This a general basic Wan 2.2 workflow for GGUF for Image 2 Video with notes in the workflow telling you where files go and links to get those files. I would say it is "new user to WAN friendly", I tried to make it as easy as possible for the used resources. But I'm no expert for sure.
It includes links to some Lora files as well for some NSFW as well.
Most of the Information is in note pads in the workflow.
I use a 4070 12gb and 32gb of ram for this workflow specifically. But i believe you can use the 3000 series as well. You may have to use a different GGUF though.
I recommend taking a look at the workflow before installing custom nodes.
I did strip it down the essentials only for the exception of power lora loader. So there is very few custom nodes.
Don't ask me on too much on trouble shooting because I dont think i can help too much on that lol. I can try, but no promises. I am still learning and have generally been learning from other peoples workflows.
You will also need to have a updated Comfyui to use the Model.
Additional Lora Recommendations:
SmoothMix Animations XXX
https://civarchive.com/models/2040641/smoothmix-animations?modelVersionId=2376136
WAN DR34ML4Y - All-In-One NSFW
https://civarchive.com/models/1811313/wan-dr34ml4y-all-in-one-nsfw?modelVersionId=2176505
Make sure to Like those models and the ones from the workflow! Without those creators, this wouldn't be possible without them!
The frame interpolation model
Description
FAQ
Comments (15)
oved your project, my friend, but I had a bit of difficulty achieving the result I wanted.
In short, I tried to remove clothes while keeping the face identical, but the video didn’t turn out well, and the face always changed a little.
I replaced your current LoRA with the Super Power Mode Loader, which I think is better for using multiple different LoRAs.
I’m a beginner, so it’s quite difficult to do certain things. I tried using a face-ID to keep the face from the photo, but it’s complicated.
Would you be able to rework this in your workflow in the future? I believe it would be a great upgrade, especially for those who know how to use it.
From my current experience as well, some lora work better on certain art styles. Such as anime vs realistic. Sometimes using only a high with no low can fix out the details before generation.
(Even bumping the strength to 2 or even 3.)
I have had some lora completely change my image and i found some new ones that just perform better on the art style at times.
I usually will do short test (2 seconds) to save time to make sure the adjustments made for my workflow are working as intended or there is no major changes to the desired output.
I dont have much experience using faceid so i'm afraid i cant help much on that. But there good resources on youtube that can help something like the faceid.
I am to also understand that comfyui has some safeguards now to make it harder to use faces, in a nefarious sort of deal using certain nodes.
Thx for the effort, but i cant get it to work. only blurry results.
were do it put the high and low loras, there is the purple box high/low "Power lora loader (rgthree)" and below the purple high/low "lora loader model only" were 2 h/l loras can be added. i gues the loewe with strengh 0.2 is for wan2.2_i2v_lightx2v_4steps_lora?
can u plz explain a bit more? thank you!
All the lora files themselves will still go to the lora directory and you will be loading 2 loras. Usually on the model page for them has a tab for a high and a low. If there is no low. Then you will only load it into the high side and load no pairing lownoise.
The group on the left in the box for High noise is for the Loras, most will be set at a strength of 0.8 to 1.0, then the box for Low Noise will be the secondary Loras for Wan2.2 usually the same strength of 0.8 to 1.0 as well.
The powerlora loader is what you can use to make it easier to switch and toggle loras easier. Reconnect the high noise on the left to the model workflow, which will be the loras on the bottom in high and for the same for the low.
Connect the top model to the left side of the high noise box, then take the output of powerlora model back down to the blue tab reroute on the left of that then select and delete loras under the tab reroute. Repeat for the low as well.
If you are not using the powerlora. The loras in boxes are set in the workflow. If you use more loras you will have to attach them in the column for each side. The purple is in bypass node mode, select and hit CTRL+B to remove the bypass to enable them.
If you get garbled messed as your output you may have changed the Ksampler, they need to be set to what i have in the image/workflow. The only thing you should really be messing with in the ksampler is the steps which will give a crisper result, but will take longer, ( i dont go past 12 really) and you will make those changes to both ksamplers. Changing the start steps and end steps in the ksamplers can cause the garbled mess. They are confusing to me on the results they make if changed too much. I haven't messed with it much though.
