🌊🍿 FlowState WAN Studio

Wan is where local video generation consolidated - Alibaba's open Apache 2.0 family, frozen at 2.2, and still the base most open video work sits on. FlowState WAN Studio is the pack's flagship: a single node that loads the models, runs the two-stage MoE sampling, decodes, and hands you a playable video. No graph of loaders, schedulers, and latent nodes to build. Type a prompt, optionally drop in a starting frame, and hit run.

The mechanism, and why it needs two models

Wan 2.2 is a mixture-of-experts video model: denoising is split between a high-noise expert that handles motion and composition and a low-noise expert that handles detail. That's why this node asks for two diffusion models - high_noise_model_name and low_noise_model_name - where you'd expect one. Internally it runs the high-noise model through the first half of the steps (KSamplerAdvanced, from step 0 to the midpoint) with noise added, then hands the partially-denoised latents to the low-noise model for the second half with add_noise disabled. It's the same split the community exploits with speed LoRAs on the low-noise pass only, just baked in so you never have to orchestrate it.

The node also:

• Uses ComfyUI's WanImageToVideo to encode your prompt and prepare the latent batch (it handles text-to-video and image-to-video via the optional starting_frame).
• Applies a fixed baked-in negative prompt - blurring, warping, blurred, morphing - so there's no negative field to fiddle with.
• Decodes with tiled VAE decode (512px tiles, temporal overlap) because decoding a full video in one pass will OOM a mid-range card.
• Assembles the frames with CreateVideo and optionally saves via SaveVideo, all internally.

Inputs that matter

• high_noise_model_name / low_noise_model_name - the two Wan 2.2 expert checkpoints. On constrained hardware, Wan's 5B dense model runs in both slots; on a bigger card you use the 14B/27B MoE pair.
• num_video_frames (25 default), fps (12 default) - length and playback speed. Keep fps near 12–16 for Wan.
• fixed_output - this one's easy to misread. True means same output every run (the seed is pinned); False means it rolls a fresh random seed each run so you get variations. It's a seed-lock toggle wearing a confusing name.
• resolution - presets plus a "Use Starting Frame Resolution" option that reads the dims off your starting_frame image, which is the sane way to do image-to-video.
• save_video (False default) - if off, the node still hands back a video output you can route into FlowState_VideoPreview; flip it on and it writes to output/video/FlowState_WANStudio (the filename_prefix supports %date%-style tags).
• pos_prompt - the whole text input. There's no negative prompt because the author hardcoded one.
• audio (optional) - mux an audio track into the saved file.

Outputs

video (VIDEO). That's it - wire it to FlowState_VideoPreview to watch, or FlowState_VideoCreator if you want to control the container yourself.

Install

Pack install is the easy half:

cd ComfyUI/custom_nodes
git clone https://github.com/flowstateeng/FlowState-Creator-Nodes

then restart, or search FlowState-Creator-Nodes in ComfyUI Manager. No pip dependencies. The heavy half is the model files: two Wan diffusion models (high + low noise), the Wan text encoder (clip_name), and the Wan VAE (vae_name). Those are multi-gigabyte downloads, and unlike the pack they're non-negotiable.

Where people get burned

The #1 mistake is loading only one diffusion model and wondering why the dropdown for the other is empty - you need both populated. Second: if fixed_output is on and you change the prompt but get suspiciously similar motion, you're seeing seed-lock, not a bug; flip it off for variety. And sage_attention will sit on "disabled" until you install Sage Attention plus KJNodes - the suite's startup log tells you which is missing. Video is also just hungry: 25 frames at 1280×720 on a 14B model will test a mid-range card, which is exactly why the tiled decode and the 5B dense option exist.