LTX-2.3 Video to Video (IC-LoRA) ⚡

The official LTX-2.3 video-to-video workflow is a ~10-node tangle: ResizeImageMaskNode → GetImageSize → EmptyLTXVLatentVideo → LTXVConditioning → LTXAddVideoICLoRAGuide → VAEEncodeAudio → LTXVSetAudioRefTokens → LTXVConcatAVLatent, all hand-wired. This node is that entire subgraph collapsed into one prep step, with LoRA selection and EditAnything reference-conditioning built straight in. It's the highest-leverage node in the LTX-2.3 section of this pack.

LTX-2.3 is Lightricks' 22B joint audio-video model (March 2026) - the release the community finally called the best locally-run video model overall, not just the fastest. Its vid2vid mechanism is genuinely different from the img2img you know: the source video's frames get VAE-encoded and appended as extra reference tokens the model cross-attends to at the same timeline position as what it's generating (comfy-core's keyframe_idxs RoPE mechanism), then cropped back off after sampling. Not partially denoised, not held as a first frame. The IC-LoRA task adapter you select (beard removal, HDR grading, motion tracking - anything trained on whole-video reference conditioning) is what makes those guide tokens mean something.

The inputs that matter

• mode (required) - t2v, i2v, or v2v, validated against what's actually connected. i2v needs images; v2v needs video; t2v rejects either. images and video can still layer on top of each other - a still for identity/framing plus a clip for the transform.
• video + ic_lora - the vid2vid core. ic_lora loads the task adapter onto model inside this node (stock LoraLoaderModelOnly, no external wiring) at ic_lora_strength (default 1.0), then injects the video as guide tokens. With ic_lora="none", a connected video only supplies length/frame_rate/audio - useful for A/B-ing whether the adapter is doing anything. guide_strength (1.0) is how hard the reference is held.
• keep_original_audio (on) - keeps the source clip's own audio untouched; off lets the model generate new audio.
• latent_downscale_factor (1.0) - only for rare IC-LoRAs trained on a downscaled reference grid. Check the LoRA's card; every official example uses 1.0. Except - and this is a real trap - the union-control LoRAs with ref0.5 in the name want 2.0, and the width/height must be divisible by 64 so the half-res guide lands on whole latents.
• editanything_lora + editanything_module_path - the EditAnything subject-injection path (see the standalone patch node). Must be set together; mutually exclusive with ic_lora. With it active, images[0] becomes the reference photo to inject (plain/white background strongly recommended), and the prompt is an imperative ADD instruction: "Add a detailed appearance, placement in frame, relation to the scene." Sample at half your target resolution, then upscale and refine - at full res the source-video guide dominates and the add never appears.

Outputs are model, positive, negative, latent, frame_rate. The model comes out possibly patched - always take it from this node when a selector is set. Feed into LTX-2.3 KSampler (distilled), then LTX-2.3 Crop Video Guide (strips the appended frames) before LTX-2.3 AV Decode. Wire frame_rate into every decode fps.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart. You'll also want the LTX-2.3 A/V kit files (GGUF DiT or fp8 safetensors, Gemma-3-12B encoder, both VAEs) from the model links in the README, plus the IC-LoRAs you use (from Lightricks or DeepBeepMeep/LTX-2). Under 🤖 CCTech/LTX-2.3.

Gotchas

The frame math is non-negotiable: width/height divisible by 32, frame count divisible by 8 plus 1 (121 is the default for a reason). When video is connected, length and frame_rate are taken from it, overriding your widgets. And if a v2v output looks like the adapter isn't firing, first check ic_lora isn't none, then check latent_downscale_factor against the LoRA's card.