⚡ FusionX Lightning — WAN2.1 ComfyUI Text-to-video and image-to-Video Workflow's
Note: Do not use the FusionX model with these workflows. Use the base WAN 2.1 model instead. FusionX has LoRAs that are already built-in, and some aren't compatible with LightX—mixing them will result in very poor outputs.
Supercharged and optimized for WAN2.1, this ComfyUI workflow lets you create stunning videos fast — just 4 steps thanks to the new LightX LoRA, dropping generation time to as low as 70 seconds at 1024x576!
🧩 Comes in 3 Versions:
Native
Native GGUF
Wrapper
VACE and Phantom coming soon..
Each version includes handpicked, fine-tuned LoRA stacks (Ingredients) for top-tier results — fully exposed and easily swappable.
☕ Like what I do? Support me here: Buy Me A Coffee 💜
Every coffee helps fuel more free LoRAs & workflows!
🧪 Text-to-Video LoRA Ingredients:
MoviiGen – cinematic motion
MPS Rewards – motion/detail tuned
Lightx2v – low-step, high-speed engine
High Speed Dynamic – boosts motion dynamics (used instead of AccVid for LightX compatibility)
+ Custom LoRAs:
Realism Boost
Detail Enhancer
🎞️ Image-to-Video LoRA Ingredients:
MPS Rewards
AccVid – temporal alignment + speed
Lightx2v
+ Custom LoRAs:
Realism Boost
Detail Enhancer
🚀 Major Upgrade: Image-to-Video Prompt Adherence
The Image-to-Video workflow now delivers way better prompt adherence — characters follow actions more accurately, with noticeably more motion and dynamic scenes than before.
Thanks to the updated LoRA stack (plus Lightx2v + AccVid synergy), you get:
Smoother motion flow
Better scene transitions
Characters that actually move the way you prompt them
It’s a big step up from older versions — especially for motion-heavy prompts. Give it a try and see the difference.
## 📢 Join The Community!
A friendly space to chat, share creations, and get support.
👉 Click here to join the Discord!
⚖️ FusionX vs FusionX Lightning?
FusionX brings the most realism.
FusionX Lightning is built for speed and low VRAM setups. With the right prompts and tuning, you’ll get results that rival the original — but way faster.
Note:
👉 If you're seeing strange particle effects in Text-to-Video, check your prompt for words like dust, particles, smoke, etc. — removing them usually clears it up.
👉 Experiencing ghosting or echoing artifacts? Try setting your shift value to 3.00 — this typically resolves the issue.
Description
FAQ
Comments (124)
When will you be available?When do you have time to arrange a VACE
Im working on it, stay tuned
I'm looking forward to it very much!
Thanks, this is awesome. This NAG node is another game-changer. Finally it follow negative and we don't have what we don't want. Also color-match another great node which I really needed.
What are the Pros/Cons of Wrapper/Native? Cheers!
The biggest one is the ability to use block swap. This lets you render at resolutions and frame counts beyond what your GPU vram could normally fit with the native workflow.
Lot of extra features. Multi-talk, enhance a video, uni3d, the list goes on. If u look at the sampler u will see a bunch of inputs for other things.
Awesome! You are a superhero in our community!👍😽😊
That is very kind!!
Nice and perfect as always, tnks! NAG yeeee!!!! ;-)
Your welcome!
Great WFs with all necessary links and explanations! Thank you very much!
Now everything follows the prompt perfectly, the scenes are dynamic, I assume it's also due to the NAG node, even with 121 frames the scene remained dynamic, thank you for your work! 👍👍👍
The workflow is very good and interesting but can i add more loras to it ? or can i replace one of the initial ones to something else without breaking the workflow ?
you can add more, just duplicate a lora and plug it into the mix.
where can i find the motionboost lora needed in the native gguf workflow ?
the link to the lora is in the workflow
good job! thank you!
