📚

ComfyUI Tutorial

Download Models Prompts Flux Workflow 3D Workflow ControlNet Audio Workflow Video Workflow LTX Workflow Upscale Workflow Manager Wan2.2
/LTX Workflow

Mastering ComfyUI LTX 2.3: A No-Nonsense Guide to High-Fidelity Video Generation

If you have spent any time in the open-source video generation space, you already know the pain. You string together thirty different nodes, pray your VRAM doesn't cap out, and after twenty minutes of rendering, you get a video where the subject grows a third arm and the background melts into a low-res soup.

Then comes the LTX-Video architecture. With the recent updates surrounding LTX 2.3, we are finally looking at a workflow that balances prompt adherence, temporal consistency (things actually staying the same shape frame-to-frame), and reasonable hardware demands.

This isn't another generic overview. We are going to build a practical, rock-solid Text-to-Video (T2V) and Image-to-Video (I2V) pipeline in ComfyUI using LTX 2.3. By the end of this guide, you will understand exactly what each node is doing, why you are tweaking specific parameters, and how to stop your GPU from bursting into flames.

1. The Reality Check: Hardware & Prerequisites

Before we start dragging nodes around, let's talk about what you actually need to run this locally. LTX is efficient, but it's not magic.

What You Need:

  • GPU: An NVIDIA GPU with at least 12GB of VRAM is highly recommended. You can scrape by on 8GB with extreme optimizations and low frame counts, but 12GB+ (like an RTX 3060, 4070, or better) is where you want to be for a smooth workflow.
  • ComfyUI: Updated to the absolute latest version. Don't use a build from three months ago.
  • ComfyUI Manager: If you aren't using the Manager node by now, you are making life unnecessarily difficult. We will rely on it to fetch missing custom nodes.
  • The LTX Models: You need the core LTX 2.3 safetensors file (usually placed in your models/checkpoints folder) and the specific VAE if it isn't baked into the main model.

Pro Tip: Make sure your system pagefile is set to at least 32GB. When ComfyUI shifts weights from RAM to VRAM, a small pagefile will cause silent crashes that look like out-of-memory (OOM) errors.

2. Core Concepts: Why LTX Behaves Differently

If you are coming from AnimateDiff or Stable Video Diffusion (SVD), you need to rewire how you think about video generation.

AnimateDiff essentially brute-forces temporal consistency by sliding a context window over standard Stable Diffusion frames. SVD uses image conditioning to guess what happens next. LTX 2.3, however, is a native diffusion video model built on a DiT (Diffusion Transformer) architecture.

What does this mean for you?

  1. Prompting is completely different: LTX understands actions, camera movements, and timing much better than older models. You don't need to spam (masterpiece, best quality, 8k, trending on artstation). You need to write like a film director.
  2. Resolution matters: Transformer models are trained on very specific buckets of resolutions and frame counts. If you try to generate a 512x512 video when the model expects 768x512, it won't just look bad—it will output absolute garbage.
  3. CFG is hyper-sensitive: The Classifier Free Guidance scale in LTX isn't as forgiving as SDXL. A small bump can lead to deep-fried, over-saturated pixels.

3. Building the LTX 2.3 Workflow Step-by-Step

Let’s build a clean Text-to-Video pipeline from scratch. Open a blank ComfyUI canvas.

Step 1: Loading the Foundation

Right-click on the canvas, go to Add Node > loaders > Load Checkpoint. Select your LTX 2.3 model. This node will output your MODEL, CLIP, and VAE.

Note: Some LTX implementations use specific custom loader nodes (like LTX Model Loader). If you installed an LTX-specific custom node pack via ComfyUI Manager, use their proprietary loader to ensure the transformer blocks are parsed correctly.

Step 2: The Text Encoders (Conditioning)

LTX heavily relies on high-quality text encoding. You need two CLIP Text Encode (Prompt) nodes. Connect the CLIP output from your loader to both.

The Positive Prompt (The "What and How"): Forget comma-separated keywords. Write descriptive sentences. Example: A cinematic wide shot of a futuristic sports car driving down a neon-lit cyber city street at night. The camera pans from right to left, tracking the car. Photorealistic, reflections on wet asphalt.

The Negative Prompt: Keep it simple. blurry, distorted, morphed, bad anatomy, low resolution, static.

