Update: WAN 2.2 Smooth Workflow v5.0 - 04/19/2026
"WAN 2.2 Smooth Workflow v5.0" update: The new version now has a First2LastFrame Workflow. Make sure to fully update your Comfyui and the extensions.
At the time of making the Showcase: Stability Matrix v2.15.7 and ComfyUI v.019.3.
Lightx2v Loras used:
HIGH with weight 1.0 -> high_noise_lora_rank64_lightx2v_4step_1022.safetensors
LOW with weight 1.5 - > lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors
Still uses the same extensions as v4.0 - No need to install more. Make sure ComfyUI, its dependencies and the Extensions are fully updated.
The Showcase is focused on the usage of First2LastFrame since the other workflows are pretty much the same. Added some examples of Loops you can make. I might take a few tries but its possible to make something simple - keep Loops at 5 seconds at most - Wan 2.2 doesn't work very well above that threshold.
Updated Notes: Added mode hyperlinks for some of the Resources and Models Extensions used in the Workflow.
The videos on Showcase used the SmoothMix I2V v.2.
All Extensions for WAN 2.2 Smooth Workflow v5.0:
Links for MMAudio usage:
MMAudio NSFW Model (fine-tuned off the base model)
Put the MMAudio models on ComfyUI/models/mmaudio. Create the folder if it does not exist.
Update: WAN 2.2 Smooth Workflow v4.0 - 03/14/2026
"WAN 2.2 Smooth Workflow v4.0" update: A new version with a new Ksampler method and more functions. Make sure to fully update your Comfyui and the extensions.
NAG: Now uses Normalized Attention Guidance (NAG) in Ksamplers. Check the github page more info. It makes the negative prompt be effective even at CFG 1.0 for improved quality and control.
Adaptive Prompts: You can use Adaptive Prompts when writing your Positive prompt now. Check the github page more info. To check the choices made by the Node just open the Positive Prompt Subgraph.
Updated Notes: Added hyperlinks for all Extensions used in the Workflow. Other notes have been added and updated. There are alot of little notes everywhere. Make sure to read them. ^^
The videos on Showcase used the SmoothMix I2V v.2 and SmoothMix T2v V.3.
All Extensions for WAN 2.2 Smooth Workflow v4.0:
Update: WAN 2.2 Smooth Workflow v3.0 - 19/01/26
"WAN 2.2 Smooth Workflow v3.0" update: Its a big workflow now. Make sure to fully update your Comfyui and the extensions.
All in one: Txt2vid, Img2vid and Audio Options! You can now try adding audio to your videos.
Select your workflow: I tried to make it very simple to change between the Img2vid, Txt2vid and Audio Workflows.
New fps Output: I had to change the video outputs to 25fps to make sure the audio and video are in sync.
Multiple Notes: There are alot of little notes everywhere. Make sure to read them. ^^
Added Extensions:
It uses MMAudio. So install - ComfyUI-MMAudio Nightly
And its requirements:
MMAudio NSFW Model (fine-tuned off the base model)
Puts the MMAudio models on ComfyUI/models/mmaudio. Create the folder if it does not exist.
Small Update - 10/16
Small change made to "WAN 2.2 S. Workflow v2.0" and "WAN 2.2 Txt2Video Workflow v2.0" That should allow for a faster VAE Decoding.
Update: WAN 2.2 Txt2Video Workflow v2.0 - 10/08
"WAN 2.2 Txt2Video Workflow v2.0" update: now its SUPER simple and easy to use.
I made it to use with the Smooth Mix Txt2Vid Wan 2.2 Checkpoint! Make sure to try it! ^^
Here are my recommend video setting when using the workflow with SmoothMix Txt2Vid:
Low/Mid end PCs:
Set Steps to "6" on both KSamplers;
KSampler 1 -> end_at_step = "3";
KSampler 2 -> start_at_end = "3".
High end PCs:
Set Steps to "10" on both KSamplers;
KSampler 1 -> end_at_step = "5";
KSampler 2 -> start_at_end = "5".
Update: WAN 2.2 Smooth Workflow v2.0 - 09/29
"WAN 2.2 S. Workflow v2.0" update: now its SUPER simple and easy to use.
I made it to use with the Smooth Mix Wan 2.2 Checkpoint! Make sure to try it! ^^
All the videos on the showcase for WAN 2.2 S. Workflow v2.0 used Smooth Mix Wan 2.2 Checkpoint.
It does not has Florence anymore. Make sure to install and update all necessary extensions listed below. No new extensions where used.