Using FusionX_T2V Lightning workflow takes 4 times longer on the Wanvideo Decode node than using FusionX_T2V workflow. The time saved on sampling speed is gained back on this node. Is this normal? Does this happen to you?
Although its four steps the generations take around the same time for me as with 8 steps on the regular wan fusion workflow Im sticking with the wan fusion ingredients workflow ive been having a lot of success with that one
Is it normal for the appearance of characters to change in FusionX_Lightning i2v Workflows, especially when using FusionX's Checkpoint? It seems to alter the character's look
You're not supposed to use the FusionX checkpoint in these workflows—pretty sure I left notes in the WF about this too. FusionX includes most of the LoRAs, but some aren't compatible with LightX, so you can't mix them. Please stick with the base WAN model—the link is already in the workflow.
I guess you were using a T2V model wrongly.
This workflow is super fast, around 150 seconds to generate a 81 frames video with 12G VRAM on 4070Ti. Thanks a lot for the sharing!
Hi buddy, how are you able to generate videos so fast with the 4070ti? What's the rest of your setup? I'm new to this.
Are you using GGUF, Native, or the Wrapper?
Have a great day :)
What is the WANvideoNAG node? I'm lacking it.
For those who have the same Error, you need to update your KJnode.
@Era1701 not work for me((
@atsanrickman2016407 did you update KJ Nodes?
Bro, this workflow is absolutely crazy. It actually outperforms your FusionX. Have you considered merging this workflow into FusionX 2.0? Absolutely epic.🔥🔥🔥
Thank's!!!! And having the loras open is better so people can experiment. So I don't think ill be making a merge. If I did a merge then I would have to make one for everything and gguf's and just don't have the time.
@vrgamedevgirl yeah this is the best way imo. I like having multiple levers to play with haha
My final output is very blurry for some reason, what could cause that do you think?
I would need more details, what CFG, shift, scheduler, res, and prompt are you using? Depending on prompt, sometimes a lower shift can cause ghosting/blurry output. I would try increasing shift to 2 or 3. Higher values can cause loss of detail though so beware.
@vrgamedevgirl Hey thanks so much for replying. It looks like changing the schedular (simple) and sampler (uni_pc) fixed the blurriness. is there a recommended schedular and sampler with this workflow?
@kilplix107 It really depends on prompt. The defaults worked well for me, but what I was doing was a bit different I think. If you check the sample images, you can see they are a different type of video so I think it really just depends on what your doing.
@vrgamedevgirl gotcha! thanks again for taking the time to reply.
@kilplix107 Your very welcome!!
awesome job! I would like to ask where can I download the lora weights? And can you share the lora weights of moviegen and accvideo before merging?
The workflow as all the lora's with the weights.
crazy work! But I want to ask a question, it seems that I don't see MoviiGen's lora in the workflow of comfyui, so how is MoviiGen integrated? In addition, the AI face in the generated video is quite serious, which lora modules have the greatest impact?
Unfortunately, according to my tests, all LoRA models except Lightx2v and AccVid would severely distort the facial features of anime-style characters. It tends to add lipstick to every character, making them extremely ugly.😢
since the lora's are open and exposed, you can just bypass them or set them to a lower strength and find the settings that work best for you.
You can keep faces stable using this Lora: https://civitai.com/models/1755105/wanfusionxfacenaturalizer
It was made for FusionX, but I've found that it works in general even without FusionX.
My gens are taking about five minutes on the base settings. I did have to bypass triton/sage attention because I don't have it installed, so not sure if that's why, or something else? I'm on a 4090, I've gone through and updated all the custom nodes, played around with settings, loras, but no change. Not super familiar with comfyui, so i know I'm missing something.
Its hard to say. have you tried using the wrapper version? Sometimes that helps because you can enable block swapping. But sagattn speeds it up alot so that could just be the issue.
