🗡️💀 DaSiWa-WAN 2.2 I2V 14B Lightspeed | FP8 Safetensors💀🗡️

No, sorry for the confusion. It is I2V, corrected that.

Sometime it helps to pause the download and then start it again.

@Hellborn I tried, but did not work. :(

In case someone needs it: I found a solution using a GPU and Google Colab to bypass Civitai’s download speed limits.

just copy paste this promt:

Hey ChatGPT, I want to download large models from Civitai without the download speed dropping to a crawl. I’d like to use Google Colab so it downloads directly to Google Drive or Colab’s temporary storage.

Please create a ready-to-run Colab notebook that:

Optionally mounts Google Drive,

Installs aria2 for fast parallel downloads,

Downloads the model using multiple connections to avoid throttling,

Saves the file with a defined name in Drive or Colab storage,

Includes comments and instructions, so I only need to enter the model URL and my Civitai token.

Make it clear that the user needs their own token.

Maybe civitai corrupted my uploads again, I have to try to download myself. This issue was also with the first upload of the last model...

@darksidewalker Amazing work, this merge superb. Super smooth and perfect details. Awesome. Congrats, and many thanks!

@H_for_Hi thank you!

I don't really know. Depends on your backend, swap and other settings.

rtx 5070 (12gb) + 64gb RAM works normal.

@saymannsk330 which workflow you are using?

@DarkAmbassador I have my own workflow

They are improvements over each other, my versioning is with names. I put the description only on high, because one need both anyways.

Yes it is possible. But the outcome is variable like with every WAN 2.2 14B model. For exact movements there are other models that mimic sound or other videos, but they are not so easy to use atm.

Mate, this is not how it works. I can not just add any and endless LoRAs in. But you can add the LoRa just fine with strength 1.0 for high and low. It works.

Here is an example, with this lora: https://civitai.com/images/103516282 - I did FLF and chained 2 videos together, really decent.

No

will you do T2V in the future?

@dehyilo397 No, not planned

I always recommend SwarmUI.

That is a problem with your settings, its just a Wan 2.2 checkpoint and therefore the same requirements as others. If you OOM its a setting or misconfiguration.

Did you tried a base Wan 2.2 and compared the results?

@darksidewalker Yes. the same prompt is heard better on the base model. We can discuss it on telegram or in or in PM, maybe I’m doing something wrong.

@saymannsk330 Yeah, would like to know if this is possible. For testing. I'll PM you, in my testing the model was much more prompt stable than base for nsfw.

It is fp8, unfortunately the setting is not available on upload to reflect that. All Wan 2.2 fp16 would be 29+GB in size.

if you took 128gb ram you ll never regret, sometimes my PC uses 90 gb ram when i generate videos

My computer is quite weak (32GB DRAM and an RTX 4060 8GB VRAM), so I enabled the Windows virtual memory setting to 96GB (set on a Gen4 SSD). Everything is good.

@PotatoChipO3O Smart :)

Nice to hear, have fun with it! :)

Were you able to fix the low quality and graininess? :D I have the same problem now

I seem to have found a solution, which is to use multiple lora, and connect the lora loader in series after the checkpoint loader, which works. As for the particle problem you mentioned, it may be caused by the low resolution. The video I generated has basically no graininess.

This is not a t2v checkpoint, you need a init image. I did not notice any gender changes without prompts. Do you have an example? If your videos are low res you will get grain or artifacts, that's a thing with all wan 2.2.

@darksidewalker Hi, I'm using I2V to generate videos, but I'm finding it doesn't always follow the prompts. For example, I have a picture of a woman standing up, and I'm trying to write prompts for a man and a woman having sex, as well as various sex positions, but it's mostly just the woman vibrating on her own. I noticed you mentioned a lot of LoRa in your introduction. Do I need to chain LoRa in?

@guotiancai You not need much extra LoRAs here, only 1-2 for specific styles that the checkpoint or wan 2.2 is not capable of. The other problem may be your prompt, wan 2.2 will not just spawn new persons in with I2V, except you use a Lora that enables that. If the person are just vibrating in motion WAn 2.2 and the finetunes are just not capable of generating that motion on its own, because it does not know it exactly. So ether be more precise with the prompt or use a LoRA that enables that specific motion. The checkpoint does not know anything on its own, that's not possible, it just understands composition better as basic Wan 2.2

@darksidewalker Thanks for your reply. I have another question. When I ask an independent woman to do some simple actions, such as stripping naked, it often generates a semi-finished male penis or semi-finished male testicles in the woman's genital area. I don't know why this is. I have emphasized the perfect pussy and perfect vagina, and I have added transsexual, penis, testicles, and male genitalia to the negative prompt words, but this phenomenon still occurs frequently.

@guotiancai Negative prompting does not work on speed-up checkpoints/loras, because of cfg1. Noted that in the description. Maybe that checkpoint is not suitable for that what you want, did you try another one or basic wan 2.2?

