SCAIL-2 Scheduled Long Video (Internal SAM)

The other scheduler in this pack is the same engine with the mask-building step swallowed whole. SCAIL-2 Scheduled Long Video (Internal SAM) does everything the external-mask version does - same segment plan, same 81-frame chunk loop, same overlap handling, same outputs - but instead of requiring you to run SAM3 tracking and the SCAIL-2 Multi Reference Colored Mask upstream, it runs SAM3 tracking inside the node and builds the colored masks itself. Feed it pose_video, a sam_model, sam_conditioning, and your reference images, and it goes from raw inputs to a scheduled, chunked, multi-reference render in one call. It's the node you reach for when you want the simplest workflow and don't care about fiddling with tracks.

The trade is spelled out in the README and it's the real decision point: with the internal-SAM variant you cannot preview or adjust the SAM tracks before generation. SCAIL-2 generation is expensive - tens of minutes per clip on consumer cards - so a mask that's wrong costs you a full run before you find out. The external-mask scheduler keeps person selection visible and editable; this one is for when you've got a clean single-subject driving video and just want frames out.

The extra inputs over the external scheduler are the SAM controls, and they map to the native SCAIL-2 behavior: object_indices (driving-video object indices after sorting, empty = all), reference_object_indices (reference-image object indices, empty = all - and for single-person references, leave it empty), sort_by (identity color ordering: none, left_to_right, area), plus sam_detection_threshold, sam_max_objects, and sam_detect_interval. The sam_model and sam_conditioning inputs are optional in the schema but required in practice - in replacement mode, which is the default. In animation mode the node skips internal SAM tracking entirely because SCAIL replacement masks aren't used, so sam_model and sam_conditioning genuinely aren't needed there.

The one gotcha the README calls out specifically is a two-person trap: if you set object_indices = 1 to select the second person in the driving video, and your reference image only has one person, do not also filter the reference by 1 - that makes the reference mask empty and your whole reference silently vanishes. Empty reference_object_indices means "all reference objects," and that's what you want for a single-person reference.

Everything else matches the external scheduler: model, clip, vae, sampler, sigmas, clip_vision, pose_video, segment_plan, seed, cfg, mode, max_frames, max_chunk_frames, overlap_frames, reference_count, color_correction, and cache_mode (the same disk cache under output/scail2_cache/long_video). Outputs are frames, used_pose_video_mask, used_reference_mask_timeline, and summary - the mask outputs come back even though you never saw the tracks, which is handy for checking after the fact.

Install is identical to the rest of the pack: ComfyUI Manager search comfyui_scail2_multi_cond, or clone https://github.com/TTPlanetPig/comfyui_scail2_multi_cond into ComfyUI/custom_nodes and restart. No models ship in the pack - you bring the SCAIL-2 weights and the SAM3 model for the internal tracking. The practical tip: keep the external-mask scheduler as your default and reach for this one when the setup is proven. When it works, it's the fastest path from a driving video to a multi-reference SCAIL-2 clip; when it doesn't, you'll wish you'd been able to look at the mask before spending the render.