Wan 2.2 Image to Image: Does It Work, How to Do It, and Better Options (2026)

Wan 2.2 has no official image-to-image mode. It is an open-source video model from Alibaba. People get stills out of it by setting the frame count to one, and they get image-to-image by feeding their own picture in as the starting latent with a partial denoise. It works in ComfyUI or through a few hosted img2img endpoints. To change an existing picture in a browser instead, use a Wan model that was built for images (Wan 2.5, 2.6 or 2.7 edit). Use Wan 2.2's image-to-video mode when you want the picture to move.

This guide is up to date as of 23 October 2026. It explains what "Wan 2.2 image to image" really means, how the ComfyUI route works at a high level, where it falls short, and what to use when you just want results.

What is Wan 2.2, and is it an image model?

Wan 2.2 is the second-generation open-weight video family from Alibaba's Wan team. The weights and inference code were released on 28 July 2025 under the Apache 2.0 licence, according to the official Wan2.2 GitHub repository. The official variants are all video tasks:

None of the official tasks is "text to image" or "image to image". The repository documents video generation only. The image use is a community discovery: a video model trained on huge amounts of footage turns out to make convincing single frames as well.

Why do people use Wan 2.2 for image to image?

Three reasons keep coming up in community write-ups:

• Realism. The model has seen a great deal of real camera footage, so skin, fabric and natural light often look less "AI-polished" than they do from some image models.
• Large native frames. One tester at Diffusion Doodles reached a latent resolution of 1920 × 1536 on a 16 GB card. That reduces the need for a separate hi-res fix pass.
• A permissive licence. Apache 2.0 weights can be run locally, fine-tuned with LoRAs and used commercially, within the licence terms.

The same tester was frank about the downsides. The portrait looked "more like a still from a video rather than a photograph". Subjects sometimes looked pasted onto the background. The ratio of good images to bad ones was low, and renders took around 170–200 seconds even with speed-up LoRAs. His conclusion was that Wan 2.2 did not clearly beat Flux for pure image work.

How does Wan 2.2 image to image work in ComfyUI?

ComfyUI supports Wan 2.2 natively. The ComfyUI Wan 2.2 tutorial lists the files: the 14B text-to-video model comes as two diffusion models, wan2.2_t2v_high_noise_14B and wan2.2_t2v_low_noise_14B, plus the umt5_xxl text encoder and the Wan 2.1 VAE. The 5B hybrid model is a single file with its own wan2.2_vae, and ComfyUI says it "should fit well on 8GB vram" with native offloading.

There is no official img2img template, so the community workflows share a common pattern. At a high level:

Step 1: Load the models

Load the high-noise and low-noise 14B text-to-video models (or the 5B model on a small GPU), the text encoder and the VAE. The high-noise expert lays out the composition in the early steps. The low-noise expert refines detail in the later steps. The repo describes this as two roughly 14B experts, with 14B active at each step.

Step 2: Encode your picture instead of starting from noise

For text to image, workflows use an empty video latent (EmptyHunyuanLatentVideo) with the length set to a single frame. For image to image, you replace that empty latent with your own picture, run through VAE Encode, so sampling starts from your image.

Step 3: Sample with a partial denoise

The denoise value (called strength on hosted APIs) decides how much of your picture survives:

• 0.15–0.3: a light pass. It cleans up skin, texture and light while keeping the composition. This is the "realism refiner" use.
• 0.4–0.6: a restyle. The pose and layout stay roughly the same, but surfaces, colours and details change.
• 0.7 and above: most of the original is replaced. You are close to text to image with a loose layout hint.

Many people run only the low-noise model for img2img. That is the expert trained on the late, detail-level steps, and the early layout work is already done by your input picture. The RunComfy Wan 2.2 image pack uses the same idea in its "low-only" mode. It generates at 1024 × 1024, upscales with 4x UltraSharp, scales back by 0.5x, then runs a 15-step KSampler pass at 0.2 denoise to refine the result.

