DiffSynth: Wan Video Inference

Scroll past the input list on this node and you'll think something has gone wrong. It's enormous. That's not a mistake or a sign of bloat - it's the union of every task the Wan family can do, from text-to-video through VACE control, speech-to-video, motion transfer and music-driven dance. You use three or four sockets. The rest are for a workflow you're not running.

The core settings

If you're doing plain text-to-video, this is the whole node:

• pipe - from DiffSynth: Wan Video Loader.
• prompt, negative_prompt - natural language, cinematic description works better than tag soup.
• seed, num_inference_steps (50), cfg_scale (5).
• height 480, width 832, num_frames 81 - the classic Wan 480p short. 81 frames is the model's native context; it's roughly three and a half seconds at 16fps, and everything the community does at 3–5 seconds starts here.
• sigma_shift (5) and switch_DiT_boundary (0.875) - scheduler and expert-boundary knobs. Wan 2.2 splits denoising between a high-noise expert and a low-noise one; the boundary is where it hands over. Leave both alone unless you're following a recipe.

Image, video and audio conditioning

• input_image - image-to-video. This is most people's actual workflow; a still plus motion beats a prompt alone for anything with a subject you care about.
• end_image - first/last-frame interpolation. Give it both ends and let it join them.
• input_video + denoising_strength - video-to-video. denoising_strength near 1 replaces the clip; drop it to keep more of the original.
• input_audio, audio_sample_rate, motion_video, s2v_pose_video - the speech-to-video family. Wants the audio processor config on the Loader.
• control_video, reference_image - ControlNet-style conditioning and subject reference.
• camera_control_direction, camera_control_speed, camera_control_origin - camera moves on the models trained for it. camera_control_direction takes Left, Right, Up, Down and the four diagonals; origin is a long transform string you should only ever paste from an example.

The rest - vace_video and friends, animate_*, animate2_*, vap_*, longcat_video, wantodance_* - belong to specific checkpoints. Connect a VACE input without VACE weights loaded and you get an error or, worse, a video that silently ignores it.

The ones that actually change your day

• tiled (on by default), tile_size, tile_stride - VAE decode tiling. Wan's 3D VAE is where big clips die, and this is why the default is on. The values are comma-separated strings (30,52 and 15,26), not tuples, because ComfyUI has no fancy string widgets.
• sliding_window_size, sliding_window_stride - how you go past 81 frames. Chunked generation; identity drift across windows is the known cost, and the community's long-video node wrappers exist for exactly this reason.
• tea_cache_l1_thresh - TeaCache, the caching speedup. Somewhere around 0.2–0.3 typically buys real time for a little drift; it's a trade, not a free win.
• cfg_merge, motion_bucket_id, framewise_decoding - niche. cfg_merge matters on Wan 2.2's MoE checkpoints; the others are model-specific.
• latents-adjacent sockets like s2v_pose_latents, audio_embeds - wildcard inputs for precomputed tensors. Ignore them unless a workflow hands you one.

Output is a single video socket, type IMAGE: a batch of frames. Wire to Video Combine or any frame consumer.

Install

Manager → DiffSynth-ComfyUI → install → restart, or clone into custom_nodes after pip install git+https://github.com/modelscope/DiffSynth-Studio.git. Weights are the real cost - the 1.3B T2V model is a few gigabytes; 14B is tens.

Common issues

• Out of memory on decode, not sampling. Tiling is already on by default; if you turned it off, turn it back on.
• A four-minute clip that loops or degrades. Native context is 81 frames. Chunk it deliberately or accept the mush.
• Wrong task weights. The Loader loads what you tell it; this node then tries to use inputs the checkpoint was never trained for. Copy a working example graph, then edit.
• Long waits on a mid-range card. 81 frames at 480p is minutes, not seconds, and 14B on a 12GB card is a different sport. Get a baseline at 1.3B first.