XB-BOX - Wan Param Bus

Long-form Wan has a wiring problem. The standard "chain 81-frame clips" workflow means every segment needs the same model, same VAE, same sampler, same resolution, same seed - and if you build that by hand, you get a noodle nest of wires and a single source of truth that lives in five places. XB_Wan_ParamBus ("XB-BOX - Wan Param Bus") is the pack's fix: one node holds every shared parameter, packs them into a single WAN_BUS connection, and every relay node downstream just reads from the bus.

It's the heart of the pack's "infinite video" system. The idea, lifted straight from the author's README philosophy of "unified parameters" and taming complex workflows, is that you configure once and the relays stay clean. If you're coming from ComfyUI's stock Wan I2V graph, think of this as replacing a wall of KSamplerAdvanced settings with a single dict you hand around.

How it works

The bus takes the two-pass Wan 2.2 architecture seriously: model_high (motion/composition) and model_low (detail refinement) are separate inputs, and high_noise_steps tells the pipeline where the high-noise pass hands off to the low-noise pass. Everything else is packed in:

• Generation basics - width, height, length, fps (these are forced inputs, so you can wire them from other nodes), negative_prompt, steps, cfg, sampler_name, scheduler, seed.
• Chunking - vae_encode_tile_size, vae_decode_tile_size, spatial_overlap, temporal_chunk_size, temporal_overlap. These are the memory knobs that let a 12GB card chew 81 frames instead of OOMing.
• Relay bookkeeping - cut_first_frame (drop the duplicated first frame between segments) and trim_head_frames (how many overlapping frames to trim).

Two optional inputs do real work: global_start_image plus clip_vision gets CLIP-vision-encoded automatically (the node scales the image to your resolution and encodes it for you - no separate encode node needed), and concat_mode (自动/强制GPU/强制CPU) picks where accumulated video gets stitched, which matters on low-VRAM boxes where GPU concat can spike memory.

The output is a single WAN_BUS. Wire it into XB_Wan_InfiniteRelayNode or the New variant, one per segment.

What to actually set

Keep the defaults until something breaks: 4 steps is the distilled/speed-LoRA regime the pack assumes, cfg 1.0 pairs with step-distilled checkpoints (crank it to 3.5+ only on a full-quality model), euler/simple is the fast default. The dials you'll actually touch are length (81 is the safe ceiling) and the chunking values if you're fighting VRAM. high_noise_steps must stay below steps - the relay nodes enforce that for you, but keep them in the right order.

Install

ComfyUI Manager → search XB_ToolBox, or git clone https://github.com/wjluoxiao/XB_ToolBox.git into custom_nodes, restart. No weights ship with it; the pack's only declared pip deps are opencv-python and easyocr. The bus is just a dict - nothing runs until a relay node consumes it, which makes it cheap to experiment with.