Sigrid is a busty blonde woman who started as a Flux character model (https://civarchive.com/models/1337046) and now she is ready for her debut as a video character with Wan. Note: The BustySigrid Flux model has been updated using the same training images as this Wan model. So if you want to generate images with Flux and do image-to-video, you can get consistent results either way.
Important: This model is 100% fictitious and is not based on images of any real person. She originated with images generated from still image checkpoints and has evolved from there using only AI rendering. No images (neither actual nor AI-generated) of any real person was ever involved in the making of this model.
Version 2 Update
Sigrid was the first LoRA I made for Wan and I'm finally releasing a native Wan 2.2 version. There aren't any big surprised in here, it's a pretty standard character model. This model doesn't have quite as many motions and actions built into it as some of the other models I've released lately, so it helps to add Boob Physics to at least the Low Noise steps, but that's up to you.
Basic Usage guidelines:
High Noise at 1.00
Low Noise at 0.75, the look is a little too strong at 1.00 and 0.75 looks just about right
Boob Physics in Low Noise, see below for details
A new version of the Boob Physics LoRA was released so I haven't had much chance to try it out yet. My sample videos all use the prior version (version 7) in the Low Noise at strength 1.00. I've made a few videos with the new version that's made for Wan 2.2 and it seems to be a little stronger with side effects than the older models. I would recommend no more than 0.50 strength if you use the high and low latest Boob Physics LoRA with Sigrid.
Sigrid has always been one of my top two or three favorite models, so I hope she gets a chance to star in some of your videos too.
OLDER UPDATE: For the best results, please give my workflow a try. This workflow is what I use to make all my videos using my models and it produces great results in Wan 2.2 using LoRAs that were created with the Wan 2.1 model.
ComfyUI workflow guidance
For my current favorite ComfyUI configuration, please refer to my article for all the necessary info you'll need to make the most of my Wan 2.1 LoRAs in Wan 2.2 videos.
This is a Text-to-Video LoRA, so I can't promise quality results if you try to use it with Image-to-Video instead.
Sigrid appears to mix well with other Wan LoRAs and can be rendered at different resolutions (tested at 320x320, 320x480, 480x720, 640x640, 720x720, and 640x960). Wan videos can be a little stiff, so I recommend pairing this LoRA with other motion enhancers like Bouncy Walk (https://civarchive.com/models/1361346). Some of her training images were topless, so she looks good with her shirt off, too.
For prompt tips, see the sample videos. I added the full prompt text to each video and the ComfyUI workspace is embedded in each video as well.
The Wan settings I used for the sample videos are [ edited to say - these settings are outdated and can be improved. See the ComfyUI guidance information at the top of this text for details. ] :
LoRA strength: 0.75 to 1.0
Frames: 81
Steps: 25-30 (at 20 steps, the quality was noticeably worse)
cfg: 3.0 - 6.0
Sampler: uni_pc_bh2
Scheduler: simple
I used frame interpolation to add intermediate frames just to smooth out the video. No upscaling was used for the sample videos.
I hope you enjoy making movies with Sigrid. If you make videos you like, please either post them to this page or mark this LoRA as a resource when you post your clip. I always like seeing what other people make with my models.
Description
FAQ
Comments (18)
Thanks very much to @setoror , @99Helion , and @chipbeard77977 for the tips!
Thanks for sharing Sigrid with us @darkroast175696 <3
Struggling with this one. Great looking difficult to deploy
What do you mean by "difficult to deploy"?
The clips you've posted look like they came from Civitai's video generator and to be honest I think that generator is not great. I've been making all the videos I post with comfyui on my home PC. If you have comfyui yourself, you can download any of the videos and drag them into the comfy browser session and it will show you my workflow so you can mess with it or recreate or whatever. Wan video generates at 16 frames per second by default and tends to have quick jerky motions. There are a few tricks you can use to deal with those things, but civitai seems to be using wan pretty much as-is. I'd encourage you to try making a video or two at home instead and see how it turns out. If your hardware can't go for the higher res, you can back it down to 480x480 and see how that goes.
My use of the 1.3b model has no effect at all. The parameters are CFG:6-8, setp: 30, and strength: 0.8-10.
Moreover, it was completely useless for me to use the prompt words of the example completely.
Are there any prerequisite requirements?
This model was trained at civitai using the 14b text-to-video model. I haven't tried it with the 1.3b model. So you may need to switch to the 14b version to get it to work. I've been using an fp8 reduced size version of the 14b to make it work in my video card.
thanks
My guess is this LoRA is totally the result of proper, textbook t2v LoRA training. It plays super well with any other LoRAs out there, and you can use it for img2vid, VACE, literally anywhere. I'm so curious how he put together that dataset.
Mostly it was sheer luck. :) And a bit of lora experience with flux and sdxl and such, but this was my first Wan lora and I've been really happy with how it turned out. If I could claim any skill here it was in choosing some good input images and I must have done a good job with the captions.
I agree, though, Sigrid has turned out to be really flexible. She looks good in just about every situation I've tried.
Thanks for the kind words!
darkroast175696 I have also recognised that this lora is a master of versatile usage and no klora keys dropped for wan2.2 generation...opposite to some others. I would appreciate if u could share at least the resources for training. I have trained loras in the past locally with onetrainer.
sikasolutionsworldwide709 I don't know if I have the exact set of training images I used for this lora still, but I have something close. I do have the training settings though. I used 54 images and these are the parameters from the civitai wan video trainer:
BustySigrid Wan training notes
published epoch 15
Params:
{
"engine": "musubi",
"unetLR": 0.0002,
"clipSkip": 0,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 32,
"numRepeats": 4,
"resolution": 960,
"lrScheduler": "constant",
"minSnrGamma": 0,
"noiseOffset": 0,
"targetSteps": 4320,
"enableBucket": true,
"networkAlpha": 1,
"optimizerType": "AdamW8Bit",
"textEncoderLR": 0,
"maxTrainEpochs": 20,
"shuffleCaption": false,
"trainBatchSize": 1,
"flipAugmentation": false,
"lrSchedulerNumCycles": 1
}
darkroast175696 Ty...Did u use the 1.3B or 14B checkpoint for training?
sikasolutionsworldwide709 The Civitai Wan trainer uses the 14B T2V model for training.
I have used this nice lora for my Wan2.2 generation. It looks very natural but I would prefer bigger bazookas. WF Wan2.2 is included in the video
I haven't even tried out wan 2.2 yet. I'm glad to hear this LoRA works with it! Have you posted the video? I'd like to see what it looks like.
darkroast175696 Yes its in the gallery somehow twice. The Lora loader in my current workflow has required to convert ur lora from musubi tuner format to comfy-ui compatible otherwise the keye are not readable but stackloraloader should work with the musubi format. Gentime is insane 20 mins for 3 sec vid comparing to fusionX recipe Wan2.1 5sec vid 386secs. so I am now testing speedup opps
@sikasolutionsworldwide709 Your wish has been granted.
A Hunyuan version would be greatly appreciated. FramePack F1 allows for very long video generation on smaller cards. Think about it!