Update: Adv. Workflow - 08/29
"Adv. WAN 2.2 S. Workflow" update: added the Node "Color Match" before 'Video Result' and 'Last Frame' to help with color consistency.
Update: Txt2Vid Workflow
Added a txt2vid workflow. Thanks for vinque for the suggestion. =)
Update: Adv. Workflow + First2Last Frame
Made a few adjustments to the basic workflow for these new version!
Both use the same tools/weights as the basic workflow so no need to install anything new if you can already use the basic one.
Both versions have "Power Loader Node" set for the High and Low Loras.
More Loras usually means more instructions put on the Positive Prompt so Florence2 is disabled by default.
To make a looping animation on First2Last Workflow make sure to set enough time for the AI to process the loop. Try 5 or 6 seconds.
WAN 2.2 SMOOTH WORKFLOW
PLEASE make sure to read everything.
Making a simple Workflow from Wan is tricky. Wan requires are more complicated set up to work. I'm going into this assuming you have read all tutorials necessary to make Wan 2.2 work on your PC as doing that here is unresonable.
WAB 2.2 is still pretty new and there is still alot to test. All images used on the Showcase use the same set of loras, weights and configurations. The only difference was the image used and the Positive Prompt. I tried using realistic, semirealistic and anime style to show that its possible to animated any style you want by using the configuration on this workflow.
So, before we begin, this Worflow uses the following:
Use the one that fits your own specs. The one used for the showcase was the "wan2.1-i2v-14b-720p-Q8_0" version.
Out of all Loras that increases generation speed this one has given me more consistenty results.
The one used for the showcase was the "Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32" version.
Of course you should test other versions or Loras that speed up generation as well as well.
Showout to @CubeyAI for the Loras. Without these making NSFW results was a pain, and we can't have that can we? Make sure to use the latest version of those Loras.
Used extensions:
==========================================================
Its divided in sections to make it easy to navigate.
PROMPTS
Here is where the magic happens. So lets start here.
There are a few ways to use it:
TYPE NOTHING ON THE POSITIVE
By default Florence will create a prompt for you and a lot of times... thats enough lol. You don't need to type anything on the Positive. For examples of results that use an empty positive prompt check this and this.
I recommend you start by leaving the Positive Prompt empty at first. After you see the results you should begin working on your prompt if you did not like the results.
USE SIMPLE PROMPTS
Don't put a giant wall of text on the Positive Prompt as Florence will do that already. The showcase has Videos with very few prompts, some with just ONE.
DISABLE FLORENCE
If you really can't seem to get what you want disabling Florence might be the way to go.
FLORENCE
Choose your Florence model. Default model is "Florence-2-base-PromptGenv.2.0". Feel free to test other version or disable all of Florence nodes if you think its not helping.
CHECKPOINTS GGUF / SAFETENSOR
Very straight foward. Just choose your Checkpoint: GGUF or Safetensors. Disable the ones you are not going to use. Set to GGUF by default.
SPEEDING TOOLS AND LORAS
Here is where Sage and the Loras are set up. Feel free to change values or disable what you don't what to use.
CLIP / VAE
Still using the same as the wan 2.1
Text encoder: umt5_xxl_fp8_e4m3fn_scaled
VAE: Wan 2.1
IMAGE
Load your Image! :)
VIDEO SIZE
By default its set to 480x720 with Upscale by 2. That results in a 960x1440 video! Default upscale method is "lanczos"!
Feel free to change anything it to fit your own specs.
VIDEO LENGTH
Here you can adjust the video length.
By default is set to 81 or 5 seconds.
VIDEO SETTINGS
All settings here were set after testing. Feel free to change to your liking! :)
VIDEO PREVIEW/RESULT/LAST FRAME
You should be able to see how your video is turning out on the preview. If you don't where the results are going just cancel the run!
The video result is set to be at 30 fps. The video will be saved on Video/(Date_video_was_made).
The last frame of the video will be saved on the same folder as the video. You can use this to extent your videos! Check this post for examples.
====================================
Other recommend extensions:
Description
FAQ
Comments (36)
Can I ask what the minimum computer specifications are?
AI is mostly GPU/Video card powered, so any card with 8-16 gigs of vram can generate these in about 5 minutes. More powerful cards can generate faster. I have an RTX 4090 and it takes about 4 minutes to generate a really high quality video for me. I can get it to generate much faster, like in 1-2 minutes, but the quality tanks at those speeds.
fox23vang226 I only 3080 12g ... 6min ？
zjunchoon969 you'll probably need to download a Q4(quantized 4) version of wan.
weird problem, i started by doing a text to image, it translated to an anime character holding a sword in a sea of flowers. then i switched nodes and every single generation from i2v is using the "last image" of the anime in the flowers picture. Despite me having deleted completely all related folders, temp, output (including relevant video folder within that folder), and input folders. And it still keeps bringing up that last image as the reference and converts every single thing i make into an anime.
