How to Generate 30-Second LTX-2 AI Videos on Wan2GP with Low VRAM

TL;DR

To generate a 30-second LTX-2 video on Wan2GP, select the distilled model and set the frame count to 721, which splits processing across two sliding windows because one window supports 481 frames. The demonstrated run took more than 11 minutes on an RTX 4060 Ti with 16 GB VRAM, but consistency declined after 20 seconds. Read on for installation, storage, speed, memory, and audio guidance.

Transcript

I made this 27 second video using  Wan2GP, and it was very simple to do so.  This will be a quick video on  running LTX using the Wan2GP project.  There are two ways to install Wan-GP,  first I will try the easier method,   which is with the help of Pinokio. The second method is the manual   installation of this project from GitHub,  I will have to... Read More

Key Insights

• Wan2GP is simpler than a node-based ComfyUI workflow for users who prefer a web interface, but it provides less workflow customization. The demonstrated Pinokio installation automatically handled requirements and model downloads, although one failed requirement had to be installed again before setup completed.
• The distilled LTX-2 model is designed to produce its main sampling result in 8 steps. The observed workflow sampled twice, first for 8 steps and then for 3 steps, and Wan2GP did not provide the customization needed to remove the second sampling stage.
• Sage Attention 2 is compatible with most CUDA GPUs in the demonstrated interface. Sage Attention 3 can reportedly improve generation speed by 50 percent with a small quality tradeoff, but accessing newer Sage Attention or Triton options requires a manually installed standalone Wan2GP setup.
• The first LTX-2 generation required large automatic downloads before rendering began. Downloads included a distilled model of about 21 GB and a Gemma 3 model of about 13 GB, while the completed Pinokio installation occupied 79 GB on disk.
• A 10-second generation took about 2 minutes and 40 seconds during the initial run on an RTX 4060 Ti with 16 GB VRAM. A later run took 3 minutes and 30 seconds, with the first sampling stage completing within 70 seconds and the second requiring about 90 seconds.
• Low VRAM use is accompanied by high system memory use in the demonstrated configuration. The test computer used an i5-14400F, 64 GB of RAM, and an RTX 4060 Ti with 16 GB VRAM, suggesting that additional system RAM remains important when GPU memory use is reduced.
• A single LTX-2 sliding window supports 481 frames, corresponding to a 20-second video in the demonstrated settings. Selecting 721 frames for 30 seconds created two windows, and the second 10-second segment reused the first window's last frame and the same text prompt.
• Audio-guided generation produced the most consistent opening example when paired with a supplied audio file and a LoRA weight of 0.70. Model-generated audio was poor in almost all tested results, and the presenter recommended supplying an audio file instead of relying on LTX-2 to create it.
• "The third option is the distilled model, and I will select the same because this will give me the result in 8 steps." (2:36)
• "As per the details here, the video generation process took around 2 minutes and 40 seconds on a 4060TI 16-gigabyte." (5:08)
• "This should produce videos easily on a system with smaller V-RAM, but you will need more system RAM though." (6:06)
• "So, it took 3 minutes 30 seconds on 4060 TI 16-gigabyte." (6:43)

Questions & Answers

Q: How do you install Wan2GP through Pinokio on Windows?

Download Pinokio for Windows and save the executable where you want the installation to occur. Run it, open Discover, select Wan2GP, and click the one-click installation button. If a requirement fails, retry it, then click Start after installation completes.

Q: Which LTX-2 model is used for faster generation in Wan2GP?

Select LTX-2 and choose the distilled model, which is described as producing a result in 8 steps. In the demonstrated workflow, sampling occurred twice, first with 8 steps and then with 3 steps.

Q: How much storage does the LTX-2 setup on Wan2GP require?

The completed Pinokio installation occupied 79 GB. Initial generation also triggered several downloads, including a distilled model of around 21 GB and a Gemma 3 model of around 13 GB.

Q: How long does LTX-2 video generation take on an RTX 4060 Ti with 16 GB VRAM?

The first 10-second generation took around 2 minutes and 40 seconds, excluding much of the model-download time. Another 10-second run took 3 minutes and 30 seconds, while the 30-second, 721-frame generation required more than 11 minutes.

Q: How do you generate a 30-second LTX-2 video on Wan2GP?

Set the frame count to 721 for a 30-second video. Because 721 exceeds the 481-frame sliding-window size, Wan2GP uses two windows: one for the first 20 seconds and another for the remaining 10 seconds.

Q: Why can consistency break in a 30-second LTX-2 video?

A single sliding window supports 481 frames, corresponding to 20 seconds in the demonstrated settings. The remaining 10 seconds are processed in another window, and the example lost visual consistency and repeated dialogue after the transition.

Q: Can Wan2GP run LTX-2 with low VRAM?

The demonstrated RTX 4060 Ti system used 16 GB of VRAM, and its VRAM usage was described as very low. However, system memory usage was high, so a smaller-VRAM system may still require more system RAM.

Q: What are the limitations of Wan2GP compared with ComfyUI for LTX-2?

Wan2GP provides a simpler interface but less workflow customization. The presenter could not remove the second sampling stage as in the simplified ComfyUI workflow. The Pinokio setup also lacked the latest Sage Attention or Triton acceleration options available through a manual standalone installation.

Summary & Key Takeaways

• Wan2GP can be installed through Pinokio by downloading its Windows executable, opening the Discover section, selecting Wan2GP, and running the one-click installer. The first attempt left one requirement missing, but retrying completed the setup. Starting Wan2GP then opened its web interface, where LTX-2 and the distilled model were selected.

• The distilled LTX-2 model uses an 8-step sampling stage followed by a 3-step stage. Initial setup downloaded numerous files, including a model of about 21 GB and Gemma 3 at about 13 GB. The complete installation occupied 79 GB, making storage capacity and download speed important practical considerations.

• Long videos use sliding windows when the requested frame count exceeds 481 frames. A 721-frame, 30-second generation was divided into a 20-second window and a remaining 10-second window. The second segment began from the first segment's final frame, but repeated dialogue and lost visual consistency at the transition point.