Pipe In Context Video (WVW) v2

Pipe In Context Video (WVW) v2 is the top of the Wan context line and arguably the most serious node in the whole pack: it carries two Wan models (model_1, model_2, both WANVIDEOMODEL) plus a standard model/clip/vae trio, loras, text_embeds, image_embeds - and then three inputs no other node in this pack has: torch_compile_args (WANCOMPILEARGS), block_swap_args (BLOCKSWAPARGS), and vram_management_args (VRAM_MANAGEMENTARGS).

That trio is the WanVideoWrapper VRAM survival kit, and it's the reason this node exists. Wan 2.x's 14B diffusion transformer is enormous for a home GPU, and the wrapper's answer is block swap (offloading transformer blocks to CPU RAM so a 24GB card can run the 14B) and torch.compile (fusing ops to cut both VRAM and inference time). You configure those behaviors in the wrapper's setup nodes and pipe the resulting args objects into this context, so the whole "how do I even fit this model" configuration rides the same wire as the models themselves. If the other context nodes are about tidy wiring, this one is about making a 14B video model run - the two-model layout points at the classic two-pass pattern (cheap first pass for motion, better second pass for detail) that people actually use to make Wan 2.2 work at home.

The context mechanics are the usual rgthree-style dict: base_ctx in, per-field overrides, inherit the rest, everything out on context plus individual outputs. The rest of the slots are the familiar video set - frame loader knobs, images_input, images_ref_start/images_ref_end, images_output, audio in/out, two masks, five images_pp_*, five any_*, steps/cfg/seed/path.

Install. ComfyUI Manager → search "RvTools" → install ComfyUI-RvTools_v2 and restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI-RvTools_v2
# restart ComfyUI afterwards

No models ship with it. Manual clones: pip install -r requirements.txt (opencv-python, pynvml, piexif, Pillow, pilgram). You must also have ComfyUI-WanVideoWrapper installed - the WANCOMPILEARGS, BLOCKSWAPARGS and VRAM_MANAGEMENTARGS inputs come straight from that pack, and without it this node has nothing to talk to.

The catches. This is the most environment-sensitive node in the pack: it's only useful to WanVideoWrapper users on tight VRAM budgets doing two-model passes. If that's not you, it's just an enormous box of unused slots. Everything else is the familiar caveat list - pack unmaintained with ComfyUI_Eclipse as the README's successor, context is pipe-typed and needs a matching Out node, and cross-pack context dicts don't merge cleanly. When it works it's great; just make sure the wrapper versions line up, because block-swap and compile arg formats have shifted across WanVideoWrapper releases.