🌀 Wan2.1_14B_FusionX — Merged models for Faster, Richer Motion & Detail in as little as 8 steps!
📢 7/1/2025 Update!
New: FusionX Lightning Workflows
Looking for faster video generations with WAN2.1? Check out the new FusionX_Lightning_Workflows — optimized with LightX LoRA to render videos in as little as 70 seconds (4 steps, 1024x576)!
🧩 Available in:
• Native • Native GGUF • Wrapper
(VACE & Phantom coming soon)
🎞️ Image-to-Video just got a major upgrade!!!!!!
Better prompt adherence, more motion, and smoother dynamics.
⚖️ FusionX vs Lightning?
Original = max realism.
Lightning = speed + low VRAM, with similar quality using smart prompts.
☕ Like what I do? Support me here: Buy Me A Coffee 💜
Every coffee helps fuel more free LoRAs & workflows!
📢 Did you know you can now use FusionX as a LoRA instead of a full base model?
Perfect if you want more control while sticking with your own WAN2.1 + SkyReels setup.
🔗 Grab the FusionX LoRAs HERE
🔗 Or Check out the Lightning Workflows HERE for a huge speed boost.
📌 Important Details- Please read the full description below because small changes to settings will provide totally different results in a bad way! I have been testing and already found better settings so just please read below! Thank you :)
💡Workflows can be found HERE (This is a wip and more will be added soon.)
🛠️Updates section has been moved to the end of the description.
A high-performance text-to-video model built on top of the base WAN 2.1 14B T2V model — carefully merged with multiple research-grade models to enhance motion quality, scene consistency, and visual detail, comparable to some of the many close source models.
## 📢 Join The Community!
A friendly space to chat, share creations, and get support.
👉 Click here to join the Discord!
Come say hi in #welcome, check out the rules, and show off your creations! 🎨🧠
💡 What’s Inside this base model:
🧠 CausVid – Causal motion modeling for better scene flow and dramatic speed boot
🎞️ AccVideo – Improves temporal alignment and realism along with speed boot
🎨 MoviiGen1.1 – Brings cinematic smoothness and lighting
🧬 MPS Reward LoRA – Tuned for motion dynamics and detail
✨ Custom LoRAs (by me) – Focused on texture, clarity, and fine details. (These both were set to very low strengths and have a very small impact)
🔥 Highlights:
📝 Accepts standard prompt + negative prompt setup
🌀 Tuned for high temporal coherence and expressive, cinematic scenes
🔁 Drop-in replacement for WAN 2.1 T2V — just better
🚀 Renders up to 50% faster than the base model (especially with SageAttn enabled)
🧩 Fully compatible with VACE
🧠 Optimized for use in ComfyUI, With both the Kijai Wan Wrapper and native nodes.
📌 Important Details for text to video:
🔧 CGF must be set to 1 — anything higher will not provide acceptable results.
🔧 Shift - Results can vary based on Resolution. 1024x576 should start at 1 and if using 1080x720 start at 2. Note: For more realism lower shift values is what you need. If your looking for a more stylized look then test higher shift values between 3-9
Scheduler: Most of my examples used Uni_pc but you can get different results using others. Is really all about experimenting. I noticed depending on the prompt that the flowmatch_causvid works well too and helps with small details.
📌 Important Details for image to video:
🔧 CGF must be set to 1 — anything higher will not provide acceptable results.
🔧 Shift - For image to video I found that 2 is best but you can experiment.
Scheduler: Most of my examples used dmp++_sde/beta and seems to work best but you can experiment.
After testing, to get more motion and reduce the slow-mo look, set your frame count to 121 and frames per second to 24. This can provide up to a 50% motion speed boost.
📌Other Important Details:
⚡ Video generation works with as few as 6 steps, but 8–10 steps yield the best quality. Lower steps are great for fast drafts with huge speed gains.
🧩 Best results using the Kaji Wan Wrapper custom node:
https://github.com/kijai/ComfyUI-WanVideoWrapper🧪 Also tested with the native WAN workflow, generation time is a bit longer but results match wrapper.
❗ Do not re-add CausVid, AccVideo, or MPS LoRAs — they’re already baked into the model and may cause unwanted results.