@vrgamedevgirl Installed sage and triton seemed like a bit of a daunting task so I avoided it, but I decided to go through with it after your response and it's definitely improved the times, thanks! 👍
The GGUF is great for our cheap video card.
but I still got an Error message "requires pytorch 2.7.0 nightly" on the GGUF workflow "Model Patch Torch Setting (enable_fp16_accumulation-ture)". but I check my pytorch is 2.7.1+CU128
Don't know what I'm missing? disable that option,still can run, I using GTX 3060 12G OK, run a 8 second video about 15 minutes.
Did u try to just change your base precision to fp16?
@vrgamedevgirl do you mean in the " Patch Sage Attention KJ - Sageattn_qk_int8_pv_fp16_triton" . I tried to select 4 types, but no luck, do I need config something or download some files into comfyUI for running the "enable_fp16_accumulation"
@DaShu999 sounds like your not using wrapper. I would try the wrapper. Or you have to bypass the patch sage nodes.
@vrgamedevgirl Thanks for your reply, I just tried your Native_GGUF (i2v and t2v). Cause video card only 12G, still no time try your wrapper workflow. (maybe I can run Wan2_1-I2V-14B-480P_fp8_e4m3fn, later) thanks,
is it possible to use sageattention with teacache in this workflow?
You can't use Teacashebecause we are already only doing 4-8 steps and teacache skips steps so you will get a very bad output. Teacache was used before these new models came out that allow for reduced steps. So like when we had to use 20-30 steps.
@vrgamedevgirl Thank you for the detailed explanation, really appreciate your effort. I would also like to know if this workflow is capable of generating videos longer than 5 seconds, or is it specifically optimized for 5-second video generation?
@qazxsw You can try to do more frames but what happens is, for every 81 frames the prompt kind of "starts over" its a wan thing. Also, you will more than likely OOM if you try for more than 121. But if your using context options you can do as many as you want, but again, the prompt starts over so you can get some strange fades every 81 frames or so.. if that makes sense?
@vrgamedevgirl @vrgamedevgirl Yes, it make sense. So I guess for now, using the last frame with i2v is the only proper way to create longer videos. Thank you for sharing this knowledge. It really helps a lot and makes things much clearer for me.
@qazxsw There is a new workflow someone else made you may be interested in. Its called endless travel and its pretty cool. If you join the Discord and ping me I can point you in the right direction.
There's no MoviiGen lora in i2v workflow??
in the description there was, am i missing something??
Sorry that was a typo on my part. The i2v does not use MoviiGen. MoviiGen is meant for t2v and adding it to the i2v has negative side effects as it does does not play with with LightX
if disable triton/sagatt i get this error
Prompt outputs failed validation: WanVideoNAG: - Required input is missing: model
Maybe when you diisabled triton sageattention you broke the chain ? you can bypass the nodes by clicking on ctrl+b when selected
If your having Triton and Sageattention errors;
https://civitai.com/articles/12851/easy-installation-triton-and-sageattention
Yeah it's working quite good, the speed gain isn't as much as I hoped, but it's still a speed gain and no apparent downgrade in quality
Wan2.1-Fun-14B-InP-MPS.safetensors brings a lot of "lora key not loaded" errors in T2V WF. Is it normal?
Yes you can ignore the error. The lora still works just fine.
Thank you
Holy shit! 45sec to generate a 93 frame with 10 step video (350x580)! Although the results are not very good, but it's blazing fast!! Anyone know how to add custom lora into this workflow?
I dont get it. Its taking around 30 minutes?
I can gen 480 I2V in about 5 minutes.
Need more info... gpu? Ram? Settings? Please join discord and i can help u there
@vrgamedevgirl Joined the discord. Asked in support. x
tried my best running it on default settings with all the links provided and all the vids come out blurry, idk what i'm doing wrong :( it gens fast but the quality is next to none
Please contact me on Discord so I can help u troubleshoot vrgamedevgirl is my username
First of all: Fantastic job on this workflow. I've recommended it to several people. I'm making 5 sec 768x512 videos in 5 minutes (with added interpolation and upscaling to double that) using GGUF_Q5_K_M on a 3080Ti 12GB. I would like to ask a few questions, if you don't mind:
1. I've only tried the "Native" workflow, but I had to bypass "Patch Sage Attention KJ" and "Model Patch Torch Settings" nodes because it would get hung up for several minutes. Do I need to install something else?
