-1.1.safetensors — Official v1.1 full bf16 precision. Best quality, recommended for 32GB.
• 32GB, training:ltx-2.3-22b-dev.safetensors — Full dev model. Use only for LoRA training or fine-tuning.

4. Download Required VAE

The TAE (Tiny AutoEncoder) is required for all LTX 2.3 workflows. Place in ComfyUI/models/vae/. For audio-conditioned workflows, also download the Audio VAE.

# Required — Download from: https://huggingface.co/Kijai/LTX2.3_comfy
# File: taeltx2_3.safetensors → ComfyUI/models/vae/

# Audio-to-video workflows only:
# File: LTX23_audio_vae_bf16.safetensors → ComfyUI/models/vae/

5. Download the Gemma 3 12B Text Encoder

LTX 2.3 uses Gemma 3 12B IT as its text encoder. Every workflow needs one of these files in ComfyUI/models/text_encoders/. Pick by VRAM, not by preference — the full BF16 file OOMs alongside the transformer on 16/24 GB cards.

• 16GB / 24GB (recommended):gemma_3_12B_it_fp4_mixed.safetensors (9.5 GB, ~90% FP4 layers) — leaves headroom for the transformer.
• 16GB / 24GB (alternative):gemma_3_12B_it_fp8_scaled.safetensors (13.2 GB) — slightly higher precision than FP4 at 3.7 GB more VRAM.
• 32GB+:gemma_3_12B_it.safetensors (24.4 GB) BF16 from Comfy-Org. IMPORTANT: rename the file to comfy_gemma_3_12B_it.safetensors after download — every workflow JSON references that exact name.
• Source:All three: https://huggingface.co/Comfy-Org/ltx-2/tree/main/split_files/text_encoders

6. Load a Workflow

Use the ComfyUI Workflow Configurator on the homepage to pick a JSON for your GPU, mode, and control type — or download official example workflows from the ComfyUI-LTXVideo repository. Drag the JSON into ComfyUI to load it. The Configurator also tells you which model file and encoder file to select in each node, so you don't have to guess.

7. Use IC-LoRAs (Control, HDR, LipDub, Motion Track)

IC-LoRAs are In-Context LoRAs that attach on top of the dev model to add structural control, HDR output, lip-sync, or trajectory-driven motion. They share one rule: download the .safetensors into ComfyUI/models/loras/, then load through LTXICLoRALoaderModelOnly (not the normal LoraLoader) so the Reference Downscale Factor of 0.5 is applied. Pair each IC-LoRA with the distilled LoRA v1.1 at strength 0.5 for speed.

• Union Control (Canny + Depth):ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors. Pick "V2V Composition" in the Configurator — it generates LTX-2.3_ICLoRA_Union_Control_Distilled.json with the right LTXICLoRALoaderModelOnly + LTXAddVideoICLoRAGuide wiring. Requires comfyui_controlnet_aux (DWPose + Canny preprocessors) and ComfyUI-DepthCrafter-Nodes for depth.
• Motion Track:ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors. Pick "Motion Track" in the Configurator. Draw motion splines directly inside the LTXVSparseTrackEditor node (bundled with ComfyUI-LTXVideo, no extra plugins). Use it to animate a still image along your trajectories.
• HDR:ltx-2.3-22b-ic-lora-hdr-0.9.safetensors + ltx-2.3-22b-ic-lora-hdr-scene-emb.safetensors. Toggle "HDR output" in the Configurator (needs 24 GB+ — uses dev FP8). Outputs linear HDR EXR. Set OPENCV_IO_ENABLE_OPENEXR=1 in your shell before starting ComfyUI, otherwise the EXR save fails.
• LipDub:ltx-2.3-22b-ic-lora-lipdub-0.9.safetensors. No Configurator preset yet — use the official IC-LoRA workflow from Lightricks/ComfyUI-LTXVideo, swap the IC-LoRA file to lipdub, and pair with the audio VAE (LTX23_audio_vae_bf16.safetensors).
• Common gotcha:Loading IC-LoRAs through plain LoraLoader breaks the Reference Downscale Factor — outputs come back blurry or with no control influence. Always use LTXICLoRALoaderModelOnly + LTXAddVideoICLoRAGuide.
• Version pinning:Always pair IC-LoRAs with the distilled LoRA v1.1 — the official JSONs ship v1.0 references by mistake. v1.0 on a v1.1 base causes stiff motion and color drift.

8. Key Parameters

• Resolution:Both width and height must be divisible by 32. Use 1280×704 (not 1280×720) and 1920×1088 (not 1920×1080)
• Frames:Must be 8n+1: 65, 97, 121, or 161 frames. Official range: 65–257.
• Steps (Distilled):8 steps, CFG=1, sampler=euler_ancestral_cfg_pp. Do NOT raise CFG — doubles VRAM without quality gain.
• Steps (Dev):20–50 steps, CFG=3–5. Required for LoRA training.
• Scheduler:euler_ancestral_cfg_pp recommended (verified from official JSONs). HDR workflow uses plain euler_ancestral with ManualSigmas.
• FP8 vs MXFP8 vs NVFP4:FP8 scaled: RTX 40xx+ FP8 matmul. MXFP8 block-32: alternative format for GPUs without standard FP8 support. NVFP4: Blackwell-only (RTX 50xx) — falls back to slow paths on older cards, so prefer FP8 on RTX 40xx.
• Upscalers:Spatial x1.5 / x2 for resolution upscaling (models/latent_upscale_models/). Always use the v1.1 spatial upscaler — v1.0 produces text/logo artifacts in the last frames. Temporal x2 to double frame count.