Nodes/XB_ToolBox/XB-BOX - ๐Ÿ†• SCAIL ๆ— ้™ๆŽฅๅŠ›็‚น (New)
ComfyUI Node

XB-BOX - ๐Ÿ†• SCAIL ๆ— ้™ๆŽฅๅŠ›็‚น (New)

Turn SCAIL into a five-minute video

By wjluoxiaoยทCreated 7 months agoยทUpdated a day agoยท 406
XB-BOX - ๐Ÿ†• SCAIL ๆ— ้™ๆŽฅๅŠ›็‚น (New)
  • wan_scail_bus
  • prev_video
  • ๐Ÿ“ฆ WAN_SCAIL_BUS (ไผ ็ป™ไธ‹ๆฎต)
  • ๐ŸŽž๏ธ ็ดฏๅŠ ่ง†้ข‘ๆต
โ—„segment_length81โ–บ
โ—„positive_promptDescribe the scene...โ–บ
โ—„previous_frame_count5โ–บ
โ—„relay_count1โ–บ
โ—„total_frames_displayโ–บ

This is the node that makes SCAIL infinite ("๐Ÿ†• SCAIL ๆ— ้™ๆŽฅๅŠ›็‚น (New)"). A single SCAIL generation produces one segment of about 81 frames - a few seconds. This relay node chains those segments: it generates one, keeps the tail frames, generates the next segment continuing from them, trims the overlap, appends, and repeats. Set relay_count to 10 and you get a coherent multi-minute character video from a model that can only natively do seconds. It's the SCAIL counterpart of the same "relay" idea XB uses for first/last-frame Wan.

How it works

Each relay iteration reads everything from the wan_scail_bus (the ParamBus node's output) and adds only the per-segment bits. It computes the current offset into the pose video, grabs the last previous_frame_count frames of the accumulated video as previous_frames (that's the tail stitching - the new segment literally starts from where the old one ended), encodes your per-segment prompt, then hands it all to the pack's XB_WanSCAILToVideoPro to build conditioning and a latent. Sampling runs through a standard KSampler with the bus's steps/cfg/sampler/scheduler, decode uses tiled decode from the bus's VAE knobs, the overlapping frames get trimmed, and the result is appended via concat_mode (CPU accumulation when VRAM is tight). If relay_count is more than 1, it loops internally instead of requiring you to copy-paste more nodes.

The inputs that matter

  • wan_scail_bus - the WAN_SCAIL_BUS output of the ParamBus node.
  • segment_length - default 81. SCAIL is trained on 4N+1 frame sections, and 81 is the standard unit; if the remaining pose frames don't fit, the last segment snaps to the nearest valid length.
  • positive_prompt - per-segment; leave blank to reuse the bus's prompt.
  • previous_frame_count - default 5; the overlap frames for tail stitching. Keep it modest; too high and you're re-generating a lot of already-done footage, too low and continuity jumps.
  • relay_count - how many segments this node generates (1 = generate one, pass the bus on to the next relay node in the chain).
  • total_frames_display - a display of the automatically computed total; mostly read-only.

Outputs: WAN_SCAIL_BUS (pass to the next relay) and the Accumulated Video Stream (IMAGE).

Install

Pack install - Manager search XB_ToolBox or clone into custom_nodes, restart. Plus the SCAIL model (the Wan 2.1 14B SCAIL fp8 from kijai's WanVideo_comfy_fp8_scaled is the community default for 16GB cards) and the ParamBus upstream. Community notes on SCAIL-2 that carry over here: it's trained at 16fps but works fine at 24fps, and you extend by chaining 81-frame sections rather than one giant generation.

Common issues

If the chain stops partway, the pose video ran out of frames - the node prints how many relays completed. A visible jump between segments means previous_frame_count is too small (or your pose/motion isn't continuous). And if VRAM climbs segment after segment, force concat_mode to CPU in the bus so the accumulated video stops squatting on the GPU.

CategoryXB_ToolBox/Pipeline

Inputs (7)

NameTypeDefaultDescription
wan_scail_busWAN_SCAIL_BUSโ€”
segment_lengthINT811โ€“8192โ€”
positive_promptSTRINGDescribe the scene...โ€”
previous_frame_countINT51โ€“8192ๆŽฅๅŠ›้‡ๅ ๅธงๆ•ฐ๏ผˆๅฐพๅธง่ก”ๆŽฅ๏ผ‰
relay_countINT11โ€“999ๆŽฅๅŠ›ๆ•ฐ้‡่ฎพๅฎš
total_frames_displaySTRINGๆ€ป่ฎก็”Ÿๆˆๅธงๆ•ฐ๏ผˆ่‡ชๅŠจ่ฎก็ฎ—๏ผ‰
prev_videooptIMAGEโ€”

Outputs (2)

NameTypeDescription
๐Ÿ“ฆ WAN_SCAIL_BUS (ไผ ็ป™ไธ‹ๆฎต)WAN_SCAIL_BUSโ€”
๐ŸŽž๏ธ ็ดฏๅŠ ่ง†้ข‘ๆตIMAGEโ€”