Installing LTX 2.3 locally on your own GPU takes about 2–4 hours start to finish — most of that is downloads. Once it's done, you can generate unlimited NSFW video with no per-clip cost. This guide walks through the full setup: ComfyUI, the LTX 2.3 weights, the Gemma 3 text encoder (abliterated for NSFW), the audio VAE, and the LoRAs you need.

If 2–4 hours of setup isn't worth your time, our AI Porn Generator runs all of this in your browser with 160 free daily tokens. If you want to install but want a one-click installer, the ComfyUI Workflow Pack bundles every model, encoder, LoRA and workflow — one command pulls everything from our Hugging Face repo into your ComfyUI install. Discord for support during setup.

What You Need Before Starting

• NVIDIA GPU with 24+ GB VRAM (4090 / 5090 / 3090 / A6000). 16 GB possible with GGUF Q4 — see VRAM requirements.
• ~80 GB free disk space for model weights, encoders, VAEs, LoRAs.
• Python 3.11 or 3.12. CUDA 12.1+.
• Stable internet — model download is ~60 GB total.

Step 1: Install ComfyUI

If you don't already have ComfyUI:

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

Open http://localhost:8188 to verify ComfyUI is running. If it loads, kill it and proceed.

Step 2: Install the Required Custom Nodes

LTX 2.3 needs several node packs. Easiest path is via ComfyUI Manager:

Restart ComfyUI after installing nodes — failures here are the most common reason workflows fail to load.

Step 3: Download the LTX 2.3 Model

Pick one quant based on your VRAM:

• 32 GB+ GPU (5090, A6000, A100): download ltx-2.3-22b-distilled-fp8_e4m3fn.safetensors from Lightricks/LTX-Video. Full quality with headroom.
• 24 GB GPU (4090, 3090): same FP8 file. Will run with text encoder offloading.
• 16 GB GPU: download ltx-2.3-22b-distilled-Q4_K_M.gguf from the QuantStack mirror. Smaller, slower but fits.

Place the file in ComfyUI/models/diffusion_models/. For comparison of quants and which to pick for your hardware, see LTX 2.3 quants compared.

Step 4: Download the Text Encoder

For SFW use, download stock Gemma 3 from the LTX repo. For NSFW use (uncensored prompt-following), use the abliterated variant — see the NSFW guide for why.

• SFW: gemma_3_12B_it_fp4_mixed.safetensors (~6 GB)
• NSFW: gemma-3-12b-it-heretic-v2_fp8_e4m3fn.safetensors (~12 GB)
• You also need: ltx-2.3_text_projection_bf16.safetensors (~50 MB)

Place all three in ComfyUI/models/text_encoders/.

Step 5: Download the VAEs

LTX 2.3 needs two VAEs — one for video, one for audio:

• LTX23_video_vae_bf16.safetensors (~700 MB)
• LTX23_audio_vae_bf16.safetensors (~300 MB)

Place both in ComfyUI/models/vae/.

Step 6: Download the LoRAs

For NSFW use, the minimum LoRA stack:

• ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors — distill base LoRA
• LTX-2.3-22b-AV-LoRA-talking-head-v1.safetensors — lipsync
• LTX2.3-NSFWMOTION_00750.safetensors — explicit motion
• LTX2.3_Soft_Enhance.safetensors — detail enhancement
• ltx2.3-transition.safetensors — for FML keyframe workflow

Find these on Civitai or our Hugging Face repo. Place in ComfyUI/models/loras/.

Step 7: Get a Working Workflow

You can find LTX 2.3 workflow JSONs on Civitai, the Lightricks GitHub, or various Reddit threads. Quality varies wildly. Two reliable sources:

• Lightricks official examples — basic I2V and T2V workflows. Good starting point but no NSFW LoRA stack.
• Our Workflow Pack — production-tuned NSFW workflows (BJ Cinema, Doggy Cinema, FML Frame Basic, Sex Cut Cinema). Drop-in JSONs with the right LoRAs and settings already configured.

Drop the JSON into ComfyUI's interface (drag-drop the file). If nodes show as missing, you forgot to install a custom node pack — go back to Step 2.

Step 8: First Generation

With the workflow loaded:

1. Drop a starting image into the LoadImage node
2. Edit the prompt in the CLIPTextEncode node
3. Hit “Queue Prompt”
4. Wait 8–15 minutes (depends on GPU and quant)

Recommended starter settings: 9 sampling steps, CFG=1.0, euler_ancestral_cfg_pp sampler, linear_quadratic scheduler. Tune from there. Full settings reference: LTX 2.3 best settings.

Common Install Errors

• “Missing custom node X” — check Step 2. Restart ComfyUI fully (kill the process, restart).
• “CUDA out of memory” — your quant is too big for your GPU. Drop down to GGUF Q4_K_M.
• “VAE decode produces black frames” — wrong VAE loaded. Make sure both video and audio VAEs are present.
• “Audio is silent” — audio VAE missing or LoadAudio node has bad file path. See audio guide.
• “Generation works but output ignores my prompt” — you're using stock Gemma 3 with a NSFW prompt. Switch to abliterated.

One-Click Installer

If those eight steps sound like a weekend, our ComfyUI Workflow Pack ships an installer that does steps 3–7 automatically. Run it once, the installer pulls every model weight, encoder, VAE, LoRA, and workflow JSON from our Hugging Face repo into the right ComfyUI folders. About 60 GB of downloads, then everything works. Discord support if anything goes wrong.

Or if you'd rather not install anything at all, the Studio runs every workflow in your browser. 160 free daily tokens, no GPU required.

Related Reading

• LTX 2.3 GPU and VRAM Requirements
• LTX 2.3 Quants Compared
• LTX 2.3 ComfyUI Workflow Guide
• Running LTX 2.3 in the Cloud
• LTX 2.3 on Mac: M-Series, MacBook, MLX