LTX-2 Inject Reference Into Start Frames ??

Here's the complaint this node exists to fix, straight from the source code comment: in LTX extension workflows, the model mostly follows the conditioning frames - the images input - and feeding a reference image into an auxiliary or empty-latent slot often has little or no effect on identity. So the "reference" wasn't really being applied. This node fixes that by pushing the reference where the model actually looks: into the start frames themselves.

It's the step-up from IAMCCS_LTX2_ReferenceImageSwitch. The switch picks which image flows; this one splices your reference into the actual conditioning frames of a segment. If you've ever wired a character still into a secondary image slot, watched the character drift anyway, and assumed the model was broken - this is the tool for that exact frustration.

How it works

You feed it the start_images (the conditioning frames, e.g. start_images from the IAMCCS ExtensionModule), a reference_image, and it overwrites or blends the reference into the last K (or first K) frames of that batch.

The controls that matter:

• mode - none (passthrough), inject (replace frames outright, strength treated as 1.0), or blend (mix with what's there via blend_strength).
• frames_to_inject - how many of the conditioning frames get modified (default 1).
• ramp - when on (default), the strength gradually increases across the injected frames instead of slamming in at full force. That's the difference between a hard cut and a smooth lead-in for the model.
• position - tail (last K frames, default) or head (first K frames).

The plumbing is automatic: a batch-1 reference is repeated to match the batch, and it's resized to the start frames' resolution before injection.

Inputs and outputs

• start_images - the conditioning frames to modify.
• mode, blend_strength, frames_to_inject, ramp, position - the injection controls above.
• reference_image - optional; batch 1 is fine.

Outputs: start_images (the modified conditioning batch) and a report STRING. Same names, same type, so it drops into an existing graph without rewiring - you just insert it between wherever start images come from and wherever they go.

Install

ComfyUI Manager → "IAMCCS", or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Restart. Pure image math, no model files.

Gotchas

Start at frames_to_inject: 1 and ramp: true. Injecting a full-strength reference into too many frames at once can over-anchor the segment and freeze the motion - the exact "static I2V" problem LTX users already fight. The tail position is the right default for extension work because it's the region the model is trying to continue from; flip to head only when you're conditioning the start of a fresh clip. And if you see no effect at all, check that mode isn't none - it's the default, deliberately, to keep old workflows untouched.