2. I can generate a maximum of 9 seconds before running out of memory, but even then results are weird. Someone else said 5s is the limit for Wan before "starting over" and my results seem to support that. Do you agree that for longer videos we need to extend from the last frame rather than create more in the beginning?
3. Is the workflow compatible with other Wan2.1 LoRAs? Can I just string in another LoraLoaderModelOnly node in there somewhere (after the DetailEnhancerV1 perhaps)?
Thank you again for an incredible workflow!
Hey, first off—really glad to hear you're enjoying the workflow and recommending it around. That’s awesome!
To your questions:
If the "Patch Sage Attention KJ" and "Model Patch Torch Settings" nodes are hanging, make sure you have Triton and sagattn installed—they’re required for those patches to work properly.
Yep, 5 seconds is pretty much the limit for Wan before it starts "resetting." So to go longer, you’d need to extend from the last frame, but fair warning—quality degrades over time with that method. There is an endless travel workflow that helps with this though—if you hit me up on Discord, I’ll point you to it!
And yes, the workflow’s fully compatible with other Wan 2.1 LoRAs. Just drop in another LoraLoaderModelOnly node and connect it right in.
Let me know if you run into anything else!
vrgamedevgirl Thanks for your reply and insight. Will Triton and sagattn speed things up or improve quality even more? I added interpolation and upscale with a final VideoCombine, and then bypassed the two VideoCombine, since I didn't really need those outputs anymore. I'm down to about 4 minutes per 5 second video. Just curious if you think Triton and sagattn would improve things beyond that.
aichataccount yes. Sagatt can give you a 30% speed boost but i have heard it actaully can cause artifacts, so does NOT provide any quality improvements. Just speed.
Why my videos are in slow movement? there's even a "Slow" in the negative prompt, but it doesn't seem to help. Beside that, Thank you for the detailed explanations on the workflow. This is neede more.
Are you doing text to vide or image to video? Can you contact me on discord on the server? its in the description. Its easier for me to help through there. thanks!
I'm a newbie, so take what I say with a grain of salt, but I had the same slow videos. I ended up adding an interpolation node (doubles frame count), then upscaled 2x, then did VideoCombine at 31fps. That solved it for me.
I have the opposite problem. The action in my videos is too fast. I thought Shift was suposed to control that, but turning it down to 1, and then up to 15 changed nothing. I don't know id the shift parameter has to be connected to something. I just don't know enough about this stuff.
thanks for your excellent job, did you notice that Kijai have updated a new series of Lightx2v? i am wandering if it would work if i replace it with the new version of Lightx2v.Have a nice day.
it seems like the lightx2v i2v version is much more motion and dynamic scences than the t2v version 2 when using the i2v WF
abc123qarqf1y yes, you can swap it out! You can also add the new pusa lora as well!
vrgamedevgirl What is pusa good for? is it on the same level as FusionX lora?
vrgamedevgirl 128rank is better than 64rank?
Zeb101 Pusa is not like fusionX at all. You would actually want to try and add it to a fusionX workflow and see if you see any better results. I have not tested it much yet.
a1161327317 not much difference besides file size really
vrgamedevgirl what is pusa？and where is pusa？
a1161327317 Its a new model and someone created a lora so you can plug it right in. I don't really know what is is but it helps with motion and prompt adherence I think.
its here
https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Pusa
Great workflow!!!! Thank you very much!!!!
I can't even get a vid to generate. After fixing links to models (you have non-standard paths in some places, like "new folder" for the main model loader). It starts, sits at loading text encoder for several minutes. Then stops. No error, nothing showing in the queue except that it finished after only a few minutes. If I try to start it again, it says "Running in another tab", but there aren't any other tabs.
