Wan Director

Every Wan video workflow starts the same way: one prompt, one clip, one flat 5-second shot. Fine for a demo. The moment you want a real video - multiple shots, different prompts per segment, audio that lines up - stock nodes leave you splicing separate generations together, or cramming everything into one global prompt and watching the model blend a "close-up of a cat" into a "city skyline" into something that is neither.

Wan Director is the answer that node-level ComfyUI has been missing: a visual timeline, rendered inside the node itself, where you drag image, text, and audio clips onto a track and it hands back the whole conditioning/latent/fps/audio bundle the sampler needs. It's the Wan-native cousin of LTX Director (the author credits WhatDreamsCost's design), rebuilt around Wan's 3D VAE - 16 channels, ÷8 spatial, ÷4 temporal - and it leans on Wan 2.1/2.2 still being the open-video base: Apache 2.0, frozen at 2.2, and the community's default for image-to-video from a strong still.

Honest framing up front: this is a young node from a fast-moving pack with essentially zero public reputation yet - no Reddit chatter, no hype train. You're an early adopter, and if that's fine, this is genuinely the most capable Wan timeline tool I've seen in a single node.

How it works

The heavy lifting is split between a JS extension and the Python node. The JS renders a dual-track timeline (image/text, audio): drag to move or resize segments, right-click to add, drop files to upload, and it serializes everything into hidden JSON string widgets (timeline_data, local_prompts, segment_lengths) that the node parses on execute.

From there it assembles the clip: it builds the Wan empty latent at the right shape, VAE-encodes your first image clip into frame 0 for i2v variants, mixes the audio segments into one waveform, and - the interesting bit - applies PromptRelay, which patches the model's cross-attention so each segment's prompt only steers its own span of frames. With 2+ clips, per-clip prompts become local prompts and global_prompt is the anchor; with a single clip it's a no-op and behaves like a flat prompt. A stack of quality toggles - dynamic CFG, PAG, NAG, AsymFlow, SLG, FETA, RIFLEx, teacache/magcache/easycache - patches the MODEL and emits a quality_recipe JSON for the sampler.

The inputs that matter

Most of the knob farm is advanced and collapsed by default. What you actually set:

• model_variant - the big one. Eight choices covering wan2.1_t2v / wan2.1_i2v, wan2.2_t2v / wan2.2_i2v, wan_fun_inp / wan_fun_control, wan_animate, and wan2.2_animate_everanimate. It gates which sliders appear and how the latent/conditioning are assembled, so pick it first.
• backend - native (ComfyUI's built-in Wan; connect model + clip) or kijai (Kijai's ComfyUI-WanVideoWrapper; connect wan_model + wan_t5).
• duration_frames and frame_rate - 81 frames ≈ 5 s at 16 fps, which is what Wan 2.x trains at. Raise the frame rate for slow-motion-ish output.
• global_prompt - persistent context prepended to every per-clip prompt.
• enable_prompt_relay - on by default; it's the whole point of the timeline. Turn it off to encode one flat prompt.

Outputs and wiring

The native branch emits model (patched), positive/negative CONDITIONING, video_latent, frame_rate, combined_audio, reference_image, and an info JSON with warnings. Wire model + positive/negative + video_latent into the Wan sampler, frame_rate into its FPS and the saver, and combined_audio into a save node that takes AUDIO. The kijai branch instead populates wan_model and wan_text_embeds - feed those straight into WanVideoSampler.

Install

It ships in the umbrella ComfyUI-CustomNodePacks pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks.git

or ComfyUI Manager → search "CustomNodePacks" → install, then restart. The node downloads no weights - bring your own Wan checkpoint and UMT5 via your normal loaders. Two extras: the kijai backend needs ComfyUI-WanVideoWrapper installed separately, and the EverAnimate variant expects an everanimate_<stage>.safetensors LoRA in your loras folder.

Gotchas

• Mixed sockets - native requires model + clip; kijai requires wan_model + wan_t5. Get it wrong and the node raises a clear error naming what's missing.
• vae_fp32_decode is on by default - Wan's VAE genuinely looks better in fp32, but it costs VRAM during decode. Turn it off if you're bumping the wall.
• Variant-specific sliders are ignored elsewhere - the 2.2 dual-CFG sliders (cfg_high_noise/cfg_low_noise) only apply to 2.2 variants, ref_strength only to Animate, the everanimate_* group only to the EverAnimate variant. Setting them on the wrong variant does nothing, despite what the tooltips say.
• The README is stale - the pack has grown well past its MaskEditControl roots and its README doesn't even mention this node. Trust the in-node tooltips and the auto-generated node reference over the marketing prose.
• Wan 2.2 speed LoRAs are a separate caution: the community consensus is they can wreck composition and motion, so use them sparingly - low-noise pass only.

If you're assembling multi-shot Wan videos with audio and per-segment prompts, there is nothing else that does this in a single node. Expect rough edges, read the info output, and keep your expectations calibrated - but the design is right.