Wan 2.2 image-to-video starts from a reference image and generates motion around that starting frame. This guide focuses on preparing that image, loading the correct native template and checking the resulting clip. It is a documentation-based walkthrough; the cover is an illustration, not a recorded FluxDraw render.
For the separate topic of dual-expert sampling and distilled adapters, use the Wan 2.2 dual-expert and Lightning guide. Establish a working standard I2V run before adding speed modifications.
Visual guide: Keep the reference image and motion goal distinct
How to use this map
- Choose the intended Wan variant before assembling files. Matching the model and VAE is part of setup, not a quality-tuning option.
- Prepare the reference at the intended framing and inspect any resize or crop. A clipped subject in the starting frame can complicate later motion review.
- Specify one observable action and restrained camera movement. Keep settings and starting frame fixed when comparing motion instructions.
- Review matching frames across the clip for identity drift and unwanted scene changes. Calculate playback duration from frame count and FPS rather than confusing it with render time.
Stop and check: A strong first frame does not demonstrate good video motion. Save the reference, prompt, frames and encoded output together.
Next practical guides: Wan 2.2 Lightning in ComfyUI: Dual-Expert Sampling and Fixes · LTX-Video in ComfyUI: Setup and Performance Checks
Implementation reference: Official native Wan 2.2 workflow. Use the matching version-specific example alongside this editorial map.
Choose I2V or TI2V before downloading
The Wan-AI I2V-A14B model card describes the 14B image-to-video model, while TI2V-5B supports both text and image inputs with a different model and VAE layout. Treat them as distinct workflows. Do not substitute a T2V expert for an I2V expert because the filenames look similar.
Load the native ComfyUI example
Use the official Wan 2.2 tutorial and select the matching I2V or TI2V template. If core nodes are missing, check your ComfyUI version and startup import errors. A native template should not lead you to install an unrelated collection of custom nodes.
| Component | 14B I2V | 5B TI2V |
|---|---|---|
| Diffusion weights | Matching high-noise and low-noise I2V experts | TI2V 5B model |
| Text encoder | Template’s UMT5 encoder | Template’s UMT5 encoder |
| VAE | Wan 2.1 VAE listed for 14B | Wan 2.2 VAE listed for 5B |
| Image input | Load Image connected to the I2V path | Enable the template’s optional image-input path |
Use the exact download links in the selected example. Repackaged ComfyUI weights and the original publisher’s Python project have different installation instructions.
Prepare the reference image deliberately
Choose one clear subject with enough space for the intended movement. Crop the image to the aspect ratio you want before starting. A tight crop can make a requested camera pullback require invented surroundings; a crowded image makes it harder to tell whether a motion error came from the prompt or the reference.
Keep an untouched original and a separate working copy. Record its dimensions. If you resize it, inspect faces, small text and thin edges afterward. Do not use a preprocessed image whose artifacts you later mistake for model-generated problems.
Design the shot before writing a long prompt
Decide what should move and what should remain stable. For a first test, request one small action and one camera behavior. An example brief is: “The plant leaves sway gently in a light breeze. The camera remains fixed. The pot and background stay in place.” This is a suggested exercise, not a prompt we measured.
The starting image already establishes appearance. Avoid asking it to transform the subject, change the scene, rotate the camera and perform several actions in the same short clip. Add complexity only after you can inspect a simple result.
Run one short baseline clip
- Select the exact model files shown by the matching template.
- Upload the prepared reference into Load Image and inspect the preview.
- Keep the template’s initial sampling and guidance settings.
- Start with a short frame count and modest output size permitted by that workflow.
- Use a fixed seed for comparisons and keep a record of the actual settings.
- Save the resulting clip and workflow together before changing anything.
A successful graph execution establishes that the pipeline ran. It does not establish subject consistency or useful motion. Watch the saved video, not only the first frame preview.
Check duration and playback separately
Frame count and playback frame rate both affect clip length. An 81-frame file played at 16 fps lasts about 5.06 seconds; at 24 fps it lasts about 3.38 seconds. This arithmetic does not specify which frame count the model supports. Follow the template’s latent-length constraints and keep the selected playback rate in your test record.
Changing playback fps alters presentation speed. Interpolating frames is a separate operation and can introduce its own artifacts. Judge the original generated clip first so an interpolation stage does not hide or confuse the failure.
Review motion with matching frames
Inspect the first, middle and last frames, then watch the clip at its saved speed. Check subject outline, hands or faces, background stability and whether the requested action occurred. A pleasing middle frame can hide a broken beginning or end.
Use the same reference and seed while changing one prompt clause. If camera drift is the problem, simplify the camera instruction first. If the subject changes shape, reduce the requested action and inspect whether the starting reference has ambiguous details. These are diagnostic steps to try, not guaranteed fixes.
Common failures and the next check
Missing models or incompatible VAE
Compare every loader with the selected template’s list. Check the model family and VAE together. Replacing one file from another variant can create a graph that looks plausible but is internally incompatible.
Out of memory
Record the failing stage, output dimensions and frame count. Reduce one of the size dimensions or clip length in a copy of the workflow. Model precision, offloading, system RAM and other processes all matter; no GPU-capacity table guarantees a particular clip will fit.
The result ignores the starting image
Inspect whether Load Image is enabled and actually reaches the I2V conditioning or latent path. Confirm you loaded an image-capable model, rather than assuming that adding an image node converts a text-only workflow.
The output is static or unstable
Check the original saved clip and actual frame count. Compare one simple motion brief with the initial settings. Keep failed clips as well as successful ones so you can trace what changed.
Keep evidence with every result
Record the model filenames, ComfyUI version, template, input image, prompt, seed, dimensions, frame count, playback fps and any adapters. If reporting speed or memory, state hardware and measurement method. This article supplies a setup and review method, not a firsthand hardware benchmark.
For video-job submission from code, see the ComfyUI API guide. For loading and validation failures, see queue troubleshooting.
