LTX-2 Load Latent Bridge ??

When a segmented LTX-2 pipeline hands off between segments, the new segment's latent starts out as noise being denoised toward the next chunk of video. Without help, that noise has no memory of the previous segment - so the model reinvents the subject's face, the lighting, everything. This node gives the new segment that memory: it loads the previous segment's latent tail from the latent bridge (IAMCCS_LTX2_SaveLatentBridge wrote it) and splices it into the leading frames of the current latent, then locks those frames with a noise mask so the sampler starts from the previous state instead of from scratch.

That's the whole trick of latent-overlap continuity, and it's why this node exists as the counterpart to the save half. The PNG last-frame bridge re-anchors an image; this one preserves actual denoised latent state.

How it works

Given the current segment's latents, segment_index, render_id, and a temporal_overlap (pixel frames → latent frames via the VAE), it loads ltx2_latent_bridge_<render_id>.safetensors and:

1. Copies the previous segment's tail into the first overlap_f latent frames of the current latent.
2. Sets the noise mask on those frames to 1.0 - temporal_overlap_cond_strength (default 0.6, so the overlap is 40% locked) - the sampler denoises the overlap but is strongly pulled toward the previous state.
3. Stashes the previous full latent and reference latent into the latent dict under iamccs_prev_latents / iamccs_reference_latents, plus the recorded iamccs_seed_offset. The IAMCCS samplers (IAMCCS_LTX2_InitLatentSampler, etc.) read these keys to continue the noise walk and normalize against the reference.

Segment 0, an empty render_id, or temporal_overlap of 0 all short-circuit to a passthrough - there's nothing to load on the first segment.

Inputs and outputs

• latents - the current segment's LATENT, which gets the tail spliced in.
• segment_index - 0 disables loading.
• render_id - must match the Save side.
• temporal_overlap - how much of the previous tail to condition on (0 disables).
• temporal_overlap_cond_strength - how hard the overlap is held (default 0.6).
• vae - optional, for the pixel→latent conversion.

Outputs: conditioned_latents (LATENT, ready for the sampler) and report.

Install

ComfyUI Manager → "IAMCCS", or:

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

Restart. No extra dependencies; reads the same bridge directory the Save node writes.

Gotchas

This node hands you a conditioned latent - but the sampler downstream has to honor it. The cleanest pairing is with the IAMCCS samplers that understand the iamccs_* keys; a generic sampler that ignores the noise mask will denoise the overlap straight away and your continuity quietly evaporates. Also, temporal_overlap: 0 disables the load (a common "why isn't anything happening" moment), and a missing bridge file doesn't crash - it reports latent_bridge=missing and passes through. Which is polite, but it means a failed save upstream looks like a silent continuity loss rather than an error. When segments drift apart, check the bridge directory first.