LTXVAddGuideMultiFc

LTX-2's secret weapon is guide frames: drop images at specific positions along the timeline and the model treats them like storyboard panels - this happens at frame 30, this at frame 60 - instead of you hoping the text prompt carries the whole clip. The capability has been buried in ComfyUI's core AddGuide node for ages; the community built wrappers around it precisely because wiring up a handful of guides by hand is miserable. LTXVAddGuideMultiFc is one of those wrappers, and it's the one this pack is built around: up to 20 guide slots in a single node, each with its own frame index and strength, appearing as a dynamic combo.

If you've seen Kijai's LTXVAddGuideMulti from KJNodes, this will feel like home - it even registers under the KJNodes/ltxv category so it sits right next to it in the menu. The difference is this one uses ComfyUI's new DynamicCombo API: you set num_guides to how many images you want and the node grows the matching image_N, frame_idx_N, strength_N, preprocess_N, and preprocess_crf_N slots. That's the input list that matters - one image, the frame to pin it to, and how hard to pin it.

How it works

Feed it your text conditioning (positive, negative), your video latent, and the LTX-2 vae. For each guide it resizes the image (that's the upscale_method combo - nearest-exact, bilinear, lanczos), runs it through ComfyUI's built-in LTXVPreprocess by default, encodes it with the video VAE, appends the encoded latent to the end of your video latent, and writes the guide's frame position into the conditioning via keyframe_idxs. The strength slider controls a noise mask - strength 1.0 pins the frame near-hard, lower lets the model drift from it.

Two settings trip people up. mask_mode of constant pins the guide across its whole span; ramp with ramp_frames expands a single-frame guide into a fade-in/out across N frames, which reads much more naturally for camera moves. And frame_idx can be negative - frames before 0 are "reference-like" context - with negative_frame_mode deciding how multi-frame guides land relative to the start (allow_cross_zero is the default; leave it).

Outputs: positive, negative, and the extended latent (samples + noise mask) go straight to your sampler, and processed_images shows what each guide looked like after preprocessing. The info string is the one to read when something's off - it lists the exact frame_idx and latent/frame id mapping for every guide.

Wiring and the gotchas

This node has to run on the video-only latent, before you merge audio with LTXVConcatAVLatent - it refuses to touch a combined AV latent (128 channels). The repo's example workflow shows the shape: EmptyLTXVLatentVideo → this node → LTXVConditioning / LTXVConcatAVLatent → sampler.

Two errors you'll actually hit. "Conditioning frames exceed the length of the latent sequence" means a frame_idx is past the end of your video - the node checks. And blurry results with overprocessed-looking guides is the classic CRF complaint: the default preprocess_crf is 33, and the community sweet spot for LTX is 30–40, dropping to 30–35 when things get soft. Conversely, if you're fighting LTX-2's notorious frozen-I2V problem, turning the CRF up is one of the levers people report actually helping.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/fangcun010/ComfyUI-LTX2-R2V

Restart, then it's under KJNodes/ltxv. Because it uses DynamicCombo, it wants ComfyUI 0.8.1 and frontend 1.33.4 or later - if the node shows up but the guide slots don't appear, update ComfyUI, that's the whole fix. No extra dependencies and no model downloads; it uses the LTX-2 checkpoint and VAE you already load.