@darksidewalker Me too

@guotiancai I also encountered this situation and am adding lora to solve it.

@jevencheung Did you solve it? I still can’t solve it, I don’t know where to put the lora node

Could you PM me the prompt that forces the model to transgender? I'm not able to reproduce that situation on my own. I could have a look into that with my next release.

@guotiancai I saw your reply saying that Lora can solve the problem, but I haven't solved it yet. I set the intensity of Lora to 2 to barely remove the female penis, but the whole picture was ruined by Lora.

I really do not understand the lora talking here, did you use the full checkpoint as lora or did you use a lora to achieve a specific outcome added to this checkpoint?
LoRAs with 1.5+ or higher will most likely burn in and produce trash at all, no matter the lora.

@darksidewalker This model automatically creates a female penis when converting from an image to a video, but I didn't include any transgender content in the prompts. I saw the original poster suggest that installing LoRa behind the model could fix this, so I tried that, but it didn't work. Only when LoRa intensity reaches around 2 does it remove the female penis, but as you said, any LoRa intensity above that level essentially ruins the video.

My prompts are simple: the camera zooms out to show a woman's full body.The picture is prepared by myself, and I want to use this model to expand the picture.

@jevencheung I'll try that prompt. Did you use a anime style or realistic image?

@darksidewalker Realistic, upper body of a naked woman

@jevencheung One thing to add, with such short prompts there are many room for WAN 2.2 to just guess things up. More descriptions = more control over details.

I tested 20+ generations at both low and high resolutions (multiple resolutions) and none of them displayed incorrect genitalia for women. I only used your sentence. The downside was that it was guessing what the hands were doing and what is on the ground. The camera movements was always a pen, instead of a zoom. Sometimes she did nothing, sometimes she played with flowers or rubbed stones, and sometimes she moved her fingers nervously.

You should really use more than one extreme short sentence. The model just does not know what to add if it is not in the image. On my end, it never added anything abstract, like the wrong genitals. In almost all samples she even was not nude, wearing stockings or panties. Your settings must be wrong somehow.

Example: https://civitai.com/posts/23238293

@darksidewalker I'm sorry for not replying to you earlier. I just came online today. Could you share your specific workflow with me? The workflow I used was downloaded from your model introduction. I wonder if the difference in our workflows may have caused the different generation effects.

@guotiancai At this time I was just using SwarmUI i2v. Nothing spacial. If you like you can try my WF and a recent checkpoint, I never got this problem.

@darksidewalker Okay, thank you very much.

More steps will do not much for you on high resolutions, maybe on low. You can elaborate that by just testing. I do not know what that "speed" is. The only valid settings for speed in WAN 2.2 are fps and frame-count.

This is not a Lora, this is a full checkpoint, you can't mix it with other full checkpoints.a

@darksidewalker Ahh, I saw Lightspeed in the title and immediately thought of lora, thanks

It is FP8, I just could not choose that option. The size of ~15gb reflects that.

Obviously it is not a WAN 2.1 VACE model, it is a WAN 2.2 model. There is no official WAN 2.2 VACE. Your best bet on something similar would be WAN-Animate at the time writing.

@darksidewalker official not, but Fun Vace for 2.2 works just good enough. And it has separate module for WanVideoWrapper, just like Wan 2.1 VACE had.

It's not a complain, don't get me wrong, you did a great job. It's just pity I can't use despite working with Illustrious style NSFW videos most of the time.

@orficks Yeah, I just wanted to say. No offence ether. Just wanted to clarify that. WAN-Animate could be the next cool v2v thing, but it is too early to say.

@darksidewalker is it possible ot extract all distinctive things of this checkpoint as a single lora in a same way FusionX did with Wan 2.1?

@orficks Possibly not atm. There are no tools for that, as far as I know.

Than just do not use the meatspin lora 🤣

@darksidewalker you been training these on cookery videos lol? Does the same for me even with no lora and an empty prompt. So weird. Must have got my comfy spaghetti wired up wrong somewhere. Noticed the vids don't have comfy workflows embedded, are there any you'd recommend to use with these models?

@thejimreaper686 The videos are not done with comfy, they are done with swarmui, so no workflow embedding. If you want NSFW in WAN 2.2 you have to use nude loras and content and explicit actions, or it will not understand. Plus no prompt makes absolute no sense with WAN.

@darksidewalker But Swarm UI is just a UI, it uses a backend from Comfi for generation, will the result really be different in Swarm?

You need to use it as Image to Video (I2V), there are plenty workflows and you could also use my basic preset/workflow.

you need to download two models - high and low (Focus on the preset, the most convenient to understand is workflow For comfyui, but there is no way to configure some parameters, unlike the preset for swarm)

You can edit every parameter in every workflow/preset out there. None of them are static, you just need to know what you are editing.