In the terminal are a bunch of load failures, which never showed on screen. Gotta troubleshoot those now...
The non standard paths are just paths to my models. You need to change them out to your own models after you download them.
I can help you if you join the discord channel. I have heard of anyone else having this issue. Once you join the discord server at the link in the description, you can ping me @ vrgamedevgirl in the support channel.
vrgamedevgirl - Yes I figured that out. But honestly for general distribution, you should use more standard paths.
I have been able to generate vids, but the quality seems very low, as is prompt adherence. Changing the CFG to a higher value in the Sampler made it unhappy.
Also, I sent you a friend request on Discord. Should be from Rat of Steel, in case you weren't sure. Thanks!
Carrera001 Hey, sorry about the path issue! I tend to keep my models organized in separate folders since I work with so many, and most of my workflows require me to point to my own models—so I didn’t realize that could cause problems for others. I did see your request but didn’t realize it was you. Best way to reach me is just to ping me on the FusionX server in the support channel — I usually ignore friend requests as of late since I mostly keep those for close friends and family.
WORKS Amazingly well. I got t2v working but i2v has zero movement. using gguf workflow. anybody got any insight on this?
In the Wan_FusionX_i2V_Wrapper_Lightning_Ingredients workflow. This error occurs. What could be the reason for this?
CompilationError: at 1:0: def triton_poi_fused__to_copy_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr): ^ ValueError("type fp8e4nv not supported in this architecture. The supported fp8 dtypes are ('fp8e4b15', 'fp8e5')") Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"
Two workflows are well done, super gain and super quality. Thank you so much for that.
Native
Native GGUF
But Wrapper refuses to work. I look at it and the nodes are different and the structure is completely different from the other two WF.
As I understand it, the main problem is TORCH, but I don't understand what's wrong.
smolusha the main part of the error is this
type fp8e4nv not supported in this architecture. The supported fp8 dtypes are ('fp8e4b15', 'fp8e5'
i would have to see the exact settings your using.
A HUGE thank you for making this. This has been the first (and so far, only) i2v/t2v workflow I've used and it works almost flawlessly. Apart from needing to re-select the LoRAs on first load it really couldn't be much simpler to use.
On my 5080 (16GB) card I was able to generate 5 second (81/16fps) videos with 4 steps at 960x720 in 10-11 minutes with the 14b models. I noticed zero difference in VRAM and system RAM usage on the 1.3b models, they both completely max-out my GPU and 64GB of RAM.
After some tweaking and managing to install Triton (on Windows) I can now generate with the same settings as above in under 2.5 minutes. I don't know if this is fast or slow for my setup, but I was amazed at the performance gain!
That being said, I've noticed that when using Triton on i2v generations the color-correction step takes ~30 seconds to save the video where it would only take a couple of seconds without... It also seems to un-load the model(?) from VRAM after each gen, which adds a small amount of time when you have multiple gens queued-up in a row.
Do you (or does anyone else) know if these are known quirks with Triton, or could it possibly be a quirk with my setup/environment?
Thanks so much for the kind words! Really glad to hear the workflow’s been smooth for you overall — and those Triton speed gains are awesome
Just a heads-up: I’ve got some custom nodes out now for Color Match and Film Grain and sharpen nodes, that are super fast. You can find them on the ComfyUI Manager under vrgamedevgirl — feel free to give them a try, should be much faster than what is in the WF right now. If you join my discord I do post updated workflows in there.
As for the Triton quirks — yeah, the color-correction slowdown and model unloading are known behaviors. Triton does some memory shuffling under the hood which can cause those kinds of pauses. Nothing wrong with your setup!
