Seedance 2.5 is live — 30-second cinematic video with native audio & real-person references
All Posts
Wan 3.0 ComfyUI: API Workflows and Setup Requirements
2026/09/14

Wan 3.0 ComfyUI: API Workflows and Setup Requirements

Set up Wan 3.0 ComfyUI with official templates or a custom reAPI adapter. Compare account requirements, media inputs, task polling and dated cost estimates.

Wan 3.0 ComfyUI workflows are available through hosted API nodes. Comfy's official announcement links three templates: text to video, image to video and reference to video. Start with those templates if your goal is to generate inside an existing graph; a local Wan 3.0 checkpoint is not a prerequisite.[1]

Running the interface on your computer does not make the generation local. The official Wan indexes checked on September 14, 2026 did not list downloadable 3.0 weights. For reAPI users, the alternative is a custom integration with the Wan 3.0 API, where your own code submits and monitors the task. That is a separate path from Comfy's built-in billing.[2][3][4]

TL;DR

  • Comfy lists three Wan 3.0 templates. Its launch instructions specify ComfyUI v0.33.4; update before looking for the nodes.[1]
  • Official Partner Nodes require a Comfy account and paid credits. A reAPI key cannot simply replace the account credentials in that workflow.[5]
  • The Comfy announcement documents 480p, 720p and 1080p. reAPI documents additional 2K and 4K tiers, with restrictions on document and webpage input.[1][4]
  • A custom reAPI adapter needs public media URLs, a generation request, task polling and an output download. No local Wan checkpoint participates in that request.[4]
  • For inference on your own GPU, use a downloadable Wan generation such as 2.2 and its native ComfyUI workflow.[6]

Wan 3.0 ComfyUI templates: choose the input first

The three official templates solve different starting points. Choose text to video when you have only a prompt. Choose image to video when a supplied frame should anchor the clip. Choose reference to video when images, video or audio guide what the model creates.[1]

Official templateTemplate identifierStarting material
Wan 3.0 Text to Video Workflowapi_wan3_0_t2vA written scene
Wan 3.0 Image to Video Workflowapi_wan3_0_i2vA source image and direction
Wan 3.0 Reference to Video Workflowapi_wan3_0_r2vReference assets and direction

The identifiers come from the links in Comfy's announcement. Open that announcement to reach the corresponding template, or find the model in the template panel of an updated installation. Do not look for a similarly named community workflow and assume it is the official one.[1]

This choice also sets expectations about control. A frame is an intended boundary of the shot. A reference supplies information the model interprets. They are different instructions, even when both begin with the same image file. Start with the smallest input set that expresses your intended shot, then add references only when they resolve a specific ambiguity.

Start the official workflow with a Comfy account

Use this sequence for the documented Comfy path:

  1. Update ComfyUI. The Wan 3.0 introduction named v0.33.4; Comfy's troubleshooting guidance recommends a current build when Partner Nodes are missing.[1][5]
  2. Open Settings → User and sign in to your Comfy account. On a supported alternative deployment, Comfy also documents login with a Comfy API key.[7]
  3. Open Settings → Credits and check the paid credit balance. Installing ComfyUI does not include unlimited hosted model calls.[8]
  4. Load the appropriate Wan 3.0 template, enter the prompt, supply its requested inputs and connect the output.[1]
  5. Review the model settings and displayed cost before running. Keep the first test short enough to assess the specific behavior you need.

If a node is missing, check the installed version before reinstalling unrelated extensions. If authentication fails, check the account and how you opened the local interface. Comfy's FAQ discusses localhost and 127.0.0.1 restrictions, while its login guide describes a Comfy API-key method for other deployments. Use the documented method for your installation.[5][7]

That key is an account credential for Comfy. It is not permission to paste an arbitrary video API key into the same field. Comfy's Partner Node FAQ does not document bringing your own model-provider key.[5]

Why a KSampler does not turn the API into local inference

A hosted Wan 3.0 task runs outside your GPU. The presence of a node in a local graph tells you where the task is organized, not where model inference happens. Comfy's Partner Nodes call hosted models; downloadable Wan 2.2 workflows instead load diffusion weights, a text encoder and a VAE.[5][6]

That distinction answers searches for wan 3 ksampler and wan 3 samplers. Connecting a local sampler cannot supply weights that the official Wan 3.0 indexes do not provide. Nor does changing a local sampler configure a setting absent from the remote API.[2][3][4]

Other nodes in your graph may still need local GPU memory. For example, a separate local image-generation stage has its own hardware requirements. The narrower benefit is that the hosted Wan stage does not load Wan 3.0 weights on your machine. For a fully local graph, follow which Wan models can run locally.

Connect reAPI through a custom adapter

Choose a custom adapter when you need to use your reAPI balance, store task IDs in your own application, or submit jobs from an existing queue. This article describes the integration contract; it does not claim that a ready-made reAPI node or tested downloadable plugin is included.

