LTX 2.3 basic GGUF 720p workflow — v1.0

Use the LTX 2.3 basic GGUF 720p workflow (v1.0), based on LTXV2, to create or extend video from text, images, audio, or video and produce saved video output; it supports T2V, S2V, V2V, and I2V with first-, last-, or middle-frame images.

The workflow requires ltx-2-19b-distilled.safetensors and ltx-2-19b-lora-camera-control-dolly-left.safetensors.

Voice-clone use takes a few seconds of audio and can crop those same seconds after processing; reference-image use takes a starting image plus an instruction for a different action, can crop the starting image, and keeps the character descriptions the same.

For video extension, provide images and audio extracted from a source video; the workflow extends the remaining length.

Before use, update the GGUF custom node and ComfyUI to the latest versions.

Suggestion · non vérifié

Place text encoder-related files in ComfyUI\models\text_encoders, audio VAE files in ComfyUI\models\checkpoints, and the upscale model in ComfyUI\models\latent_upscale_models.

Suggestion · non vérifié

Listed ComfyUI nodes include LTXVConditioning, LTXVImgToVideoInplace, LTXVPreprocess, LTXVAddGuide, LTXVCropGuides, LTXVSeparateAVLatent, LTXVConcatAVLatent, LTXVEmptyLatentAudio, EmptyLTXVLatentVideo, LoadImage, LoadAudio, CreateVideo, SaveVideo, VAELoader, VAEDecode, VAEDecodeT

Other listed ComfyUI nodes include DualCLIPLoaderGGUF, UnetLoaderGGUF, LoraLoaderModelOnly, KSampler, KSamplerSelect, SamplerCustomAdvanced, LTXVScheduler, LTXVLatentUpsampler, LatentUpscaleModelLoader, ImageScale, ImageScaleBy, CFGGuider, RandomNoise, TorchCompileModel, and the.

Listed model files include google_gemma-3-12b-it-qat-Q6_K.gguf, ltx-2-19b-distilled.safetensors, ltx-2.3-22b-distilled-Q6_K.gguf, LTX23_video_vae_bf16.safetensors, ltx-2-19b-ic-lora-pose-control.safetensors, ltx-2-19b-lora-camera-control-dolly-left.safetensors, and ltx-2-spatial-

For T2V, set bypass image on; for I2V, set bypass image off.

Suggestion · non vérifié

Use the distilled model with distilled-embedding, or the dev model and dev-embedding with distilled-lora.

Suggestion · non vérifié

For low-resolution output, you can bypass the upscale node and start with a lower length, perhaps 9.

Suggestion · non vérifié

Before use, verify the listed node names and model filenames against what is installed in your ComfyUI setup.

Suggestion · non vérifié

With a reference image, review whether the generated action matches your intent; the character descriptions remain the same.

Suggestion · non vérifié