🎨 You can use other LoRAs for additional styling — feel free to experiment.
📽️ All demo videos were generated at 1024x576, 81 frames, using only this model — no upscaling, interpolation, or extra LoRAs.
🖥️ Rendered on an RTX 5090 — each video takes around 138 seconds with the listed settings.
🧠 If you run out of VRAM, enable block swapping — start at 5 blocks and adjust as needed.
🚀 SageAttn was enabled, providing up to a 30% speed boost. (Wrapper only)
Workflows for each model can be found here: HERE
🚫 Do not use teacache — it’s unnecessary due to the low step count.
🔍 “Enhance a video” and “SLG” features were not tested — feel free to explore on your own. -- Edit. I did test "Enhance a video" and you can get more vibrant results with this turned on. Settings between 2-4. Experiment! SLG has not been tested much.
💬 Have questions? You’re welcome to leave a message or join the community:
👉 Click here to join the Discord!
📝 Want better prompts? All my example video prompts were created using this custom GPT:
🎬 WAN Cinematic Video Prompt Generator
Try asking it to add extra visual and cinematic details — it makes a noticeable difference.
⚠️ Disclaimer:
Videos generated using this model are intended for personal, educational, or experimental use only, unless you’ve completed your own legal due diligence.
This model is a merge of multiple research-grade sources, and is not guaranteed to be free of copyrighted or proprietary data.
You are solely responsible for any content you generate and how it is used.
If you choose to use outputs commercially, you assume all legal liability for copyright infringement, misuse, or violation of third-party rights.
When in doubt, consult a qualified legal advisor before monetizing or distributing any generated content.
### 🧠 More GGUF Variants
- 🖼️ [FusionX Image-to-Video (GGUF)]
- 🎥 [FusionX Text-to-Video (GGUF)]
- 🎞️ [FusionX T2V VACE GGUF (for native)]
- 👻 [FusionX Phantom GGUF (for native)]
###🧠 fp16 Versions can be found here:
-- 🖼️fp16 FusionX Models
📌gguf comparisons!
I'm slowly adding to this list, but here you can see how the models compare against the main model.
Text to video:
--------
🛠️Update 6/8/2025 - Image to video model is published! Settings that I use in the example videos: Steps = 10 / cfg = 1 / shift = 2 / schedular = dmp++_sde i'll post a WF soon.
🛠️Update 6/7/2025 - Published a i2v phantom model that can take up to 4 reference images and combine them into a video. Posting workflow soon
Phantom WF is getting uploaded soon.
🛠️Update 6/6/2025 - Added a new gguf model! If you want the highest quality and have enough VRAM get the V1.0 model otherwise gguf is the next best thing! When using the gguf's it will take longer to generate even on an RTX 5090.
Description
FAQ
Comments (356)
Showing latest 264 of 356.
FusionX_Text2Video- Is this what version 1 was? Or a new model?
Its just V1 text to video. i had to Rename it because it was causing confusion :)
@vrgamedevgirl and the FusionX_PhantomV1 model, is it t2v or i2v?
@vrgamedevgirl and the FusionX_PhantomV1 model, is it t2v or i2v?
@_RUST_ Phantom is basically a image to video model yes but its a bit different since its multi image.
Hi, I'm new to this, but why does the resulting clip have nothing to do with the originally uploaded photo?with video card NVIDIA GeForce RTX 4060 8GB....if it matters!!
I'm trying this model - Wan2.1-14B-T2V-FusionX-Q4_K_M.gguf!A 5 second clip is generated in about 6-7 minutes
use image2video, not text2video
that model is for text to video. Not image to video.
@vrgamedevgirl , And which model could I use? Thank you.
@tany6666372 which workflow are you using?
@vrgamedevgirl , Wan_FusionX_i2v_Native
@tany6666372 the i2v model. The link to it is in the notes in the wf
Is it just me or does the PhantomV1 not implement faces at all for generations? I love your T2V Fusion model, but your Phantom doesn't seem to want to clone/copy any face that I add to the image inputs like the vanilla phantom does.
I realized it was my base precision setting. For some reason, it needs to be bf16 for your Phantom model, but works with fp16 on the vanilla. Amazing work!
@blastermaster123 Oh strange! Good find! I don't know what I have my WF even set to for that. I will have to check.
does Riflex work with this?