Let me know if you find any other oddities or have suggestions — always happy to improve things
wow, I do 4 step 440x800 129fps just about 8 minutes + 2X enlarge(5 minutes) on my cheap 3060 12G. Do I really need a 5080 upgrade? or anyone know nvidia NVIDIA DGX Spark 128G, It really run the full video model like wan?
Very very nice work. Thank you so much works beautiful. t2v and i2v genereate in 2 minutes with 4070 ti super. I only dont understand one thing my pictures in i2v changing a lot, they nothing like the original. Waht setups i can tweak so it stays more like the input picture. For all the people who didnt get the Attention to work this one helpt me alot.
https://www.reddit.com/r/StableDiffusion/comments/1jdfs6e/automatic_installation_of_pytorch_28_nightly/?share_id=6B4t_c0afrnklQcp7Tbgy&utm_content=1&utm_medium=ios_app&utm_name=ioscss&utm_source=share&utm_term=1
Hey! Can you show me an example of the i2v issue? Do you have discord? i'm vrgamedevgirl
please reach out and I can help you. I have not had any issues with that.
why is the workflow file an image not json?
The workflow is embedded in the PNG image. Just drag it into comfyUI or open it just like you would a JSON.... You can save it as a JSON if you really want it that way.
drag it onto comfyui
Any chance for a Wan 2.2 version of this workflow?
Really love the workflow! Other workflows always gave me issues, but this one is like working so perfect. I just don't understand how people make smooth ai videos (everyone makes amazing ai videos in the Gallery here, specially the smooth ones), while mine is nowhere near smooth, even raising frame_rate to 24, it just made the video speed up (googled it and i thought that was the solution, i was wrong), even adding slow motion to the prompt didn't work. I would appreciate advices if it's possible. I'm still new to t2v and i2v, but it's really fun to do for own entertainment.
(forgot to add, i do use Triton and Sageattention)
I think some people run video's through topaz or the like. All my video's are raw and straight from the WF though. I don't upscale or interpolate any sample vids.
also glad its working great for you!!
I'm trying this now. I thought it was more than twice as slow as using FusionX Q8 GGUF on my RTX 4070 Super 12GB VRAM machine, however it may be that it requires less than half the number of steps? Not sure yet. Everything is the same as on your GGUF workflow except that my base model is wan2.1-i2v-14b-720p-Q8_0.gguf, I use umt5_xxl_fp8_e4m3fn_scaled.safetensors with FusionX model but the GGUF version here and I don't have Triton/SageAttn but I don't use those with FusionX Q8 GGUF either. The FusionX Q8 GGUF is awesome by the way, thanks, I'm only trying this workflow as you say it might be adapted to work with Wan 2.2, otherwise still very happy with my Wan 2.1 based FusionX Q8 GGUF :) A couple of observations, when I used a lot of steps (20) I had white speckles all over the final video, but not when I only used 10 steps. Also, I disabled AccVid to see how much difference that would make and it made no difference in speed whatsoever, so AccVid was doing nothing? Most of the speed increase comes from lightx2v I think (I tried the new lightx2vWan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v.safetensors in place of the T2V one but it made no discernable difference).
How do I change it to use it in Mac Studio? It doesn't work. What's your opinion?
I am not familiar with Mac Studio.
Hello, thank you very much for your great work. I need your advıise about your workflow. Is it possible to add a node to your workflows in order to apply prompts based on time stamps or frames. I.e 0-5 sec. "woman starts walking towards camera" 5-10 sec. "she smiles to camera". I want to control the scenes If possible. So far I could not make it. I am not expert on comfyui. I am just a copy / paste person following experts like you :) So please let me know there is no node for this or there is a simple way to do it with adding some commands to your workflows. Many thanks in advance.
anything with wan 2.2 with high noise low noise loras?
I regretted that I downloaded this process. There are simpler and more understandable processes in which you can use the same lores.
aha.. you should name and link them ;-)
I keep getting the "KSamplerNo module named 'sageattention'" Could you help? Thanks!
Very well documented! Thanks for the download links and the instructions!