Step 3: Latent Configuration (The Canvas)

This is where most people mess up. You need an Empty Latent Video node (or the LTX-specific equivalent).

Here are the golden rules for LTX 2.3 latents:

  • Width and Height: Stick to multiples of 32. 768x512 or 512x768 are the sweet spots for testing. Going straight to 1024x576 will likely nuke your VRAM.
  • Batch Size (Frames): This dictates the length of your video. Start with 17 or 33 frames. Why odd numbers? Video models often require an anchor frame (1 + 16, or 1 + 32) for the conditioning context.
  • Frame Rate: Usually handled downstream, but keep in mind that 33 frames at 8fps gives you a nice, manageable 4-second clip.

Step 4: The KSampler (The Engine)

Add a standard KSampler node. Connect your Model, Positive Conditioning, Negative Conditioning, and the Empty Latent Video.

Let's dial in the settings:

  • Seed: Set to Randomize for exploration, Fixed when you are tweaking a video you already like.
  • Steps: 20 to 30. Going above 40 rarely improves LTX outputs and just wastes your time.
  • CFG: Keep it low. Start at 3.0. If the video ignores your prompt, bump it to 4.0. If you see burned-in colors or heavy artifacting, drop it to 2.5.
  • Sampler Name: euler or euler_ancestral. DiT models generally respond incredibly well to Euler.
  • Scheduler: normal or sgm_uniform.
  • Denoise: 1.0 (Since we are doing pure Text-to-Video).

Step 5: Decoding and Output

Finally, pull the LATENT output from the KSampler into a VAE Decode node. Connect the VAE from your very first Checkpoint Loader to this node.

Take the IMAGE output from the Decoder and plug it into a Video Combine node (you can get this from the ComfyUI-VideoHelperSuite custom node).

  • Set the Frame Rate to 8 or 12 or 24 depending on your preference.
  • Set the format to video/h264-mp4 so it plays natively in your browser.

Hit Queue Prompt. Grab a coffee. If your setup is correct, you should see a surprisingly cohesive short video pop out.

4. Advanced Tactic: Image-to-Video (I2V)

Text-to-Video is cool, but Image-to-Video is how you actually get usable assets for projects. You want to generate a perfect still image in Midjourney or SDXL, and then bring it to life with LTX.

To convert the workflow we just built into I2V:

  1. Add a Load Image node.
  2. Add a VAE Encode node. Connect your image to it, and pass the resulting latent into the KSampler (replacing the Empty Latent Video node).
  3. The Catch: LTX needs to know this is an image context. You must use the specific LTX Image Conditioning custom node. You feed your loaded image into this node, and it injects the visual data directly into the MODEL or CONDITIONING stream (depending on the specific wrapper you are using).
  4. Denoise adjustment: Set your KSampler denoise to 0.8 or 0.85. If you leave it at 1.0, it will completely destroy your input image and generate something entirely new based on the text prompt. If you set it too low (like 0.4), the video won't move at all.

5. Troubleshooting the "Jank"

Even with a perfect setup, things go wrong. Here is how to fix the most common LTX 2.3 headaches:

Problem: The video turns into a grey, noisy mess halfway through.Fix: Your CFG is too high, or your prompt is asking for an action the model doesn't understand. Lower the CFG by 0.5. If it still happens, simplify your prompt.

Problem: The subject moves, but the background stretches like a rubber band.Fix: This is a classic DiT artifact. Add negative prompts like warped background, distorted perspective, panning. You can also try lowering the total frame count. Sometimes, asking a model trained on short clips to generate 60 frames pushes it out of its latent comfort zone.

Problem: CUDA Out of Memory (OOM).Fix:

  1. Close your browser tabs (especially YouTube or anything hardware-accelerated).
  2. Lower your resolution to 512x512.
  3. Lower your frame count to 17.
  4. If you have the ComfyUI-Manager, install the FP8 weights of the LTX model instead of the standard FP16. It cuts VRAM usage almost in half with a very minimal hit to quality.

Conclusion

Getting LTX 2.3 running smoothly in ComfyUI requires a bit of patience, but the payoff is immense. You are moving away from the jerky, AI-fever-dream aesthetics of older models and stepping into genuine, controllable video synthesis. Start small with low resolutions and short frame counts, lock in your prompting style, and only scale up when you have a seed and composition you absolutely love.