You have ever tried creating smooth transition videos, you know the struggle. You want your clip to begin with one scene and naturally flow into another, but the tools out there either give you choppy results or require hours of manual editing. For creators, this means wasted time and a final output. Now, this can be solved using Fun Inpainting with Wan2.2 model released by Alibaba PAI team.
You upload a starting frame and an ending frame, and the model fills in the gaps that smoothly, intelligently, and in film like quality. It puts creative control back in your hands without the headache of traditional editing.
This supports multiple resolutions like 512x512, 768x768, and even 1024x1024. It's released under the Apache 2.0 license, which means not only is it open source, but you can also use it commercially without worrying about restrictions.
Installation
1. First, update ComfyUI from the Manager by clicking on Update All to avoid any custom node errors.
Then, do the Wan2.2 Image to Video(I2V) Setup in ComfyUI. You can also use Wan2.2 I2V with LightX2v LoRA model to do faster generation.
2. Now, download any of the pair of models(high and low) from the list given below. Choose the pair as per you system requirements.
Official Fun Inpaint models by Alibaba Pal Team (Base variant) need high end VRAMs
(a) Wan2.2 Fun A14B Inpaint High Noise (diffusion_pytorch_model.safetensors)
(b) Wan2.2 Fun A14B Inpaint Low Noise (diffusion_pytorch_model.safetensors)
Compressed by ComfyUI Team (BF16 variant) for high Vrams
(a) Wan2.2 Fun A14B Inpaint High Noise BF16 (wan2.2_fun_control_high_noise_14B_bf16.safetensors)
(b) Wan2.2 Fun A14B Inpaint Low Noise BF16 (wan2.2_fun_control_low_noise_14B_bf16.safetensors)
Compressed by ComfyUI Team (FP8 Variant) for low VRAMs
(a) Wan2.2 Fun A14B Inpaint High Noise FP8 (wan2.2_fun_inpaint_high_noise_14B_fp8_scaled.safetensors)
(b) Wan2.2 Fun A14B Inpaint Low Noise FP8 (wan2.2_fun_inpaint_low_noise_14B_fp8_scaled.safetensors)
Rename them to some thing relative that will be easy to differentiate.
Save them into ComfyUI/models/diffusion_models folder.
3. Restart and refresh ComfyUI.
Workflow
1. Download the Workflow -Wan2.2 Fun Inpainting (Wan2.2_14B_Fun_Inpainting.json) from our Hugging face repository.
Alternatively, you can get the workflow by moving to ComfyUI dashboard from Workflow (top left corner) > Browse Template> Video > Wan2.2 Fun Inpaint and Wan 2.2 Fun Control.
2. Drag and drop into comfyUI. The workflow has two types-
(a) With Lightx2V lora models for faster generation
(b) Without Lightx2V lora models for high quality generation. Choose as per your use cases.
3. Load the Wan2.2 fun Inpaint high and low models. Remember to choose the similar pair. Load Lightx2V lora models.
4. Load your first and end frame (image) into Load image node.
Here, the first and end frames converted to the latent data that again transfers to the KSampler. By default, into the WanFunInpaintToVideo node the frame sizes are set to smaller for low VRAM users.
You can change to higher for better quality if having adequate VRAMs.
To get the natural effect double the FPS with Frame interpolation.
5. Add positive prompts, what you require for the video motion. Make sure you put the long detailed prompting to get the perfect output from it. More detailed instruction for Wan2.2 prompting can be found in our Wan2.2 installation guide.
6. Click Run to start the execution.
Let's say you are working on a product ad. You can start with a close up shot of the product (start frame) and end with a lifestyle shot of someone using it (end frame). Wan2.2 Fun Inpaint will generate all the connecting frames, giving you a smooth, cinematic transition something that normally take a professional editor hours to achieve.You know what we mean.
Or maybe you are into storytelling. You can illustrate a character in two different scenarios, feed the start and end frames, and watch the model to generate the in between journey as video. It's like having a mini-animation studio at your fingertips.