Step 4: Decode and save one frame

VAE Decode the latent and save the single frame as a PNG. Speed-up LoRAs such as LightX2V cut sampling to about 4 steps per expert, at some cost to consistency.

Can I run Wan 2.2 image to image online?

Yes, through a hosted endpoint. fal.ai lists a Wan 2.2 14B image-to-image API priced at $0.05 per image. It has a strength parameter (default 0.5, where 1.0 means "fully remake" and 0.0 means "preserve original"), 27 default steps, and separate guidance values for the two model stages. That is a developer API rather than a consumer app.

Kenerate does not run a Wan 2.2 image model. On Kenerate, Wan 2.2 is a video model. What we do run are Alibaba's own Wan image models, which were trained for still images and support editing with your uploads as references.

Which Wan model should I use for image to image?

Here is how the options compare as of 23 October 2026. Kenerate prices come from our model configuration; 30 credits equal $0.10.

The real difference is in how the model treats your picture. ComfyUI img2img is noise-based: the model re-draws your picture, and the denoise value decides how much it changes. The Wan 2.5–2.7 edit modes are instruction-based: you upload the picture and describe the change ("same woman, same pose, now in a rain-soaked neon street at night"). The model keeps what you tell it to keep. For most everyday edits, the instruction-based route takes fewer tries.

How do I edit an image with a Wan model on Kenerate?

1. Open the image generator at kenerateai.com/app/image and sign in. New accounts get 25 free credits, enough for two Wan 2.7 images.
2. Pick Wan 2.7 (or Wan 2.6 / Wan 2.5) from the model list.
3. Upload your picture as a reference. Wan 2.7 takes up to 9 images, so you can combine a character from image 1 with a location from image 2.
4. Write what stays and what changes. Start with the parts to keep ("same face, same outfit, same framing"), then the change (new background, time of day, art style, lighting).
5. Choose a size and generate. Wan 2.7 supports sizes up to 2048 × 2048 and has a thinking mode for harder compositions. Use the seed to repeat a result you like with small prompt tweaks.

For more on how image to image works across every model we run, including the ones on the free plan, see our image to image AI guide. All of Alibaba's Wan models, image and video, are listed on the Wan model page.

Can Wan 2.2 turn my image into a video instead?

Yes. That is what its image-to-video mode was built for, and it is the most reliable way to use Wan 2.2 with a picture. On Kenerate, the Wan 2.2 page runs it in the browser:

Both take 16:9 or 9:16 and need your image as the start frame. For prompt wording that keeps the first frame stable, see the Wan 2.2 image-to-video prompt guide. To run it locally instead, our how to use Wan 2.2 guide covers both routes. If you are choosing between generations, Wan 2.2 vs Wan 2.7 vs Wan 3 compares them side by side.

What are the limits of Wan 2.2 image to image?

• It is a workaround, not a feature. Results depend on community workflows, which change often and can break when ComfyUI or a custom node updates.
• Hardware. The official repo's single-GPU command for the 14B models needs at least 80 GB of VRAM. Consumer cards need FP8 or GGUF quantised files plus offloading, or the 5B model.
• Video-still look. Testers report frames that look like movie stills rather than photos, along with weak background depth and occasional anatomy errors such as a doubled head.
• Two models to manage. The 14B pipeline uses separate high-noise and low-noise files, and LoRAs often need to be applied to each at different strengths.
• Slow iteration. Minutes per image on a mid-range GPU, compared with seconds on a hosted image model.

Is Wan 2.2 image to image worth it?

It is worth it if you already run ComfyUI, like experimenting with LoRAs, or want a free low-denoise "realism pass" over images from another model. Running only the low-noise model at 0.2–0.3 denoise is the most useful trick here.

It is not worth it if you just want to change a picture and move on. A Wan model built for images does the same job in a browser, with plain-language instructions and no GPU. Use Wan 2.2 for what it was made for, which is turning your picture into motion.