WAN 2.2 Image Guides (All-in-One)

If you just want to feed WAN 2.2 a handful of storyboard frames and get guided conditioning out the other side, this is the node to grab first. WAN22MultiImageI2VGuide is the pack's all-in-one: the guide picker, the frame timeline, and the conditioning/latent output all on a single canvas node. No manager-to-apply handoff, no wiring two nodes together just to see your first keyframed clip.

It comes from comfyui-helto-wan22, and it's basically the pack's other two nodes fused - the WAN22ImageGuideManager's shot-list UI and shared settings, plus the WAN22ApplyImageGuides conditioning machinery - collapsed into one. The display name is "WAN 2.2 Image Guides (All-in-One)," which is confusingly close to the Manager's name. Don't overthink it: the Manager outputs a reusable WAN22_IMAGE_GUIDES payload; this node outputs conditioning and a latent, ready to sample. The naming difference is the whole distinction.

How it works

Everything about the mechanism matches WAN22ApplyImageGuides: your guide images get spliced into a neutral-frame timeline, VAE-encoded once, and attached to the conditioning via WAN's native concat_latent_image / concat_mask / concat_mask_index. WAN I2V models expose dedicated image-conditioning channels, and this node speaks directly to them. The difference is where the settings live. On the Apply node, timing and strength settings travel inside the payload from a separate Manager. Here, the settings are plain inputs on the node itself - fps, timing_mode, resize_mode, duplicate_policy, pad_color, global_strength, start_images_strength, and structural_repulsion_boost - so you tune the whole stage in one place. Same knobs, different home.

The custom UI is the same storyboard interface: folder browser, thumbnails, per-guide frame position and strength, enable toggles, saved guide sets. Everything serializes into the hidden guides_json widget, so a saved workflow carries your shot list with it.

Inputs and outputs that matter

Required inputs: positive/negative conditioning (from the WAN text encoder), a WAN 2.2 vae, width/height, length (frame count; WAN latent length is ((length - 1) // 4) + 1), batch_size, and the hidden guides_json that the UI writes. The optional trio - start_images (a sequence inserted from frame 0), ref_image (encoded to reference_latents), and control_video (encoded into the first concat slot) - are the same across the pack.

Outputs: positive_high, positive_low, negative (conditioning), and latent. Wire positive_high into your high-noise sampler, positive_low into the low-noise sampler after the switch step, negative into both, and feed the empty latent to whichever sampler runs first. This is a stage node - it won't give you frames until a sampler and VAE decode come after it.

How to install

Same pack, same drill. No requirements.txt, no bundled models:

cd ComfyUI/custom_nodes
git clone https://github.com/helto4real/comfyui-helto-wan22

Restart ComfyUI (or install "comfyui-helto-wan22" via ComfyUI Manager). The WAN 2.2 I2V high/low models, WAN VAE, and umt5 text encoder still go in your models/ folders the normal way.

Where people get burned

• The classic duplicate-frame error. Default duplicate_policy is error, so two guides landing in the same 4-frame group will stop the run. Set keep_first, keep_last, or offset_next once you have overlapping shots.
• It's easy to mistake this for a generator. No sampler inside - the output latent and conditioning need real sampler nodes (or the pack's WAN22GenerateAllInOne) to become frames. If you've wired it to nothing and see nothing, that's expected.
• Guides reference files by path, so moving a source image after arranging your timeline breaks the run at encode time.

The split-node route (Manager → Apply) still wins when you'll reuse the same shot list across several workflows - the payload is the reusable artifact. But for a self-contained keyframed I2V stage, this all-in-one is the fastest path to a first result, and you can always refactor into the split later.