SkyReels Image-to-Video in ComfyUI on Low VRAM

SkyReels Image-to-Video in ComfyUI: Run It Locally on Low VRAM

SkyReels V1 is an image-to-video model built on Hunyuan Video: you give it a photo and a prompt, and it animates the photo. This guide installs ComfyUI with Triton and SageAttention on Windows using our installer, then runs the SkyReels workflow on low, medium and high-end GPUs.

First published February 2025 on the old PixelAI Labs blog. Refreshed in October 2026: the steps below are checked, the links point to the current project pages, and newer options are mentioned where they help.

Newer image-to-video models (HunyuanVideo 1.5, Wan and MiniMax H3) have arrived since, but SkyReels is still a good way to learn local image-to-video, and the installer steps apply to many video workflows.

Prerequisites

Install these first:

Microsoft Visual Studio with the C++ build tools (download). Check it works by typing cl in Command Prompt.
CUDA Toolkit (archive).
Python 3.12 (python.org).
Git (git-scm.com).

Install ComfyUI with the installer

Download the installer from the downloads and put it where you want ComfyUI.
Run the batch file. It clones ComfyUI, scans for Python versions; pick Python 3.12.
Wait for the dependencies and press Enter when asked.
Choose the Triton version: Triton 3.2 for PyTorch 2.6 or newer (recommended), Triton 3.1 for PyTorch 2.4 to 2.5.
Wait for SageAttention to install, then press Enter to finish.

Set up the workflow

Start ComfyUI with the run ComfyUI batch file.
Load the image-to-video workflow from the download.
Install the missing custom nodes with Manager → Install Missing Custom Nodes, choosing the latest versions, then restart ComfyUI and refresh the browser.

Models:

text

https://huggingface.co/Kijai/HunyuanVideo_comfy            (VAE)
https://huggingface.co/Kijai/SkyReels-V1-Hunyuan_comfy      (SkyReels model)
https://huggingface.co/spacepxl/skyreels-i2v-smooth-lora    (smooth motion LoRA)

Building an AI influencer? The UltraRealistic AI Influencer Pro course shows the complete system with today's models: design one character, train her own LoRA, and make KREA 2 photos and MiniMax H3 talking videos, on the cloud or your own GPU.

Example prompt

Describe the motion and the camera, not the whole picture again (the image already shows it):

text

the woman slowly turns her head toward the camera and smiles, her hair moves in a light breeze, handheld camera, natural motion

Speed notes

RTX 3090 (24 GB): about 20 minutes for 3 to 4 seconds of video.
RTX 4090: about 10 minutes.
Smaller GPUs with quantized models: 20 minutes or more.
The first run downloads the text model, so it takes longer. Later steps get faster after the first few iterations.
Lowering steps from 30 to 25 or 20 saves time, at some cost in quality.

FAQ

Does SkyReels work on 12 GB GPUs?

Yes, with the quantized models and patience. Keep clips short and the resolution low.

What do Triton and SageAttention do?

They speed up attention, the slowest part of video models. The installer sets them up so you don't have to compile anything.

Is there a newer model I should use?

For new projects, also try HunyuanVideo 1.5, Wan or MiniMax H3 image-to-video in ComfyUI.

Take it further

Local image-to-video is the next step after realistic photos. If you want a ready-made, tested system instead of assembling one yourself, the UltraRealistic AI Influencer Pro course walks through character design, LoRA training, KREA 2 photos and MiniMax H3 videos step by step, and the InstantClone AI Influencer Suite creates consistent images and motion videos from one character image. More free guides are on the PixelAI Labs blog.