FL LTXV Masked IC-LoRA Guide

If you're working with LTXV's in-context LoRAs - the adapters that condition video generation on a reference image or control video rather than just a text prompt (Lightricks ships official ones for pose, depth, canny, plus a growing community set for things like colorization and inpainting) - you eventually hit the same wall: you want to hand the model a guide image at a specific point in time, and you want a mask to control exactly how much that guide influences the result, and where. That's what this node is for. It's flagged Work-in-Progress by the author, straight in the README's WIP section, so treat it as an experimental building block rather than a finished, battle-tested tool.

How it works

LTXV latents run on their own temporal cadence - you can't just drop a guide frame in at an arbitrary index, it has to snap to the model's actual latent frame boundaries. This node handles that snapping: give it a frame_idx and it aligns your image (plus optional mask) to the right spot in the latent sequence, then writes it in using LTXV's own _INPAINT_GREEN marker convention - a specific placeholder color the model's inpainting path recognizes as "generate here, guided by what's around it." From there, three separate strength dials control how the guide influences generation: guide_strength for the overall conditioning weight, denoise_strength for how much the guided region gets regenerated versus preserved, and attention_strength for how hard the guide pulls on the model's attention specifically. mask_mode decides whether your mask gates the latent noise, the attention, or both - which matters because those are genuinely different effects: gating noise controls where pixels get regenerated, gating attention controls where the model looks when deciding what to generate.

The inputs and outputs that matter

• image, mask (optional), frame_idx - the guide content, its region of effect, and where in the video's timeline it lands.
• guide_strength, denoise_strength, attention_strength (all 0–1) - the three levers described above. Start conservative; these stack, so cranking all three at once tends to overpower whatever else is driving the generation.
• mask_mode (both / latent_noise / attention) - pick based on whether you want spatial control over pixels, over the model's focus, or both at once.
• inpaint_preprocess and write_source_to_latent - both default true; they control whether the node does the green-marker inpaint prep and whether your source image gets baked into the starting latent before sampling.
• Outputs positive, negative, latent feed straight into your sampler - this node reshapes conditioning and the starting latent, it doesn't sample itself.

How to install it

Via ComfyUI Manager: search "ComfyUI_Fill-Nodes", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI_Fill-Nodes

You'll also need an LTXV setup already working - this node assumes an LTXV model, VAE, and conditioning are already wired up; it's an add-on to an existing LTXV graph, not a starting point.

Common issues & troubleshooting

It's WIP - expect rough edges. The author put this in the README's Work-in-Progress section for a reason. If behavior seems inconsistent between versions of the pack, that's the expected state of an actively-changing experimental node, not necessarily user error.

Guide has no visible effect. Check mask_mode against what you actually want - if you set it to attention only but expected pixel-level control, that's a mismatch, not a broken node. Also verify frame_idx actually lands inside your latent's frame range; an out-of-range index against LTXV's cadence can silently clamp to an edge frame.

Tiled encode settings (use_tiled_encode, tile_size, tile_overlap) exist for VRAM reasons. If you're guiding at high resolution and hitting out-of-memory on the VAE encode step, that's the pair of knobs to reach for before giving up on resolution entirely.