IAMCCS LTX 2.5 BBox Backend

If IAMCCS_CineShotboardPlannerV3B is the screenwriter, this node is the director's first assistant. It takes the compiled shot plan off the cine_linx bus, turns every animated bounding box into LTX-native regional conditioning, applies it to your model and latent, and hands you back a fully wired-up model, positive, negative and latent - ready for a sampler. It's the heaviest node in the V3B pair and it orchestrates several LTX-native nodes under the hood, so when things go wrong they usually go wrong here.

How it works

The flow, straight from the source:

1. Pull the plan off the bus - width/height, total frames, fps, the bbox project, and the reference images.
2. If use_first_timeline_image is on and there are image paths, it respects the actual first clip on the timeline (not slot order - you can drag media around in the V3B editor without breaking which frame is the I2V anchor) and runs LTXVImgToVideoConditionOnly to seed the latent for image-to-video.
3. Compiles the bbox project into normalized keyframes and feeds LTXRegionalBBoxAnimator to draw the animated control frames.
4. Applies regional conditioning with LTXApplyRegionalConditioning, blending regional vs. global prompt influence.
5. Runs LTX's ICLoRA-style guide (LTXAddVideoICLoRAGuide) at control_strength so the bbox control actually steers generation.

Inputs that matter

Mostly model plumbing plus a few dials:

• cine_linx - the V3B planner's bus. Feed it anything else and it fails fast.
• model, clip, video_vae, video_latent, negative - standard LTX sampling inputs.
• regional_prompt_weight (0.85) and global_prompt_weight (0.15) - how much each per-object prompt vs. the global prompt matters. This is the artistic knob: drop regional to ~0.7 if the boxes are making things rigid.
• control_strength (1.0) - how hard the bbox control pushes.
• use_first_timeline_image (true) and i2v_strength (0.7) - whether/how the shot starts from a reference still.
• latent_downscale_factor (1), use_tiled_encode (false), tile_size / tile_overlap - VRAM management for the control-guide encode. Turn tiling on if you're close to OOM on long frames.

Outputs

• model, positive, negative, latent - feed these straight into your sampler.
• control_images - the actual bbox control frames (check these before rendering).
• regions - the LTX_REGIONS object the regional conditioning built.
• global_prompt, compiled_bbox_json, report - the prompt text, the compiled normalized keyframe JSON, and a status report.

Install

Part of IAMCCS-nodes (ComfyUI Manager → IAMCCS, or the git clone below). It delegates to ComfyUI's native LTX nodes, so you need LTX-2.5 support in a current ComfyUI plus an LTX model/CLIP/Video VAE on disk.

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

Where people get burned

• "Shotboard V3B BBox control is disabled." bbox_enabled is false on the planner. Flip it.
• "has no enabled BBox objects/keyframes." You enabled control but built no animated objects. The backend won't fake a render.
• Wrong cine_linx. This node only reads the V3B planner's bus; wiring it to an H3 shotplan bus fails with confusing downstream errors. Match planner to backend.
• Pushing i2v_strength too low with use_first_timeline_image on gives you a latent that ignores the anchor still - keep it ~0.7 and tune regional weights for structure instead.