This guide turns FutuTek’s LTX 2.3 walkthrough into a practical setup-and-run reference. You will load the required models, choose one of six workflow modes, provide the right input, and run cinematic AI video generations from one ComfyUI workflow.
By the end, you should have a working LTX 2.3 all-in-one ComfyUI workflow that can produce:
Generate a cinematic clip from a written scene prompt.
Use an audio clip plus a visual prompt to generate a video around the sound.
Animate a source image using a prompt and optional dialogue.
Combine a starting frame, an audio track, and a performance prompt.
Create a transition between a beginning image and ending image.
Transfer motion from a source video onto a new character or subject.
The video description lists these model/resources. Put each file in the matching ComfyUI folder, then restart ComfyUI or refresh model lists.
| What | Destination | Notes |
|---|---|---|
| All-in-one workflow/resources | Import into ComfyUI | Download from the creator’s Google Drive folder, then drag the workflow JSON into the ComfyUI canvas. |
| Low-VRAM LTX 2.3 GGUF model | ComfyUI/models/diffusion_models/ | Use for 12GB-class GPUs. |
| FP8 scaled LTX 2.3 model | ComfyUI/models/diffusion_models/ | Use when you have more VRAM, typically 16GB+. |
| Gemma text encoder | ComfyUI/models/text_encoders/ | Needed for LTX 2.3 prompt understanding. |
| LTX 2.3 text projection | ComfyUI/models/text_encoders/ | Keep the filename unchanged. |
| Audio VAE and video VAE | ComfyUI/models/vae/ | Required for audio/video decoding in the workflow. |
| Tiny VAE preview file | Workflow-dependent preview VAE location | Optional/recommended for faster previews; follow the file notes from the source page. |
| Spatial upscaler | ComfyUI/models/latent_upscale_models/ | Use the ltx-2.3-spatial-upscaler-x2-1.0.safetensors file. |
| Distilled LoRA and camera movement LoRAs | ComfyUI/models/loras/ | Used for speed and camera moves such as dolly-in effects. |
checkpoints unless your specific workflow loader asks for that.1 through 6.Use this for a full scene generated from text. In the video, the creator sets length and resolution, then prompts a blonde woman and little boy on a train traveling through old countryside. The workflow selects the text-to-video group automatically.
A blonde woman and a little boy riding a vintage train through an old countryside village, cinematic lighting, realistic faces, gentle camera movement, warm nostalgic mood.
Load an audio clip into the Load Audio node, then describe the visual subject and setting. The video example uses a YouTuber standing in a park introducing LTX.
A tech YouTuber standing in a green park, speaking to camera, natural daylight, realistic facial expression, subtle handheld camera movement.
Upload a source image and write a prompt that tells the model what movement and performance to add. Keep the prompt aligned with the source image so the identity, clothing, and setting do not drift too far.
A YouTuber standing on a rainy street, speaking to camera, raindrops visible on jacket, neon reflections, subtle head movement and natural gestures.
Set a short duration first, load the starting frame, choose an audio track, and optionally select a camera-movement LoRA such as Dolly In. The video uses a 12-second length and lower resolution for faster processing.
A singer performing emotionally on stage, close-up portrait, dolly-in camera movement, expressive eyes, realistic mouth movement synchronized to the song.
Load a beginning image and ending image. Use this when you need a controlled transition, then stitch multiple segments together into a longer scene.
A cinematic walk through a flower field, camera moving forward, soft wind through flowers, dreamy golden-hour lighting, smooth transition from the first frame to the final frame.
Load a source video for motion, then define the new starting frame/character and optional prompt. The video example transfers a Latin dance motion to Wonder Woman and Iron Man.
Wonder Woman and Iron Man dancing together, dynamic Latin dance movement, cinematic lighting, playful energy, realistic body motion.
| Problem | Most likely cause | Fix |
|---|---|---|
| Missing nodes when opening the workflow | ComfyUI or required custom nodes are outdated | Update ComfyUI, update custom nodes through ComfyUI Manager, restart, then reload the workflow. |
| Model not found | File is in the wrong folder or renamed | Move the file to the destination shown in this guide and keep the downloaded filename unchanged. |
| CUDA out of memory / VRAM failure | Resolution, length, or model choice is too heavy | Use the 12GB GGUF model, lower resolution, shorten duration, reduce upscale use, and close other GPU-heavy apps. |
| Audio-to-video runs but the scene is wrong | Prompt does not describe the visual subject clearly | Describe who is speaking, where they are, camera angle, lighting, and the kind of motion you expect. |
| Lipsync looks off | Duration, audio, or face framing is mismatched | Use a clear front-facing image, clean audio, shorter clips, and avoid extreme head turns on the first pass. |
| Motion transfer identity drifts | Source motion is too complex or the target frame is ambiguous | Use a clearer target frame, simpler source video, and a prompt that reinforces the subject’s outfit and environment. |