Output example of using both workflows: https://civarchive.com/images/97448768
Credit goes to DonutsDelivery & Daxamur, refer to them for the unmodified workflows by the creators: https://civarchive.com/models/664292 & https://civarchive.com/models/1853617
These are the production workflows that I use to generate the images in SDXL and then use the image for the Wan2.2 I2V videos.
My production workflow is text2img SDXL -> img2video Wan2.2
This will require installing custom nodes, model files and may require you to do manual node installs from github. Its been awhile since I've installed all the dependencies so it may be a pain in the ass but as of 8/31/2025 these workflows work.
All nodes, dependencies, and software is the most current versions.
Tested on Windows OS RTX 5090 with latest CUDA + Torch + Sage
Description
I fixed some node issues and removed nodes that were not relevant to my production.
Modified Version of https://civitai.com/models/664292
Credit: DonutsDelivery
FAQ
Comments (11)
WARNING- scaled FP8 models (on Blackwell at least) are utterly broken- use GGUF instead, and this includes Clip (the LLM). FP8 will appear to work, but gives vastly worse results - and astonishingly scaled FP8 clip models are responsible for ignoring your prompt details- I literally could not believe this when trying A-B testing. Scaled FP8 has been wasting my time for months!
Do you have a fixed version? Does this still apply as of Jan-2026? 2.9.1+cu130 (5090)
Ty!
Mate, this was an awesome workflow. I tried to amend it to get F2F working but its only taking the first frame. Any ideas what I messed up?
https://drive.google.com/file/d/1-DORCjSZ0k0tx8ZxbnxplP3n_BpyZfM5/view?usp=drive_link
@becausereasons Credit goes to the author: https://civitai.com/models/1853617 It looks like you've modified the workflow and maybe broke something. I'd just re-download and start for there.
@Kierkegaard420 I modified it on purpose, trying to add F2F (Frame to Frame) functionality to it.
try this using Dax's workflow with MOE sampler
https://ibb.co/fzJXWxwb
Best one
It's not simple, it's very complicated and difficult.
This is true. Its definitely an advanced workflow. Newbros be warned!
@Kierkegaard420 What are the optimal CFG and Motion CFG settings? I’d really appreciate the advice. I’d test it myself, but it’s really time-consuming
Very interesting workflow, however it seems broken now - maybe some of the nodes have been updated?
The break appears to be at the CFGGuider node. It requires positive and negative conditioning inputs, but there is too much spaghetti around the conditioning blocks for me to figure out which ones.