In Norse mythology, Ratatoskr (Old Norse, generally considered to mean "drill-tooth" or "bore-tooth") is a squirrel who runs up and down the world tree Yggdrasil to carry messages between the eagle perched atop Yggdrasil and to the roots of the tree to Nidhogg.
WAN 2.1:
This WAN 2.1 Model is trained based on parts of the Ratatoskr Helheim / Flux Database
its made to animate Ratatoskr image outputs
so the model is better with animate , Animals , Anthro , Furry and Fantasy Creatures than the standard model.
the model can do SFW and NSFW
the model can handle different resolutions between 480p and 720p (even slightly above)
WAN V1.0:
I also recommend using any Ratatoskr based Pictures to generate Videos with it.
WAN 2.2 Hybrid V1.0:
This is an Hybrid model between parts a customized Wan 2.2 (WAN2.2-14B-Rapid-AllInOne T2V V1) and the V1 Version of Ratatoskr WAN 2.1
It's made with specific block merging and for now an wip model i will sure try also train in the future.
Like in the original WAN2.2-14B-Rapid-AllInOne Model the refiner is merged inside the model
so you can generate the image in with one ksampler but it needs 6 Steps instead of the 4 steps of the original Wan 2.2 Custom model and other Settings:
6 -10 Steps
Sampler Recommended: Uni PC BH2
Scheduler: Beta or sgm_uniform
cfg between 1 and 1.2 is tested
Overall this model is better with anthro and furry creatures than original also better in the nsfw area than standard WAN 2.2
Note: if you like T2V i made a Lora for that you can download and test it here: https://tensor.art/models/872785610920689931/Ratatoskr-WAN-Alpha-T2V-WIP-V0.4-WIP
There is also an Lora that i have build before i trained the model. i recommend to check it out.
the Model is trained at around 2700 pictures and 230 Videos.
if you wanna support my work please leave a like.
Description
This is an Hybrid model between parts a customized Wan 2.2 (WAN2.2-14B-Rapid-AllInOne T2V V1) and the V1 Version of Ratatoskr WAN 2.1
It's made with specific block merging and for now an wip model i will sure try also train in the future.
Like in the original WAN2.2-14B-Rapid-AllInOne Model the refiner is merged inside the model
so you can generate the image in with one ksampler but it needs 6 Steps instead of the 4 steps of the original Wan 2.2 Custom model and other Settings:
6 -10 Steps
Sampler Recommended: Uni PC BH2
Scheduler: Beta or sgm_uniform
cfg between 1 and 1.2 is tested
the comfy workflows i used can be found in the videos download them and open them in comfy
Overall this model is better with anthro and furry creatures than original also better in the nsfw area than standard WAN 2.2
FAQ
Comments (64)
Can't this be just a LORA?
In this case definitely no the wan 2.1 version is a trained model not a Lora. The wan 2.2 is a hybrid version that's based in the mentioned Models. If you need a Lora please visit my Furry Enhancer Video.
freek22 hey, how do i go about using this? could you share a workflow or quick guide? thanks for your reply in advance
reikoawoo789 all workflows are embedded in the videos
which VAE do you reccomend?
much thanks
Can you upload a comfy workflow for WAN 2.2 Hybrid I2V please ?
well in any of the gallery videos is the comfy workflow in metadata just drag the video into comfy
@freek22 can we use this for i2V? I get the t2v workflow loaded ok, but what about i2v?
@grasshopper85116 you can use this workflow or better the new one of https://civitai.com/models/1782485/furry-enhancer-video (in the training folder )
@freek22 That is the LORA version, what about this Ratatoskr WAN 2.1 / WAN 2.2 Hybrid I2V where is the WF for this?
@grasshopper85116 in theory it's in the gallery videos metadata if it's not working I upload it separately when I'm back from work
@freek22 please do thank you
@grasshopper85116 i added a new workflow like a mentioned to get the old one just download any video of the original gallery post and put them into comfy
Does this model go in the High or Low spot for Wan2.2?
hi in both trained a low and a high variant because its a model that does both at once for workflow download one video and put them into comfy
@freek22 so we put it in both high and low?
@zombycowvonzombe7433 no this model dont have high and low it works in high and finish the video in one go
What text encoder node/model do I use for this? I can't figure it out. Was converting a wan2.1 workflow into wan2.2 but that part is saying it's incompatible with your model.
https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main and for the vae the wan2.1 vae you can also download a video and put into comfy for the workflow
Can someone share t2v workflow?
well this version is I2V but i am currently working on a t2v version when it is released i will provide a workflow for t2v
@freek22 Can't wait very excited for that I can't seem to get the hang of wan 2.2, other version seem to be more simple. I hope it goes well for you!
@orisio actually it did the model is out of training and is in testing i plan to release it asap
@freek22 I'm wondering how is the process going of the workflow, maybe I miss it? O:
@orisio if you mean where you can find the Workflow its inside the picture or in the training folder attached to the model
@freek22 Oh but it's not text to video :(
@orisio because this lora here is i2v but for example i have the furry enhancer with a t2v versionhttps://civitai.com/models/1782485?modelVersionId=2144066
I was surprised, but it works on my GeForce 3060 12 GB and 16 Gb RAM, although generating one animation takes about half an hour.