Hello, thx for your workflow. I'm here for some help. I've downloaded every model and node you list, and run it several times. But the output video was totally blurred. I changed the cfg value of KSampler for different results, but it didn't work for the blurry. I'm the beginner, how can i make it work? TvT
You should only change the steps and not CFG. I would not recommend going higher then 12 steps due to time it takes to run. I would also not recommend really changing anything in the ksampler other than the steps.
Use the modelsamplingsd3 above the ksampler instead.
@WintersDawn thx for your recommend, it worked! The blur is resovled by pull up the steps, but motion of the outcome chara is very odd. For example, what i give the prompt is "high definition, not blurry, not pixelated, natural motion, masterpiece, best quality, very aesthetic, the girl in the picture take off her clothes and show off her naked body, flat breasts with pink nipples, pussy have no pubs." , but the outcome video just the girl swaying in situ, nothing else happened. Can you help me? TvT
@HeavenSword It looks like it is the way you are prompting. The Wan model uses a more natural language when prompting.
Example for yours:
"The woman in the video takes her shirt off and is completely nude revealing her naked form, her breast and tits are visable. She is smiling at the viewer as she takes her shirt off. She takes her pants off after she reveals her breast to show her pussy. She changes her expression to a more content and sensual expression as she removes her pants while looking at the viewer. The scene is vivid, full of emotional expression, with smooth animation and excessive jiggle physics. The scene is high quality. The camera is steady."
You could also specify underwear color after removal and then a action after as well. Or the beast details after the removal prompt part. (You may also want to remove the jiggle part for your breast sizes)
Not all seed numbers will give the desired results, but it's more about prompting. I tested this prompt and it worked for the results you are looking for i think.
I had the nsfw lora on at 1.0 in both high and low noise. And the lightning speedup lora. (The NSFW could possibly need to be lowered in strength)
Recommend to also include the lora SmoothMix Animations XXX. (Linked in description) this lora can cause a bunch of movement at times, sometimes it needs to have a 0.3 strength in both high and low.
In addition. You want a specific scenery, you would build the scene first then the person. Like, "the scenery is an old tavern..." then describe the tavern details you would want, such as lighting, materials that make up the tavern, such as is it stone or wood walls, is there people there and so on.
There are limitations to the model itself based on training data it has in it, since GGUF is a smaller stripped model, things may look weird in detail if it wasn't really trained on the data or if the model has been merged to include the data, like a lora that is trained.
@WintersDawn thx, i'll try to regroup prompt later! May i ask what's the strength value of loras recommended? I stuffed at least 4 loras after the loader xD. Because of the test always failing, never succeeding, i'm not sure which step is right, or wrong.
@HeavenSword most Lora's, your high noise one and the low noise one will have the same setting strength, make sure they are in the correct loaded sides to corresponding sides, wan 2.2 trained Lora's will usually have 2 loras that need to be downloaded and put in the lora location, if there is only a high noise, you will load it on the high side only at 1.0 (you can load wan 2.1 lora this way and it supports them usually, but not always) Most of the time you can set them to 1.0, but some recommend other strengths, such as the smoothanimation xxx recommends 0.3.
generally they need to be between 0.8 and up to 1.0.
Changing the modelsamplingsd3 to high values, such as 12 to even 20 (or higher even) will cause more movement but can cause bad results without extra steps in the ksampler.
And sometimes loading to many Loras will damage the final results too. It has various degrees, sometimes dropping strengths on them as you load more can fix those issues, but may also cause results to not take place such as a action will be minor or even not happen properly.
@WintersDawn thx a lot!
Thank you! This works really well. Just needed a quick fix for one missing connection and now I'm getting great results, much cleaner than other workflows I've tried.
Good to hear that it works well for you.
I actually think I am going to put up a new workflow soon up in a week or so (maybe sooner?)
It has 3 prompt fields that are stitched together - making a 18 second clip with better transitions and flow. (Also takes more time obviously as well) But I have been testing it and it seems to be holding up well for about anything you put in it. It is a matter of the use is to flipping a switch and setting a number for that switch and clicking run.
The longest part will be loading all the loras into it for a person to use it. (Which i will try to link to downloads for in it)
The integer part of the various node isn't working for me. Any ideas?
It could be due to the version I made the workflow in (older now) try rightclick on the integer node, about 5 down should be a option "reload node". It may unlink though and will need reconnected.
Alternative is to delete the integer node itself and manually set the value, i just had it there to make changes a little easier.
Another option is to double click a empty space and type in "int" and select integer and manually recreate the node that way.