Wan Sampler

Wan Sampler (WanVideoSampler_F2) is the KSampler of this pack - the node that actually denoises the video and hands back your frames. But unlike a generic KSampler, it's Wan-aware in a way that saves you a lot of plumbing: feed it a start image and it runs image-to-video; give it a start and end image on the right model and it runs the Fun-InP inpaint workflow; leave both empty and it's plain text-to-video. One node, three modes, decided by what you plug in.

How it works

It reads its marching orders from the Wan Configure node's config (frames, resolution, steps, CFG, prompts) - so Configure has to be in the graph and wired through the patcher, or there's nothing to sample. Then it:

• encodes your prompts with the auto-loaded umt5 text encoder (for I2V it also runs the image through CLIP-Vision),
• builds an empty latent with the right temporal shape - frames = duration × 16 + 1, spatial dims divided by 8 - and, for I2V, encodes your start frame into a concat_latent_image that gets tacked onto the conditioning,
• samples with your chosen sampler/scheduler (defaults: uni_pc / sgm_uniform, with 44 samplers and 9 schedulers to pick from),
• optionally runs a denoised_output pass - an extra 5-step dpmpp_2m refine at 0.49 denoise, for when you want a cleaner finish at a real time cost,
• and VAE-decodes, either in one shot or tiled (vae_decode_type + vae_tile_size) for VRAM relief, with a live TAE preview whose resolution you control via preview_resolution.

Model-mode detection is automatic and worth understanding: if the loaded model name contains i2v (or fun + inp), it treats start_image as an anchor; the Fun-InP inpaint models take both start_image and end_image and fill the video between your first and last frame. The Fun control models are a different story - their image-encoding function is literally pass # todo in the source, so don't expect the control variants to work yet.

The extend_video_count from Configure plugs in here too: when it's 2+, the sampler runs the clip, then re-seeds each subsequent segment from its own last frame, chaining them into one longer output.

The inputs that matter

• model - from the Wan Model Patcher.
• seed - the usual; lock it for reproducible runs.
• sampler / scheduler - uni_pc/sgm_uniform are fine defaults. The KB's Wan advice: res_2s is the quality sampler at roughly double the time, euler + beta is the fast option. Both are in the list.
• start_image / end_image (optional) - from the Resize Image node. Wire start_image and you've built an I2V workflow; add end_image on a Fun-InP model and you've built an inpaint-between-frames one.

Output is a single images tensor - feed it to the Wan Frame Enhancer for upscaling/interpolation, then into a VHS_VideoCombine node to save your clip (the pack's example workflows use VideoHelperSuite for that; the pack doesn't ship its own video output node).

Installing

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

Troubleshooting

• Resolution surprise: for I2V the sampler takes its width/height from the start image itself, not the Configure node. Resize your start frame to the resolution you actually want.
• 81 frames is the wall. At 16fps that's ~5 seconds. Going much past it produces warping and looping - that's Wan, not the node. Chain via extend_video_count or move to an extension workflow.
• The group=1 channel error: Given groups=1, weight of size [5120, 36, 1, 2, 2]... is the community's signature error with this pack - flow2-wan-video's global patches have broken native Wan workflows for a lot of people. If you're running the pack's own pipeline and see it, reinstall clean; if another workflow breaks while the pack is installed, the pack is the prime suspect.
• Text encoding takes a while on first run - the umt5 fp8 encoder is loading; it gets cached after.