Wan Image to Video: Run It Locally or on a Hosted Service

Because Wan’s weights are open, the same image-to-video model is available in two very different places: on your own machine through ComfyUI, and behind a web form on hosted services such as Replicate, fal and Krea. Creatives often use both, drafting on one and finishing on the other. This guide walks through each route, compares them honestly, and shows how to keep your prompts and settings organised so a clip made on one can be reproduced on the other.

What is the same on both routes

The model does not know where it is running. Whether you are in a ComfyUI graph or a hosted form, Wan image-to-video takes the same inputs: a still image, a prompt describing motion and camera, a negative prompt, a resolution, a frame count, sampling steps, a guidance value and a seed. The Wan video model guide explains each of those. If you record them for every clip, you can move between routes freely. What differs is cost structure, speed, control and how much of the setup is done for you.

Route one: local in ComfyUI

ComfyUI is a node-based interface that runs on your own hardware and has native support for Wan. The one-off effort is higher; after that each clip costs only electricity and time.

1. Install ComfyUI and update it. Wan support arrived in a recent version, so an old install will not have the nodes.
2. Download the Wan image-to-video checkpoint plus the text encoder and VAE files the model card lists, and place them in the folders ComfyUI expects. Pick the variant your graphics card’s memory can hold; quantised builds exist for smaller cards.
3. Load an example workflow. ComfyUI ships Wan templates. Start from one rather than building a graph from scratch.
4. Drop in your still at the resolution the checkpoint prefers, write the motion prompt and negative prompt, set a low frame count and the default steps, fix a seed.
5. Queue the render and watch memory usage. If it runs out, lower resolution or frame count or switch to a smaller variant.
6. Save the output with the workflow embedded. ComfyUI can write the whole graph, including every setting, into the saved file’s metadata. Keep that on. It is the best record you will get anywhere.

Local strengths: no per-clip cost, control over every node, private inputs, and upscaling and frame interpolation in the same graph. Local weaknesses: setup time, hardware limits on which variant you can run, and slow renders on mid-range cards. See the ComfyUI beginners guide if the interface is new to you.

Route two: hosted services

Replicate, fal, Krea and others run Wan on their own servers. You upload a still, fill in a form and pay per generation or per second of video, with pricing on each service’s current page.

1. Find the Wan image-to-video model on the service. Check the version number against the model card you have read; hosts often offer several.
2. Upload the image and write the prompt. The fields map directly onto the settings above, sometimes with friendlier names and sometimes with a few hidden behind an advanced toggle.
3. Use the defaults for the first run. Hosts tune them to the model card. Change steps, guidance or frames only after you have a baseline.
4. Copy the request parameters before you leave the page. Most services show the exact inputs used; some give an API snippet. Paste it into your notes with the downloaded file name.
5. Download promptly. Outputs may be kept for a limited period, and the file itself carries none of the settings.

Hosted strengths: no setup, the largest variant on fast hardware, a shareable link for approvals, an API for volume. Hosted weaknesses: per-clip cost, fewer exposed controls, uploads leaving your machine, and hosts changing the version under the same name.

Settings at a glance

Keeping work reproducible across routes

The point of open weights is that nothing locks you in, but that only holds if you keep records. A simple system:

• One folder per shot containing the still, every output and a plain text file of parameters.
• Name files by seed and version, for example pier_wan22_i2v_seed1234_f81.mp4, so the recipe is half-visible in the name.
• Keep the embedded workflow on locally and paste the host’s parameter block into the text file for hosted clips. Both should list the same fields.
• Note the exact model file or host version. "Wan" is not enough; the checkpoint name or the host’s version string is what lets you match results later.
• Expect small differences. The same seed on different hardware or a different sampler build produces a near-identical but not pixel-identical clip. Treat the recipe as a starting point for a match, not a guarantee.

Try it

Still: a paper lantern hanging in a dark wooden corridor, warm glow, slight haze
Prompt: The lantern sways gently on its cord. Its warm light flickers softly on the wooden walls. Slow push in along the corridor. Static background, calm motion.
Negative prompt: blurry, low quality, distorted, text, watermark, flicker, fast motion
Frames: lowest valid count. Steps: default. Guidance: default. Seed: 2048

Run it once on a hosted service and once locally with the same values, then compare side by side. The differences are hardware and sampler, not the model; that tells you how far a hosted draft predicts a local final.