LTX-2.5 Video to Video (IC-LoRA) ⚡

This is the do-everything prep node for LTX-2.5: text-to-video, image-to-video, and - the headline feature - video-to-video via an IC-LoRA task adapter that loads inside the node. No external LoraLoaderModelOnly wiring, no "is the LoRA attached" boolean to guess at. You pick a task adapter from a dropdown, feed it your source clip, and the node handles the whole guide-injection mechanism. Beard removal, HDR grading, motion tracking, deblurring - any LTX IC-LoRA trained on whole-video reference conditioning - this is the node for it.

The mechanism is genuinely different from the i2v first-frame hold. 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 (ComfyUI core's keyframe_idxs RoPE mechanism), then cropped back off after sampling. The official examples wire that by hand across ~10 nodes; this node collapses all of it into one prep step. Because those guide tokens are appended, the family's Crop Video Guide node strips them off after the distilled pass - and before the latent upscale, because the upscaler can't tell guide frames from output frames and would double both.

The inputs that matter

The mode dropdown declares the base behavior and is validated against what's connected: t2v (nothing connected), i2v (images required), v2v (video required). Then the interesting set:

• video - your source clip. When connected, length and frame_rate are taken from the clip rather than the widgets.
• ic_lora - the task adapter, straight from your models/loras folder. Loaded onto the model here at ic_lora_strength. Set to none and a connected video is used only for length/frame-rate/audio - useful for A/B-ing whether the adapter is doing anything.
• guide_strength - how hard the reference is held; 1.0 is the official default.
• keep_original_audio (default on) - keeps the source clip's own audio unchanged in the output. Off means the model generates new audio from scratch.
• reference_audio - drive generation from a voice or sound clip with no source video; length_from_audio sizes the video to it.
• latent_downscale_factor - ignore it unless your IC-LoRA's model card says it was trained on a downscaled reference grid. Most use 1.0.
• img_compression - the 2.5-specific widget, the H.264 CRF round-trip on i2v holds (official 18).

Outputs: model, positive, negative, latent, frame_rate. Chain: this node → KSampler (distilled) → Crop Video Guide → Latent Upscale x2 → KSampler (refine) → AV Decode.

Installing it

Part of the ComfyUI-GGUF-Loader pack (ChrisColeTech's fork of city96's ComfyUI-GGUF). ComfyUI Manager: search "ComfyUI-GGUF-Loader". Or:

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

Restart. The IC-LoRA weights themselves go in models/loras.

Common issues

Two worth flagging. First, the on-disk IC-LoRA adapters are 2.3-trained - they load cleanly onto 2.5 (same key layout), but cross-version quality is whatever the weights give you; don't be shocked if a 2.3-tuned adapter behaves differently here. Second, remember the crop: feed the distilled output straight into the upscaler and the guide frames get doubled into the final resolution. The node's own recipe order exists for a reason.