I'm using this article to share ComfyUI workflows I use with others. I'm trying to keep them up-to-date with what I'm actually using - running locally. These are designed to work on low-end hardware - I run them on a low-mid range gaming laptop (ASUS ROG Zephyrus G14 - model no: GA403UV) with a 8GB VRAM RTX4060 GPU, 16GB of system RAM, and whatever SSD the laptop shipped with. Windows 11.
Last Update - 2026-03-13
Updated list of what's attached:
WAN-FLF2V.json - WAN 2.2 Image-to-Video (I2V) using First-Last-Frame-to-Video feature
This workflow will create 5 second videos that start and/or end with a provided image. WAN2.2 generates video at 16fps, so I'm using "FILM VFI" to interpolate that up to 32fps. I find that I can get somewhere around 480x848 resolution before I start running into OOM crashes. It also has previews for the first and last frames of generated video, so you can optionally save those and use them in another run to extend video in either direction. I use ffmpeg from the command line to join videos together when they are done - that's how I create longer videos.
This runs the Q4_K_M GGUF of the WAN 2.2 14B model - it consists of 2 models that you run in succession, high-noise first, then low-noise. The workflow uses the lightx2v model distillation LoRAs to provide low-step generation. I'm using the latest '1022' builds of those LoRAs that don't exhibit the slow-motion problems earlier builds did (so no need to do a 3-phase run as long as you prompt enough action to fill 5 seconds, you shouldn't see the slow-motion effect. Adding other LoRAs for certain repetitive movements like dancing or walking or... - if that's what you're looking for - also helps - I'll include a link to a sample video generated with no additional LoRAs applied.) This also assumes you have sageattention installed - if you don't, you can bypass the "Patch Sage Attention KJ" nodes.
Model files and links (place under ComfyUI directory):
models\loras
models\text_encoders
models\unet
models\vae