Greetings travellers! Stay awhile and listen!

DasiWa WAN 2.2 "TastySin" arrives!

A ritual was performed. Summoning the most delicious dishes... topped with the finest ingredients. Tainted with a pinch of forbidden desire. Do you dare to try it?

Do you remember the big news about "MidnightFlirt"? ... How much improvement it was? ... That ... that is now an old story.

This soooo~ sweet smelling, alluring, warming and almost forbidden "TastySin" will arise!

(https://civitai.com/models/2190659)

I felt the need to get it done before Christmas. Think of it as a little Christmas present. 🎄 I'm happy and extremely excited to tell you more! Just read ahead! ~ I hope you can enjoy this like I do and that this will enable even more people to use the tools to create their art!

Here comes a new fairytale to tell! With a plot twist!
I was able to improve my methods of refining and merging so much! Also with the help of many nice donations and submissions it was possible for me to bring the DaSiWa WAN 2.2 I2V model to a whole 🌟🌟🌟new level 🆙 🌟🌟🌟

Again, I collected valuable feedback and listen. Elaborated with other awesome creators and many people on my fresh launched discord server. So I was able to adopt things even further. Well not anything is perfect, but WAN 2.2 will probably never be. But can it be more awesome?

~ Yes! ✨

Remember I told you that a gguf quant is trash for my checkpoint? Well, that was totally valid/right with the FP8-scaled checkpoint processing. I got so many requests about GGUF, compatibility to older hardware and support for low-end systems. With an FP8 base it was just not possible to do while maintaining my quality standards.
But with tools, knowledge and support evolving ... 🥁 ...

Here comes the plot twist! ~ I'm so exited to say this ... really excited!!! ~ That does not apply any more with a full checkpoint processing with FP16 precision and all my optimisations!

So, 🎉 ... GGUF becomes valid and ensures quality, more compatibility and to an extend support for GPU-poor systems!

Well this was a huge step for me and a steep learning curve with tools and dependencies.

Oh ... let's go a bit technically, right?

Lets break it down to the following:

High quality checkpoint in Q8 GGUF, like full FP16
Compatible with older GPU's
Quants down to Q2 (with equivalent quality loss)
No 'fake' GGUF dequant from FP8 (made by others) needed

Well, because I'll provide most quants (Q8 till Q2), I'll have to do a new approach of releasing it. So I'll make a new model page just for that release with all the quants inside.

I'll tag the model pages with GGUF with (G) and the model pages for safetensor with (S).

Here are some technical facts about the new checkpoint:

Only open-source resources
Fast, with only 4 steps - lightx2v r64-1022 distillation (latest/best)
As always tested with many situations (NSFW/SFW/realistic/anime)
Full FP16 base
Baked with CFGZeroStar
High compatible with LoRAs
Baked with reward attention for best guessing of the situation
Excluded CLIP
- Better control over offloading and less memory dependent from the view of the single checkpoint if used with custom clip
- smaller size obviously

🧿 Zero prompt i2v are possible! The understanding is really awesome! Try it! 😉

What to expect from this release:

Superior details with anime/realistic images (lips, eyes, ears, breasts, genitals, ...)!
Natural and smooth motion
High success on guesstimating details
Better results (compared to MidnightFlirt)
Better compatibility with LoRAs (compared to MidnightFlirt)
Usable with your preferred/custom CLIP (if compatible)
Reduced hallucinations (compared to MidnightFlirt)
Upgraded prompt adherence (compared to MidnightFlirt)
Better texts (compared to MidnightFlirt)
Better camera control (compared to MidnightFlirt)

Because examples show more than 1000 words:

See for yourself - the release will showcase:

Examples ~

in various situations
SFW and NSFW
Anime, semi-realistic, realistic (illu)
Some with zero prompting
All samples without any LoRA
Initial images made with my Illustrious checkpoint
Made with my workflow
All showcases are 1st try (on various WAN 2.2 resolutions), to showcase the capabilities!
(This time also the text front-video!)
All include metadata/workfow

Comparison of Q8 to Q2

Comparison examples for the Q8 till Q2 quants
The comparison is below ⏬
Settings
- Same prompts, seeds, initial images
- Metadata included
- frames (97) and resolution (medium 608p) to showcase the not native capabilities and prompt adherence on average settings with not just standard 81 frames
- Complex scenes
- anime style, realistic style

As always this is meant to be as versatile as possible, but enhance the understanding of WAN in NSFW and not trained situations.

Want to make NSFW? Realistic? Semi-Realistic? Anime? ... all in! 💯

The goal remains to create something useful and simple, yet powerful in understanding and consistency.

What not to expect:

The checkpoint is not meant to replace every LoRA-Concept out there, it will be great on its own and even more awesome with a specialised LoRA!
So make sure to support the awesome LoRA-Creators!
Arbitrary perfect guessing of NSFW details out of visibility in every situation (WAN 2.2 is not trained on that at the end)
To be made for concepts that are against the ToS like "nudifying" real individuals. If this works out by coincidence, it would simply be a basic understanding of WAN 2.2.
Guessing the lower genitals in every situation and style, you may need a LoRA. (Forcing this into the model results just in a total bias to the chosen gender)
To understand every situation and NSFW concept possible

Workflow

There is 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.

Support and early access

I hope you have fun and you can bring your ideas to life with that awesome tool! 😉

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. Maybe Buzz / EA, everything helps :)

All checkpoints will be available through EA and free after. I'll lower the buzz tag from previous releases to about 50% to give more people the opportunity to have access, have fun with it as soon as possible (Happy Christmas time!🎄) with this option to support me if desired, but fully available before Christmas.
Besides the amount of work, tuning, testing, and spend resources was huge on my end.

💡Price hint: Even if someone intends to use the buzz, civitai will likely cut about 50-60% off!

In the end it will be your decision to spend some buzz and support further development or wait till the EA ends.

Comparison Q8 to Q2

Anime style, 97 frames, 624*912, prompt:

A woman wearing a black wedding dress with blonde-pink hair, light-blue eyes, and a blue rose in her hair.
She confidently touches her hips, tracing her hands over her waist.
She swiftly turns around to swing up her skirt.
With the swift turn the black wedding dress becomes a white wedding dress.
The blue rose turns into a red one.
She sways her hips, moving her dress to emphasise her beautiful figure.

Realistic style, 97 frames, 624*912, prompt:

A woman wearing a black, red trimmed gothic clothes. 
It is raining, her hair and clothes clinging wet on her body.
Her black make-up is running from her face like tears.
She turns around, facing her back to the camera. 
She walks away, down the street. 
The camera is following her in moderate speed.

Anime style with text, 97 frames, 624*912, prompt:

She is a woman with dark red eyes and slit pupils, black-red hair, cat ears and a cat's tail.
A glowing, dark-red, slowly spinning halo is above her head.
She touches her lips, tracing them with one finger. 
Soft, glowing, dark-red text appears on the monitor: 'DaSiWa TastySin', surrounded by glowing heart symbols.
Playfully, she licks her finger upwards.

Closing words and sources

💌 Spacial thanks to the resources, videos, and works of all these awesome creators!

Make sure to check out their stuff! ✨

Listed resources may be partly used or part of the process to create that checkpoint if open source or free of use:

Upcoming release of DaSiWa Wan 2.2 i2v "TastySin" Lightspeed | GGUF