This is with all Florence 2 nodes disabled.
Seeing same issue.
Wtf!? I wish I could offer you guys some direction but I have used this dozens of times and nothing like this have ever happened. o.O
I guess try disabling all the stuff thats not necessary one by one. Since you already disable Florence2 try the same with the Loras, Last Frame or even the Negative Prompt
DigitalPastel The only thing i can think of is that there's another folder somewhere that it's sourcing the "last image" that's not the ones i listed. I even had a clean backup of comfy, with everything set exactly how it was a day ago (i like doing this so i never ruin anything). And upon completely restoring the backup, it still brought up that damn last frame of the sea of yellow flowers and the person from the video holding a sword. (Which logically says that it has to be getting that exact image from a folder outside of comfy ui portable itself)
Unfortunately i can't scan though a millions folders and files in folders outside of comfy ui throughout my entire system to find that last frame so i can finally be rid of it.
whitespider9999 Thats so weird. This is starting to sound like some creepy pasta stuff! lol
If anyone can figure out how to summon this anime girl please let me know so we can figure it out how to exorcise her.
DigitalPastel I definitely see the humor in it :D LOL . Using your workflow would be good too :D
Ah, interesting! I've had a similar issue which I mention in my comment. I haven't had the problem of it pulling from a previous image (I think). But what it would do was cross-dissolve part way through the gen I'd requested, changing to a long shot with my subject (and what she was wearing), same semi-realistic style... holding a sword in a field of flowers. 0_0
This was after me bypassing everything in the Florence group (as the node didn't work for me). However, if I leave the "concatenate" node, it seems to generate as normal. I think without that node, it breaks the positive prompt completely (as I'm sure it has "value" in there now where it didn't before, and final prompt preview has "text" which was likewise empty.)
Edit: Did another test - and for me, it's definitely the concatenate node. It needs to be active, even if all the other Florence nodes are bypassed.
syntec Oh so its the Concatenate node messing things up? Great find! =) I will update the workflow without that node (and positive encode node just to be safe) out of that zone and reupload the workflow. Thanks!
What are weird issue. I mean... an anime girl holding a sword in a field of flowers? Where the hell did that come from?
DigitalPastel IKR - the only thing I can think of is that wan2.2 somehow has a default that it outputs when it gets confused by a weird prompt format or something. That's my rationalization to try not to think about machine ghosts anyway :D
DigitalPastel I found this while looking through the json, a prompt embedded into the positive encode, likely culprit.
"title": "Positive encode",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.3.26",
"Node name for S&R": "CLIPTextEncode",
"widget_ue_connectable": {
"text": true
}
},
"widgets_values": [
"A girl in a white dress stands in a field of yellow flowers, holding a sword, under a sunlit forest. Slow and small Movements. Idle Animation"
],
"color": "#222",
"bgcolor": "#000",
"shape": 1
"
fox23vang226 Removing that quoted prompt text from the json fixed it for me. The updated workflow file did not fix it btw. That line of text needs to be removed.
whitespider9999 Thank you for the help! I removed it and updated the workflow! ^^
For those who want this to be easy, follow this guys tutorial for an easy sage attention installation... https://www.youtube.com/watch?v=CgLL5aoEX-s&t=177s
It works with this workflow.
The only thing this workflow reeeeally needs is rgthree power lora loader. I just removed the two single lora loaders and replaced them with it and it works fine.
The simplicity is beautiful and can easily be expanded.
would love to see your workflow for this
Yes. This was made with an easy expansion in mind. My first draft of this workflow included the Power Lora Loader Node. Since there isn't many Loras for 2.2 out yet and a lot of stuff for it is still being updated (specially the Loras related to speeding up generations) I chose to not include it and keep it simpler.
ResistAi so far I just swapped the lora loaders. Next I'm going after the loopers and will see what the best way to integrate them is.
I have another workflow I like allot but it doesn't use sage attention. The fact this uses sage attention and creates an end frame is what's really awesome. Too many of the cool workflows lack them and they are not as nicely layer out as this one.
I'm very confused, all the tutorials on the internet are for the portable version.
