LTX 2.3 Image: Image-Only LTX Checkpoints Get a ComfyUI Path

LTX 2.3 has been a video-and-audio model so far, but it turns out the same backbone makes a capable still-image generator. A new pull request against ComfyUI by elismasilva adds a dedicated image-only model path (ltxv_image) for LTXV 2.3 checkpoints, and a companion Hugging Face repo ships four unified safetensors checkpoints pruned for text-to-image.

LTX 2.3: the audio-video foundation model the image-only checkpoints are pruned from

What the PR adds

The pull request routes image-only LTXV 2.3 checkpoints through a new LTXVImage / LTXVImageModel pair, selected only when checkpoint metadata marks the model as image_model="ltxv_image". The existing LTX video path is untouched and keeps instantiating the current LTXVModel. Beyond the loader, the PR also brings:

• Optional gated attention in LTXV transformer blocks (off by default).
• Per-block STG perturbation routing for LTXV self-attention.
• LTXVGuidanceRescale, a new node implementing Diffusers-style guidance rescale.
• Two blueprints: Text to Image (LTX-2.3 Image Dev) and Text to Image (LTX-2.3 Image Distilled), registered under Image generation and editing/Text to image.

The PR is open as of August 28, 2026 and has not merged yet. Until the ltxv_image path lands upstream, the checkpoints can be tested on the author's ComfyUI branch.

The checkpoints

The four files in elismasilva/ltx2.3-image-comfyui are derived from Lightricks' LTX 2.3 family, pruned to remove the video and audio execution paths. No new training is involved: the weights are converted into unified ComfyUI checkpoints and dropped into models/checkpoints/.

The 13b name refers to the effective size after pruning from the original larger LTX 2.3 package. The int8 convrot builds trade some fidelity for lower VRAM use.

Suggested settings

The model card gives starting points for each variant:

Both blueprints in the PR expose prompt, resolution, steps, guidance, and the PAG / guidance-rescale controls, so the settings above map directly onto the shipped workflows.

Why an image path for a video model

Lightricks' LTX-2.3 is a joint audio-video foundation model, and image-only pruning is the same direction other video families have taken: strip the temporal and audio weights, keep the spatial prior, and you get a fast text-to-image checkpoint that inherits the base model's prompt adherence. For ComfyUI users already running LTX 2.3 for video, the image checkpoints reuse the same text encoder and loading conventions, so trying text-to-image costs little more than a checkpoint download. If the PR merges, LTX 2.3 would cover both stills and video inside core ComfyUI with no extra node packs.