V2.6 LTX-2.3 DEV & Distilled Video with Audio + LTX Director workflow
Minor update to LTX2.3 Image to Video with Ollama workflow:
sampling Preview implemented
NAG to allow negative prompt for Distilled model (CFG=1)
Included a Workflow supporting LTX Director V2.0 node, which is awesome, it supports:
First, Mid, Last or whatever frame, basically any frame is a keyframe
Prompt Relay which allows to generate frames prior to Input Image
more precise editing: What shall happen when in the clip?
Audio Import to create lipsynced clips
Image or Text to Video, even both in one process
more Info: https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI
=> added a longer example video at the end.
V2.5 LTX-2.3 DEV & Distilled Video with Audio
Image to Video and a Text to Video workflow, both can use own Prompts or Ollama generated/enhanced prompts.
works with latest LTX 2.3 Distilled model (8 steps, CFG=1) or Dev model (20 steps, CFG=3)
Updated the processing for DISTILLED and DEV model, select the DIST or DEV model in loader node and switch to dedicated DIST or DEV processing pipeline, so each model has its own processing.
DIST model pipeline: Standard Guider and Basic Scheduler, follows the manual sigmas issued by Lightricks
DEV model pipeline: MultiModal Guider and LTX Scheduler + Distilled Lora on latent upscaler
Included a workflow version with "RTX Video Super Resolution" node, which upscales videos in highspeed.
Tip: With latest Comfy and LTX updates, the processing got faster for me, so I can increase the scale_by in sampler node from 0.5 to 0.6 or higher to have crisper videos with minor impact on render time.
V2.3 LTX-2.3 DEV & Distilled Video with Audio
Downloads for LTX 2.3:
update : April 14th 2026 : Lightricks has updated their LTX 2.3 distilled model to 1.1 (and Lora):
Models Dev & Dist (1.1 fp8 scaled by Kijai or better use the newer ...int8_convrot version): https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/diffusion_models
dist. Lora 1.1 : https://huggingface.co/Lightricks/LTX-2.3/tree/main
Smaller, alternative Desitilled Lora by Kijai (1.1): https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras
LTX-2.3 Distilled & Dev Models (fp8_scaled)Outdated, replaced by above 1.1.:https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/diffusion_modelsTextencoder1: (fp8_e4m3fn, same as LTX-2): https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/tree/main
Textencoder2: (projection_bf16): https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/text_encoders
Video & Audio Vae: https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/vae
Preview VAE (taeltx2_3): https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/vae
Loras:
Spartial upscaler (x2-1.1): https://huggingface.co/Lightricks/LTX-2.3/tree/main
Detailer Lora (same as LTX-2): https://huggingface.co/Lightricks/LTX-2-19b-IC-LoRA-Detailer/tree/main
Ollama Model (prompt only, fast): https://ollama.com/mirage335/Llama-3-NeuralDaredevil-8B-abliterated-virtuoso
alternative model with Vision (reads input image+prompt, slower): https://ollama.com/huihui_ai/qwen3-vl-abliterated
other model with Vision (great for I2V): https://ollama.com/huihui_ai/qwen3.5-abliterated
smaller LTX 2.3 GGUF Dev or Dist. models work as well. (replace Checkpoint loader node with Unet loader node from this custom node: https://github.com/city96/ComfyUI-GGUF ):
models: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main
save to models/unet/
V1.5 LTX-2 DEV Video with Audio including latest 🅛🅣🅧 Multimodal Guider
Image to Video and a Text to Video workflow, both can use own Prompts or Ollama generated/enhanced prompts.
Replaced the Guider node with latest Multimodal Guider node, see more details in WF notes or here: https://ltx.io/model/model-blog/ltx-2-better-control-for-real-workflows Before we had 1 CFG parameter for audio and video. With multimodal guider, we now can tweak audio and video seperately with even more parameters...
added a Power Lora Loader node to inject further Loras
use Image to Video Adapter Lora to improve motion for I2V: https://huggingface.co/MachineDelusions/LTX-2_Image2Video_Adapter_LoRa/tree/main
replaced a node to no longer require comfymath custom nodes
V1.0 LTX-2 DEV Video with Audio:
Image to Video and a Text to Video workflow with own Prompts or Ollama generated/enhanced prompts.
setup for the LTX2 Dev model.
uses Detailer Lora for better quality and LTX tiled VAE to avoid OOM and visual grids
2 pass rendering (motion+upscale). Upscale process uses distilled and spatial upscale Lora
setup with latest LTXVNormalizingSampler to increase video & audio quality.
Text to Video can use dynamic prompts with wildcards.
Download LTX-2 Files: (Workflow V1.0 and V1.5 only Outdated, replaced by LTX 2.3 above)
Find Model/Lora Loader nodes within Sampler Subgraph node.
