IAMCCS Cine Info H3 V2V - Source + Pictures

If the H3 V2V route is a pipeline, IAMCCS_CineInfoH3V2V is the intake desk. You plug in one source video plus up to four reference pictures, and this node attaches them to your cine_linx bus with all the instructions about how that source should be used - which slice of it, fitted how, guided by what, paired with which audio. It's deliberately an attach node: nothing is encoded or generated here, but every downstream decision in the V2V backend inherits the choices you make on this node, so it's the cheapest place in the whole graph to get things wrong.

The required source_video is the footage you're restyling, and source_fps (24 default) tells the pipeline how to interpret its timing. Then the interesting stuff:

• guide_mode - from_shotboard, raw_only, raw_pose, raw_depth, or raw_depth_pose. This decides what structural guides the V2V pass gets. Set a pose/depth mode and the node expects you to wire the corresponding pose_guide / depth_guide inputs, whose tooltips are refreshingly blunt: preprocess upstream with DWPose for pose and Depth Anything V2 for depth. The node won't run those for you - that's the price of keeping it dependency-light.
• source_range_policy - timeline_segment (use the shotboard's trim), sequential_requested (walk through the source sequentially), or repeat_from_offset. With source_offset_seconds this is how you point H3 at the right part of a long clip.
• source_fit - native_adapt, canvas_pad, canvas_crop, stretch. Same crop-vs-letterbox debate as every V2V tool, same advice: native_adapt if you want the model to decide, explicit crop/pad if you're matching a locked canvas.
• audio_pairing - from_shotboard, pair_with_source_video, standalone_reference, off. Wire source_audio and decide whether the source's audio is a generation input or just a reference.

The reference block mirrors CineInfoH3 but with V2V-specific roles: subject_identity, wardrobe_object, environment, style (or disabled) per slot. Notice there's no keyframe here - in V2V the source video itself carries the temporal structure, so the stills are about identity and look, not hitting beat numbers.

The two optional guide inputs (depth_guide, pose_guide) are the ones that trip people up: they're only consumed when the active guide mode demands them, but the tooltips make clear the preprocessing is on you. There's no hidden automatic DWPose - that's why pose_mode and dwpose_enabled live over on the Shotboard planner side of this family instead.

Output is a single cine_linx bus carrying the source, its audio pairing, and every one of these decisions, which the V2V backend routers and conditioning stage then read per segment. Nothing decodes, nothing generates - if you queue this node alone, it just re-emits the bus and you'll wonder what happened. That's normal.

Installation is pack-standard (search IAMCCS in ComfyUI Manager, or clone the repo into custom_nodes, restart), and it inherits the H3 family's real requirements: a current ComfyUI with native MiniMax H3 AV conditioning, the H3 video/audio VAEs, and - only when you actually use pose/depth guide modes - a DWPose setup and Depth Anything V2 somewhere upstream. The pack's H3 workflow doc is the dependency matrix to read before assembling the full graph.

The recurring failure mode: selecting raw_depth_pose without wiring depth_guide or pose_guide, then wondering why the output ignores your structure entirely. The guide modes are promises you make - the node takes your word for it and the backend just passes your (possibly missing) guides through.