The first step is an ordinary request. Keep the API key in the server environment, then submit:

curl --fail-with-body https://reapi.ai/api/v1/videos/generations \
  -H "Authorization: Bearer $REAPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan3.0-video",
    "prompt": "A ceramic cup turns slowly on a wooden table, morning light, quiet room ambience",
    "resolution": "720P",
    "size": "16:9",
    "duration": 6,
    "audio": true
  }'

This request follows the published schema and incurs a charge when executed successfully. It is a request example, not a report of a generation test. Wan 3.0 generation on reAPI requires an API key; the model Playground configures requests but does not submit browser generations.[4]

Read the task identifier from the response and poll GET /api/v1/tasks/{id}. When status becomes completed, retrieve output.video_urls. If the task fails, expose the returned error. Preserve an existing task ID when polling is interrupted instead of submitting a second generation automatically.[4]

An adapter also needs to translate graph inputs into the correct public fields:

Graph inputreAPI request fieldImportant boundary
Opening or closing frameimage_with_rolesUse first_frame or last_frame explicitly
Identity/style imageimage_with_rolesUse reference_image; do not mix with frame roles
Motion/reference footagevideo_urlsReference input, not a dedicated edit source
Reference soundaudio_urlsPublicly reachable media URLs
Documentfile_urlOne document; incompatible with link_url
Public webpagelink_urlNo login requirement; incompatible with file_url

All media fields require public HTTP(S) URLs. A Comfy image object, filesystem path or base64 data URI is not a valid replacement. Add an upload step before submission if your graph starts with a local file. Keep frame and reference families separate, and stay at 1080P or below for document/webpage input.[4]

Budget the complete request and choose a path

On September 14, 2026, the reAPI Wan 3.0 price table displayed $0.035/s at 480P, $0.068/s at 720P, $0.136/s at 1080P, $0.199/s at 2K and $0.229/s at 4K. These are displayed rates; the page's structured offer retains more decimal precision.[9]

Using the displayed 720P rate, ten output seconds suggest $0.680 before reference-video additions and task-level credit rounding. That is a budgeting estimate, not an exact invoice. Reference-video duration is billable too. A six-second reference plus eight output seconds represents fourteen billable seconds, and the input-video/output total must remain within thirty seconds.[4]

Set an explicit output duration when budgeting. reAPI documents that duration: -1 uses the thirty-second billing ceiling. Turning off generated audio does not reduce the per-second rate, and failed jobs receive an automatic refund.[4]

DecisionOfficial Comfy templateCustom reAPI adapter
SetupExisting official templateIntegration work required
Payment accountComfy creditsreAPI credits
Request controlsControls exposed by the selected templateFields documented by reAPI
Batch organizationUse the workflow's available execution toolsImplement your own queue, task storage and retry policy
Cost comparisonCheck Comfy's current node quoteCheck the dated model table and actual usage

There is no verified price comparison here that establishes one path as universally cheaper. Choose the official template for its setup convenience; choose the adapter for control over the surrounding integration. A self-built queue does not create unlimited concurrency or exempt calls from rate limits.

FAQ

wan 3.0 comfyui

Yes. Comfy's official announcement provides text, image and reference workflows for hosted Wan 3.0 generation. Start with one of those templates.[1]

comfyui wan 3.0

Update ComfyUI, sign in to a Comfy account and check the credits balance. Then load the matching template. Official Partner Nodes use Comfy's account system.[1][5]

wan 3 ksampler

A local KSampler does not run hosted Wan 3.0 weights. Use the remote model node, or choose a downloadable Wan generation for a local diffusion workflow.[6][2]

wan 3 samplers

Use only the controls exposed by the chosen integration. The reAPI Wan 3.0 schema does not expose a sampler-selection field.[4]

wan 3.0 local tutorial

For a locally running interface, follow the official API template steps above. For model inference on your own hardware, use a Wan 2.2 native workflow and its required model files.[1][6]

wan 3.0 local ios

The checked official sources do not provide a Wan 3.0 iOS weight package. Accessing a hosted interface from a device is different from running the model on that device.[2][3]

Pick the workflow that matches where generation should run

Use the official template to get hosted generation into ComfyUI with the least integration work. Use reAPI's model page and API docs when you need your own request lifecycle. If local inference is essential, continue with the open-weight status guide and local video model comparison. A Wan 3.0 ComfyUI workflow is useful once the execution location, account and model controls are explicit.

References

  1. Comfy. Wan 3.0 in ComfyUI: Native 30-Second Video with Omni-Reference Control. Retrieved September 14, 2026 from blog.comfy.org/p/wan-30-in-comfyui-native-30-second
  2. Wan-Video. Wan. Retrieved September 14, 2026 from github.com/Wan-Video
  3. Wan-AI. Wan-AI (Wan-AI). Retrieved September 14, 2026 from huggingface.co/Wan-AI/models
  4. reAPI. Wan 3.0 | reAPI. Retrieved September 14, 2026 from reapi.ai/docs/wan-3-0
  5. Comfy. FAQs about Partner Nodes - ComfyUI. Retrieved September 14, 2026 from docs.comfy.org/tutorials/partner-nodes/faq
  6. Comfy. Wan2.2 Video Generation ComfyUI Official Native Workflow Example - ComfyUI. Retrieved September 14, 2026 from docs.comfy.org/tutorials/video/wan/wan2_2
  7. Comfy. Log in to your Comfy account - ComfyUI. Retrieved September 14, 2026 from docs.comfy.org/account/login
  8. Comfy. Credits Management - ComfyUI. Retrieved September 14, 2026 from docs.comfy.org/interface/credits
  9. reAPI. Wan 3.0 — 30-Second Omni-Modal Video Generation. Retrieved September 14, 2026 from reapi.ai/models/wan-3-0
All Posts

More Posts