Hello community and my precious followers! 👋
I'm happy and exited to announce the upcoming release of my version "LureNoir" of my DaSiWa WAN 2.2 i2v checkpoint series.
This was a truck load of work and fun the last weeks, fine-tuning parameters, testing with over 400+ samples, situations and various prompt-refinements.
Testing new stuff, nodes, technologies...
At the beginning I started to do it just for me, making a checkpoint that fits my style and taste, easy to use and making good results without endless try and error.
The goal was to create something useful and simple, yet powerful in understanding and consistency.
A great journey so far! - I learned so much and I'm thankful of all the positive and constructive messages that arrived to me!
That even drove me to write a whole in-depth guide to WAN 2.2 ... crazy!
And doing a comfyui workflow ... even I was not using comfyui actively before
I got so much support, driving me forward 🤟
You are awesome! - All these posts, all these fantastic ideas and creative art! Plus all the supporters! Thank you! 🖤
Here are some facts about the new checkpoint:
tested 20+ situations (NSFW and SFW), with over 100 samples
good details even on low resolutions
consistent understanding of situations (as versatile as possible)
Baked with fitting BF16 VAE
Baked with CFGZeroStar
as compatible to other LoRAs as manageable
reward attention for best guessing of the situation
Excluded CLIP (again)
Better control over offloading and less memory dependent from the view of the single checkpoint if used with custom clip
smaller size obviously
Interoperated most compatible and best working lightx2v tech (4 step) in various tests
What to expect from this release:
Reduced detail transformation with anime/realistic images (lips, eyes, ears, ...)
Even better motion
It can guesstimate better details out of frame
As close to RL motion-speeds as I could get with speed-up tech
Less tries for good results
Better compatibility with LoRAs ( + tested with 💦 LoRAs)
Usable with your preferred/custom CLIP
Reduced hallucinations
What not to expect:
Some things still need a extra LoRA to be awesome, but the checkpoint will support them.
Arbitrary perfect guessing of NSFW details out of frame in every situation (WAN 2.2 is not trained on that at the end)
The perfect "splash/shoots" - This is just not a thing it can do without killing other understandings
With the release I'll also upload a fitting optimised workflow on my latest conclusions, so make sure to try it out! 😸
Remember. WAN 2.2 is not perfect, neither any merge/fine-tune will be. So there are always some bad samples. Think back of Wan 2.1 beginnings .. what a mess 😁
I hope you have fun and you can bring your ideas to life with that awesome tool! 😉
As long as it does not violate the ToS! 😡
As I'm doing this in my spare time and I'm limited on resources, consider buying me a ko-fi, if you want to support me. Everything helps :)
Release: Already on EA.
Initial upload got corrupted. This was fixed!