🅛🅣🅧 LTXV Add Guide Advanced Attention

This is how you tell LTX "at this exact frame, the video should look like this image." It's keyframe conditioning - you drop a reference frame in at a chosen position and the generation is steered toward it. The "Advanced Attention" part is the twist: on top of the usual strength control, you also get a knob for how much the model's attention leans on that guide, plus an optional spatial mask so the guidance only applies to part of the frame.

How it works

Per the node's own description, it does the same preprocessing as the plain Add Guide Advanced node - encoding your image into the latent space at the frame you specify - and then adds attention_strength and an optional attention_mask on top. So you're conditioning in two ways at once: strength biases the denoised result toward the guide, and attention_strength controls how strongly the model attends to it during generation. That second lever is what lets you have a firm anchor image without the video freezing on it, which is exactly the failure mode LTX's I2V was notorious for early on.

The inputs that matter

• image and frame_idx (default 0) - the guide frame and where it lands. Negative indices count from the end, so you can pin a last frame as easily as a first.
• strength (default 1) - how hard the output is pulled toward the guide. Drop it if the guided frame dominates and kills motion.
• attention_strength (default 1) - the attention lever. Lower it for a looser influence that anchors composition without locking every detail.
• crf (default 29) - simulated compression on the guide, 0–51. LTX was trained expecting compressed input, and a crf around 29–40 is a known sweet spot; too clean an image can actually hurt.

It also takes positive/negative conditioning, the vae, a latent, and image-fit options (blur_radius, interpolation, crop). The optional attention_mask restricts the guidance spatially. Outputs are updated positive, negative, and latent - chain them into your sampler.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

Needs an LTX checkpoint and VAE loaded; both pull on first use.

Common issues

The guided frame freezes the clip. Classic over-conditioning. Bring down strength, and if it's still stiff, back off attention_strength too - that's the combination the advanced node exists to give you.

The reference looks too crisp and the result is worse. Counterintuitive but real for LTX: nudge crf up toward the 30s. The model expects some compression.

Frame index does nothing / errors. Make sure frame_idx is inside your clip's frame count, and remember LTX's grid rules - frame counts follow the divisible-by-8-plus-1 pattern on 2.3, so an index past the end just won't behave.