Use the LightX2V lora, it will cut that time down to about 3 minutes
Hey, Freek! A legend as always. Would you consider creating some GGUF Quantized versions of your models?
thanks ;-) i would but didn't found a script yet for conversion
This would be amazing, still thank for all your work
@freek22 Would this by chance work? I confess I'm woefully ignorant of these things
https://github.com/city96/ComfyUI-GGUF/tree/main/tools
@atrociraptor @freek22 I'm testing this conversion now, however I have no idea what I'm doing really. it did convert to GGUF, I'm running the fix, and next I'll try to quantize it.
I read the readme a bit wrong, I'm going to try again later
I don't have the VRAM for this as is.
@atrociraptor I'll be damned, that gave me a Q2_K_S version.
I was able to use this within 8 gigs of vram.
I just had to do the fix after quantizing it.
@door_zekor Well I'm happy to have found that for ya then :)
@door_zekor if you do that please link to the used original models ;-)
@door_zekor i will test that when iam back home sadly iam since 4 weeks in hospital and come back home in 1 week to test that.
@freek22 Oh goodness, I'm sorry to hear that! Do get well soon!
@freek22 I'm not going to uploading anything I don't have the resources to make good example videos and don't know what info to put in the description. there is a branch of the github with an auto convert that can just process the quants to "all" auto_convert.
I just wanted to get these things for personal use and then confirm that it works for those that know this platform.
@freek22 Just want to say I'm so sorry you've been in the hospital and wish you a speedy recovery and an uneventful return to recuperating at home.
How do I get more control over the movements? I either get the most gentle smile, the joker about to destroy me or jeff the killer level of smiles (or other movements).
I admit, I have very little experience yet on video gen but descriptions seem to be not effective.
I come back after some absense and this model has only 220 upvotes????
Literally the best video model out there and only 220???????????????????????????????????????????????????????????????????
Hey, can you provide an IVT workflow? Because I keep getting errors when running this model on both high-noise and low-noise data (since the workflow I have requires two noise models), and I can't find the workflow in the video below the page either.
there is an workflow in the training folder its also possible that comfy or some nodes has changed over time and other settings are required now.
i get zero motion for any character. hardly any movement at all. any other model works 100x better.
if so use the model you prefer .. or train your own one the model was good for its time it did get released . i also can recommend the furry enhancer video lora that i have released later that lora has some improvements that was not implemented when this model (wip) model was released
its rather the opposite, after trying 12 different variants, only 2 produce the results im looking for, and this one is one of them. could you post your workflow so we can ensure it's not an error on your end?
@freek22 I actually do find this checkpoint has the best fur hair texture I have tried, but the motion is just slow to start. I will try tweaking. very good quality hair, especially with the FFGO lora.
@daspin335 I might have just needed a lora for my specific action. I will try something more simple. It worked on something more basic.
@freek22 i tried the furry enhancer, and although it gives amazing motion results with this checkpoint, the background quickly dissolves into a red painted mess around the characters. Any tips on how to avoid that? Happens every time.
@kronos1959777 the problem seems is comfy had change over time some parameters this issue pop over sometimes when they change some stuff do you have used the newest workflow in the training folder ?
@freek22 i use wan2gp.
The fur and hair detail is amazing. combined with FFGO lora, this works great but not when using the furry enhancer lora at the same time :( otherwise great quality. Using i2v Painter helps increase motion. So I would recommend this.
It is weird, when I use your furry enhancer lora with this checkpoint, the motion is amazing but the background just distorts immediately and turns to a weird paint smear. without the lora the motion is just ok, not quite as fast.
Can someone give me a hint, please? EVery WAN2.2 workflow asks for High and Low checkpoints, which one is this? Help please... thank you.
were you able to figure it out? in the same predicament
@LipsThorogufareEvening This has both sides baked into one. Just plug it into both the high and low noise samplers along with whatever loras you want at ~6 steps total.
You can make a workflow that only uses just one sampler but then you'd have issues with loras that use high and low models.
@LipsThorogufareEvening Slammed into both fields with mild changes in the weight, it has odd results tho, however I found a way to requiere only one at a time, isn't that great but works
@Samethingai Thank you, I figured that out later, but it gives odd results somehow, it seems needs more mods in CFG and size apparently.
its a merged model so you only do one ksampler setup, there is no high/low with these, see
https://cdn-uploads.huggingface.co/production/uploads/631be8402ea8535ea48abbc6/tW8lXhRrAXzluvjNPudag.png
hey im new to comfyui, just got the desktop app and i have a 9070xt with 16gb vram and 32gb system memory. Can i use this? i need some help setting it up or finding a workflow that works with this model, everything ive tried has given errors but im sure that im doing something wrong
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.