elismasilva/ltx2.3-image-comfyui

LTX 2.3 Image for ComfyUI

This repository contains image-only LTX 2.3 checkpoints packaged as unified ComfyUI safetensors files.

The checkpoints are derived from the LTX 2.3 model family by Lightricks and are pruned for still-image generation. The original video and audio execution paths are not included in these image-only variants. The weights remain derived from LTX 2.3 and inherit the original LTX Video 2 Open Source License.

Files

File	Description
ltx-2.3-13b-image-dev.safetensors	Image-only Dev checkpoint in bf16 precision.
ltx-2.3-13b-image-distilled.safetensors	Image-only Distilled checkpoint in bf16 precision.
ltx-2.3-13b-image-dev-convrot-int8.safetensors	Image-only Dev checkpoint quantized with convrot int8.
ltx-2.3-13b-image-distilled-convrot-int8.safetensors	Image-only Distilled checkpoint quantized with convrot int8.

What Changed

• Converted into unified ComfyUI checkpoint files.
• Pruned to keep the image generation path only.
• Removed weights that are only required by the original video/audio execution paths.
• Preserved the shared checkpoint components needed for ComfyUI loading, including the image transformer path and supporting model components.
• Not a new training run and not a dataset modification.

Variants

Dev

Use the Dev checkpoint when prompt flexibility and quality are more important than speed. This variant is intended to be used with classifier-free guidance and more sampling steps.

Recommended starting points:

• steps: 20 to 40
• cfg: 3.0 to 5.0
• pag_scale: optional, commonly around 0.8
• guidance_rescale: optional, commonly around 0.7

Distilled

Use the Distilled checkpoint for faster generation with fewer sampling steps. This variant is intended for low-step CFG-free style inference.

Recommended starting points:

• steps: 8
• cfg: 1.0
• pag_scale: optional, commonly around 0.8
• guidance_rescale: optional, commonly around 0.7

Convrot int8

The convrot int8 files are quantized variants intended to reduce memory use. They may trade some output fidelity for lower VRAM requirements.

ComfyUI Usage

Temporary ComfyUI Branch

These image-only checkpoints require ComfyUI support for the ltxv_image model path. Until this support is available in the upstream ComfyUI release, test them with this branch:

https://github.com/elismasilva/ComfyUI/tree/ltxv-image-only

After installing or switching to that branch, copy one or more .safetensors files into your ComfyUI checkpoints folder:

ComfyUI/models/checkpoints/

Then restart ComfyUI or refresh the model list.

These checkpoints are intended for LTX 2.3 image-only ComfyUI workflows that load the model through CheckpointLoaderSimple and route it through the LTXV image model path.

Suggested blueprints:

• Text to Image (LTX-2.3 Image Dev)
• Text to Image (LTX-2.3 Image Distilled)

The image-only checkpoints use metadata to identify the pruned image model path. ComfyUI support for this model path must be present in your ComfyUI build or branch.

Notes

• These files are for ComfyUI usage as unified checkpoint files.
• The checkpoints are unified safetensors files and should be placed in the ComfyUI checkpoints directory.
• Image-only pruning means video and audio generation are intentionally unavailable from these files.
• The 13b name refers to the effective image-only checkpoint after pruning from the original larger LTX 2.3 model package.

License

These weights are derived from Lightricks/LTX-2.3 and are released under the LTX Video 2 Open Source License:

https://huggingface.co/Lightricks/LTX-2.3/blob/main/LICENSE

Please review the original LTX 2.3 model card and license before redistribution or commercial use.