Skip to content

Text to Video

Video generation pipelines built on LTX 2.5 and MiniMax H3 Turbo GGUF quants.

Setup

Video models use a frame-aware latent that carries a temporal dimension. The sampler steps over the full sequence rather than a single frame.

Load the diffusion model and text encoder with the GGUF loader nodes from comfyui-gguf-reboot. The loader nodes give the best performance.

Models and formats

ModelRepositoryRecommended formats
LTX 2.5molbal/LTX-2.5-GGUFQ8_0, Q8_CR, Q4_K_M
MiniMax H3 Turbomolbal/MiniMax-H3-Turbo-GGUFQ8_0, FP8, Q4_K_M

Loading instructions

  1. Download the GGUF files from Hugging Face.
  2. Load the diffusion model and text encoder with the GGUF loader nodes.
  3. Use a frame-aware latent with the desired frame count and resolution.
  4. Configure the sampler for the model's step count and schedule.
  5. Decode frames with the VAE and assemble them into a video.

VRAM with video

Video latents multiply memory per frame. On 16GB cards use Q8_0 or the U16G mixed build. See Quant Formats.

Showcase

Video examples are on molbal's Civitai profile.

GGUF quantization and open-source tools.