wanx-troopers

LTXV

Capabilities

Our APIs support up to 20s at 25fps at the moment. Depending on hardware of course.

Setup

• Main model HF:Lightricks/LTX-2
  • ltx-2-19b-distilled
  • ltx-2-19b-dev-fp8
  • ltx-2-19b-distilled-lora-384
• Text Encoder
  • HF:google/gemma-3-12b-it-qat-q4_0-unquantized for custom node, see bellow
  • HF:Comfy-Org/ltx-2:split_files/text_encoders gemma_3_12B_it adapted by team Comfy
  • HF:unsloth/gemma-3-12b-it-GGUF gemma-3-12b-it-Q4_K_S.gguf possibly not suitable?
• Alternatively HF:GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn
  • gemma_3_12B_it_fp8_e4m3fn - The fp8 converted text encoder from comfy, goes in CLIP folder
  • ltx-2-19b-dev-fp4_projections_only - Extracted projections from LTX-2 model to allow loading with DualClipLoader node, goes in CLIP folder
  • ltx-2-19b-dev-fp4_video_vae - The video vae, can be loaded with VaeLoader node, goes in VAE folder
  • ltx-2-19b-dev-fp4_vocoder - The vocoder model, not useful separately currently
• If using Comfy native nodes and workflows
  • code and template workflows merged to ComfyUI main branch (nightly)
  • Comfy native workflows do not require custom nodes from Lightricks/ComfyUI-LTXVideo
• If using LTX developers’ nodes and workflows
  • GH:Lightricks/ComfyUI-LTXVideo nodes
  • GH:Lightricks/ComfyUI-LTXVideo:example_workflows
  • docs.ltx.video/open-source-model/integration-tools/comfy-ui
  • still need Comfy nightly (likely)

does the model go in diffusion_models folder? no, in checkpoints folder

what folder does the spatial upscaler go in models/latent_upscale_models

From LTX developers:

If you are using the custom node (from the LTXVIDEO repository) you must download the entire model folder from Hugging Face: HF:google/gemma-3-12b-it-qat-q4_0-unquantized after downloading, place the folder in your models directory and configure the ComfyUI node to point to: gemma-3-12b-it-qat-q4_0-unquantized/model-00001-of-00005.safetensors Do not delete or move the other model-0000X-of-00005.safetensors files.

Advice

euler simple or ddpm 2 karras are safe combos

Using the full model with the distill lora at 0.6 makes skin far more natural than just using the distill model. Shame it’s so RAM heavy.

Tools For LTX-2

• HF:Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-Left

Outdated Tools For LTX 0.9

• HF:Lightricks/LTX-Video-ICLoRA-depth-13b-0.9.7
• HF:Lightricks/LTX-Video-ICLoRA-detailer-13b-0.9.8
• Simple WAN refiner workflow which can be used after LTX main run and upscaler: Experiment_upscaler_wrapper.json

This site is open source. Improve this page.