Brilliant, thank you!
when i run this comfy its end with a txt file with the prompt generated by florence. and it dosent generate the video.
just turn off Florence, you dont need it at all imo. check out my comfyui workflow in the video i posted below on how i bypassed florence.
Nice workflow, love the labels and explanations.
I had some nasty error with Florence though (Unrecognized configuration class) which I've not been able to find a solution to. Pity, I was interested to try it out with wan2.2. I've bypassed it for now, but learned not to bypass the concatenate node as when I did my gens would dissolve mid-way, putting my subject in a field of flowers with a sword for some reason...
Is it the 'Cache only has 0 layers, attempted to access layer with index 0'? I had that one. had to follow this: Florence2Run Index Issue · Issue #175 · kijai/ComfyUI-Florence2
I'm using stability Matrix so I to deleted the folder "ComfyUI-Florence2" in "ComfyUI/custom_nodes/" folder. And reinstalled the node with the Stability Matrix extension tool.
DigitalPastel My one was "DownloadAndLoadFlorence2Model Unrecognized configuration class <class 'transformers_modules.Florence-2-base-PromptGen-v2.0.configuration_florence2.Florence2LanguageConfig'> for this kind of AutoModel: AutoModelForCausalLM." with a looot of other stuff after it, model types. I found a reddit thread (https://www.reddit.com/r/comfyui/comments/1jh2rca/miaoshouai_unrecognized_configuration_class_any/) from months ago with a similar issue that said it was a transformers version problem. The issue on the github that the thread links to is still open. There's a few suggestions on the github so will have another look a bit later, I can live without Florence for now :D
I kept getting swords in all my generations, then I noticed this in the json file
" "title": "Positive encode",
"properties": {
"cnr_id": "comfy-core",
"ver": "0.3.26",
"Node name for S&R": "CLIPTextEncode",
"widget_ue_connectable": {
"text": true
}
},
"widgets_values": [
"A girl in a white dress stands in a field of yellow flowers, holding a sword, under a sunlit forest. Slow and small Movements. Idle Animation"
],
"color": "#222",
"bgcolor": "#000",
"shape": 1
"
Some random prompt embedded in the positive encode.
"A girl in a white dress stands in a field of yellow flowers, holding a sword, under a sunlit forest. Slow and small Movements. Idle Animation"
So I followed someones advice here to disable everything from florence and only enable the concatenate node and the swords stopped showing up. Im not a fan of florence so i disabled it, dunno if it takes up extra resources but i prefer to prompt myself.
I dont know what im doing wrong, but my videos always come it a bit blurry, not ultra sharp like a lot of the examples here...is it because you guys use a really high initial resolution? Im always conservative, using a low 480p resolution. Would the blurryness go away with a 720p resolution?.
Its mostly always the face and eyes that are blurry.
I hit the same “mystery swords” issue. The culprit was a hard-coded positive prompt inside the CLIPTextEncode (“Positive encode”) node in the workflow JSON (widgets_values): “…holding a sword…”. I gave the JSON to ChatGPT; it cleaned the file and removed that stray text - after that, no more swords. You don’t need to disable Florence; just clear that embedded string (or recreate the node).
1- unlink green line from positive encode in prompt group to concatenate node in florence group.
2- erase the random prompt "A girl in a white dress stands in a field of yellow flowers, holding a sword, under a sunlit forest. Slow and small Movements. Idle Animation" in positive encode.
3- link again green line, positive encode with concatenate.
4- to use without florence, bypass the group florence only with concatenate node enabled (autor note says "disable all")
Is anyone else getting a snowing effect when they change the CFG and/or Steps?
likely too high of a cfg? i keep my cfg between 1 to 1.5, its because im using a lightning lora/speed up lora
Can someone tell me why is my laptop (Scar 18, rtx 4080) crashing when I want to run this workflow?
overheating maybe, laptop GPU's are notoriously bad
The laptop itself crashes or just comfyui? I'm having no issues. Use a hardware monitor to keep an eye on your memory usage and temperatures. You may be overheating. Make sure you use the correct quant of wan2.2 for your VRAM.
Great workflow many thanks !
For all your work which i love ;)
It took me a while to get it working as it was my first video generation i had a hard time compiling sageAttention finally resolved here https://github.com/thu-ml/SageAttention/issues/228#issuecomment-3172388431
One question tho I have issues with faces not very good if you have any insight ?
i used a lot stable diffusion but not much of comfyUI consider me as a noob ;)
EDIT: Better results on face when i pay a little more attention on being coherent with dimensions ;) also i m full comfyUI now not going back to stable diffusion probably XD