- LTX2 Dev Model (dev_Fp8): https://huggingface.co/Lightricks/LTX-2/tree/main
- Detailer Lora: https://huggingface.co/Lightricks/LTX-2-19b-IC-LoRA-Detailer/tree/main
- Distilled (lora-384) & Spatial upscaler Lora: https://huggingface.co/Lightricks/LTX-2/tree/main
- VAE (already included in above dev_FP8 model, but needed if you go for GGUF models): https://huggingface.co/Lightricks/LTX-2/tree/main/vae
- Textencoder (fp8_e4m3fn): https://huggingface.co/GitMylo/LTX-2-comfy_gemma_fp8_e4m3fn/tree/main
- Image to Video Adapter Lora (more motion with I2V): https://huggingface.co/MachineDelusions/LTX-2_Image2Video_Adapter_LoRa/tree/main
Save Location:
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 checkpoints/
│ │ ├── ltx-2-19b-dev-fp8.safetensors
│ ├── 📂 text_encoders/
│ │ └── gemma_3_12B_it_fp8_e4m3fn.safetensors
│ ├── 📂 loras/
│ │ ├── ltx-2-19b-distilled-lora-384.safetensors
│ └── 📂 latent_upscale_models/
│ └── ltx-2-spatial-upscaler-x2-1.0.safetensors
│ └── 📂 Clip/
│ └── ltx-2.3_text_projection_bf16.safetensors
Custom Nodes used:
https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI (RTX VSR Version)
Text 2 Video only: https://github.com/KoinnAI/ComfyUI-DynPromptSimplified
LTX Director only: https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI
Ollama help:
Install Ollama from https://ollama.com/
download a model: Go to a model page, chose a model , then hit the copy button, i.e. https://ollama.com/huihui_ai/qwen3-vl-abliterated
open terminal and paste the model name, i.e.: ollama run huihui_ai/qwen3-vl-abliterated
model will be downloaded and can be selected in green comfy node "Ollama Connectivity". Hit "Reconnect" to refresh.
Example longer Video
Description
Minor update to LTX 2.3 I2V workflow with Ollama
Additional workflow with LTX Director (including LTX Director V2.0)
FAQ
Comments (7)
Dev and Distilled Pipeline, how does it work?
LTX 2.3 uses a 2 stage processing:
Stage 1 = Main process generates the motion. It scales down rendering by what you set in the scale_by parameter (default = 0.5)
Stage 2 = upscaling process with spatial upscaler in latent space will always upscale x2.
Render at half resolution, then spatial upscale x2 it, does the magic and is why LTX is considered "fast".
The upscaler in Stage 2 will always need to run with a distilled setup with 3 steps and CFG=1 (either by using the Dist model or the Dev model together with Dist. Lora, which is managed by the pipeline switch)
Stage 1 can use Dev (20steps, CFG=3) or Dist (8+steps, CFG=1) model. Another way is to use the Dev model together with the distilled Lora and dist. settings.
The PowerLora Loader (on the right) will be applied to both stages, main and upscaler.
Scenarios:
1. to have a Dist Pipeline (8+steps, CFG=1) there are 2 options:
a) Load the Dist model, set pipeline switch to Dist model (= default when you load the WF).
b) Load the Dev model, set pipeline switch to Dist model, load a distilled Lora in PowerLora Loader.
2. to have a Dev Pipeline (20steps, CFG=3): Load the Dev model, set pipeline to Dev model.
Dist models are faster, Dev models are more precise, better prompt following and have more variation, but take way longer to process.
The RTX Video Super Resolution will be applied at the end and scale the clip to a desired final resolution (1920x1080 as default), ensure to update these settings i.e when you do portrait format clip.
Not sure if this helps or is even more confusing ...:)
Does this make multiple clips and combine them allowing for longer videos?
LTX Director creates 1 clip combined of as how many Images with prompts or just prompts you throw at it. Like in the video example by the creator of the node.
The length depends on Vram avail.
LTX Director V2.0 is out, making V1 outdated, therefore WF had to be updated. If you previously downloaded this WF and get errors after updating to LTX Director V2.0, just re-download the WF, I have included a new one for V2.0.
Check this IC Edit Lora to Edit Clips by adding, removing or replacing items (edit_anything_v1.1...safetensor):
https://huggingface.co/Alissonerdx/EditAnything/tree/main
Load the Loras with the "LTX Director Guide" nodes for both stages.
Quenstion, the bass player in your examples , how did you make sure the strings do not deform, as im trying and trying but i get bad deformations on guitar strings, any help would be greatly appreciated.
not sure what setting prevented string distortion, My guess: Steps =11, detailer Lora from download link, OmniNFT Lora (TX-2.3-OmniNFT-RL-Lora_bf16.safetensors) & 30fps.
ComfyUi now supports the Int8_convrot quants for models, those are faster, have better accuracy and seem to use less Vram compared to i.e. fp8.
Check the newer uploaded int8_convrot model by Kijai: https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/diffusion_models
i.e. :
ltx-2.3-22b-distilled-1.1_transformer_only_int8_convrot.safetensors