Wan Configure

Wan Configure (WanVideoConfigure_F2) is where you actually steer a generation in the flow2-wan-video pack. Prompts, resolution, length, steps, CFG - it's all here, and the rest of the pack's nodes mostly read from it. Think of it as the control panel that KSampler keeps in six separate widgets; this pack pulls all of that into one bundle and hands it down the line.

How it works (and the quirk)

The node doesn't send your settings down a wire. It stores them in a class-level Config that the Wan Model Patcher and Wan Sampler read at runtime, and the patch output it returns is effectively a boolean trigger - wire it into the patcher's patch input so the pack knows the config is ready. Two practical consequences: the Configure node has to be in your graph (no Configure, no settings, no render), and if you add a second Configure node, whichever one runs last wins. Don't keep two of them around for A/B testing - change values in place instead. It's the one design quirk of this pack that will bite you, and now you're warned.

The inputs that matter

• positive / negative - multiline prompts. The placeholder on positive says it all: "if positive is empty, the model will decide for itself." Wan is a flow-matching model, not SD-era CLIP, so write descriptive prose; for I2V, keep the image description out of the motion prompt.
• width / height - 512 default, up to 1280, stepped in 16s. 480p (832x480) and 720p (1280x720) are the standard lanes.
• duration - seconds, and this is where Wan's math lives: frames = duration × 16 + 1, because the base runs at 16fps. Default 3s = 49 frames; 5s = 81 frames, which is the native context limit. Push past that and quality degrades.
• guidance_scale - 6.0 default. Full-quality Wan wants CFG 3.5+; if you're running a speed-distillation LoRA you'd normally drop to 1.0, but this pack's defaults are tuned for plain models, so start at 6 and adjust from results.
• sampling_steps - 20 default; the community's full-quality range is 20–30.
• flow_shift - 6.0 default. It's the sigma shift the patcher applies to the model; leave it unless you know why you're moving it.

The advanced row (guidance_percent, enhance_strength, cfg_zero_steps, skip_layer, skip_start_percent, skip_end_percent, extend_video_count) is worth knowing exists but not worth touching on day one. Two notes for later: enhance_strength is the FETA temporal-enhancement that the patcher applies during sampling (not the same thing as the separate Wan Frame Enhancer post-processing node), and skip_layer guidance literally raises an error unless TeaCache is enabled in the patcher - the source code enforces it.

The last one, extend_video_count, has the author's own tooltip and it's the feature people miss: "If this value is 2 or greater, the video will be extended using the last image." Set it to 2+ and the sampler chains clips, re-seeding each one from its own final frame - a poor man's video extension.

Wiring it up

Outputs are patch, width, height. patch goes to Wan Video Model Patcher's patch input; width/height can be driven by the Resize Image node's outputs so your first frame and your resolution agree.

Installing and troubleshooting

Part of the flow2-wan-video pack - install via ComfyUI Manager (search flow2-wan-video) or git clone https://github.com/Flow-two/flow2-wan-video into custom_nodes, then pip install -r requirements.txt and restart. The models auto-download on first run.

If a render comes out wrong, the usual suspects apply: 81-frame ceiling, resolution not a multiple of 16 (this node steps it, but only for the frames), and - the pack's known global gotcha - if you also run native ComfyUI Wan workflows, flow2-wan-video's monkey-patching has broken those for plenty of people. It's the pack's reputation, sadly, not a secret. When in doubt, move the pack out of custom_nodes and test your other workflows with --disable-all-custom-nodes.