Many powerful video models are available through APIs or closed platforms. While these services make generation convenient, they can also introduce limitations such as per-generation costs, usage restrictions, dependency on external infrastructure, and limited ability to customize the underlying model.
Generating a single impressive video clip is one thing. LTX-2.5 is an open-weight model designed for local execution and fine-tuning. Its primary capability is generating synchronized, high-fidelity video and audio from text, image, and video inputs.
Installation
1. Make sure you have the latest version of ComfyUI installed on your system. If you haven't updated it yet, simply update ComfyUI from the Manager.
You need to share your information before downloading these models.
2. Download the required LTX-2.5 diffusion model from the LTX-2.5 Hugging Face repository.
(a) ltx-2.5-22b-dev-transformer-comfy-int8-convrot.safetensors
(b) ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors
(c) ltx-2.5-22b-dev-transformer-bf16.safetensors
(d) ltx-2.5-22b-distilled-transformer-bf16.safetensors
(e) ltx-2.5-22b-distilled-transformer-nvfp4.safetensors
Save the following model into your ComfyUI/models/diffusion_models folder.
3. Download the required text encoders from the LTX-2.5 text encoders repository. You need both of the following files:
(a) gemma4_e2b_it_bf16.safetensors
(b) gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors
Save both files into your ComfyUI/models/text_encoders folder.
4. Download the required VAE models from the LTX-2.5 VAE repository. Get both VAEs files:
(a) ltx-2.5-video-vae-bf16.safetensors
(b) ltx-2.5-audio-vae-bf16.safetensors
Save both of them into your ComfyUI/models/vae folder.
6. Finally, download the LTX-2.5 latent spatial/temporal upscaler from the LTX-2.5 latent upscaler repository. These are optional models to do the video upscaling.
(a) ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors
(b)ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors
Save it into your ComfyUI/models/latent_upscale_models folder.
7. This is optional. Download distilled lora model bf16 (ltx-2.5-22b-distilled-lora-450-bf16.safetensors and save this into ComfyUI/models/loras folder.
7. Once all the models are in their respective folders, restart or refresh ComfyUI.
Workflow
1. Download the LTX2.5 workflows from our Hugging face repository:
(a) LTX-2.5_i2v.json (Image to Video workflow)
(b) LTX-2.5_t2v.json (Text to Video workflow)
(c) LTX-2.5_flf2v.json (first frame-last frame to Video workflow)
2. Drag and drop the workflow into ComfyUI. You can follow official prompting guide for LTX 2.5 to get the better results.
(a) Image to Video workflow
-Load image into Load first frame node.
-Put the prompt into prompt box. Use detailed prompting(like-motion with the holding character and audio sync ) to get the best out of it. You can also use Prompt Enhancer- expands short prompts into rich cinematic instructions,but it will take longer to generation video.
-Add and do settings for duration/width/height/frame_rate(defaults work good)
-These are multiples of 32 gets auto managed. Supported resolutions:
0.2mp -608 x 352
0.3mp-736 x 416
0.4mp-864 x 480
0.5mp-960 x 544
0.6mp-1056 x 608
0.7mp-1152 x 640
0.8mp-1216 x 672
0.9mp-1280 x 736
0.98mp-1344 x 768
1.0mp-1376 x 768
1.2mp-1504 x 832
1.5mp-1664 x 928
1.8mp-1824 x 1024
2.0mp-1920 x 1088
-Finally hit run button to start generation.
(b) Text to Video workflow
-Put the prompt into prompt box. Use detailed prompting(like-motion with the holding character and audio sync ) to get the best out of it. Simply use Prompt Enhancer- converts short prompts into rich cinematic instructions but adds longer time to generate.
-Add and do settings for duration/resolutions/width/height/frame_rate(defaults work good)
-Finally hit run button to start generation.
(c) First frame-last frame to Video workflow
-Load first image into Load first frame and second image into last frame node respectively.
-Put the prompt into prompt box. Use detailed prompting(like-motion,scene,lighting, camera style shots etc) to get the best out of it. Enable/disable Prompt Enhancer- expands short prompts into rich cinematic instructions at near-zero extra compute
-Add and do settings for duration/width/height/frame_rate(defaults work good)
-Finally hit run button to start generation.
LTX-2.5 is interesting because it focuses on a problem that is becoming increasingly important in generative video: ownership and control. As video models become more capable, creators don't necessarily want to send every generation to an external service.
They may want to run models locally, fine-tune them for their own visual style, integrate them into custom applications, or keep sensitive creative assets inside their own infrastructure.
){kind=link}