Have been playing with the t2v version for a few days and it's amazing!
Any chance of having a diffusion forcing version? I mean like Skyreels V2 DF. I don't know if it's even possible, but it would be nice for generating longer videos.
I wouldn't even know where to begin lol. This is just a merge not really a new model. U can use context options in the wrapper though to get longer videos. Join the discord channel and I'll show u😀 link in description
Extending videos is possible with this model when using VACE
You just provide WanVideo VACE Encode node with few frames from video you want to extend keeping other frames grey. Also you provide masks (black for frames from video, white for grey frames) and get extened video with great quality
The native workflow works fine, but the WanWrapper gave me blurry, low-quality results. The base precision was set to fp16_fast, but since my PyTorch version is outdated, I changed it to fp16. Could that change have affected the results?
That should not have effects on the results. I use wrapper. Join discord listed in description and we can chat
Suberb model, I am using the Q8 version!
Model is nice and quick but has some issues when using it with a pretty basic workflow, it struggles with text and probably tries too hard to create motion?
CivitAI post/review upload logic is broken as hell (saying the IMAGE im uploading is too big..) so find my 3 test generations here:
https://drive.google.com/file/d/1o-LaeZ2V2mB7NDdGbXQ95A9Vp9hzoghf/view
https://drive.google.com/file/d/1jqie2nlEThNVICI7K5l8ciOYteLBZqTG/view
https://drive.google.com/file/d/1I0_40ftLK-irWzcZ9HWKBAwdTpIg6o7j/view
(Have reached out to civitai support to hopefully have them fix the post/review upload logic)
In the case of the above video's not containing the used workflow etc (i honestly don't know) the prompt used for these videos (so you can more clearly see the issues) is this:
A scene from a movie made around the year 1979, in the style of 'starsky and hutch' or 'dukes of hazard', the scene starts with us looking up at a billboard advertisement sign in detroit, the billboard has the text "BUS STOP PLEASE", behind the billboard is a blue sky. From the left of the frame a classic american car (a white chevy) is flying into the billboard with great speed, on impact the car deforms and bounches off, both the car and the billboard sign crumble and shatter and release lots of particles and bits, the scene is highly cinematic and action packed.
- Text needs work
- It seems to have ignored my attempts to make it look like a scene from some old movie
- Like mentioned before it seems to try a bit too hard to create motion, in the last 2 clips its like the collision creates a black hole 🤭
This was made with the Phantom v1 version using comfyUI and a basic workflow i repurposed to test this model with sage attention like was suggested in one of these posts or the description or something.
(This is based on a scene from "Bus stop please" by Fatboy Slim & Daniel Steinberg, a quite lovely AI-movie-fever-dream hehe)
The text2video model is noticeably better and doesn't seem overly eager to force a lot of motion, text still needs work but is also better! 👍
(the below 3 are same prompt and workflow, just using the older text2video model, these turned out larger than 100mb so google drive doesnt wanna autoplay them.. 🤦♂️ but if you download and open in chrome u can see them in full)
https://drive.google.com/file/d/1IDK0SKgek7gMaEciR2Ta6i7kKk-h5-8n/view
https://drive.google.com/file/d/1EtMge4ujEtJtytx2yTC3JRmfq3QicKbg/view
https://drive.google.com/file/d/14miNlztPhxLKAhiKyBYzDIdxfQSAM8Xm/view
- Has better understanding of the front and the rear of a car being different things (but still pretty random)
- Better text (but still not perfect)
- Less motion, in a good way
Also, just incase this is misunderstood, i very much appreciate and applaud that you are trying to make custom wan models, keep up the great work! Hope this silly stuff helps you a bit with making future models better! 👍
@baconmessenger can you come over to discord and chat? I see your issue and its the prompt. Click here to join the Discord!
Wan is picky when it comes to prompting. I trained a GPT on Wan documentation so It works really well.
Also, the reason your having issues uploading is because your using WEBP. WEBP is huge and technically an image file. I know its video but CivitAI sees it as a huge image. You need to use my workflow I shared which uses the video combine mp4.
I just made your video and it looks really good. FYI please reach out and i'll share the prompt.
@vrgamedevgirl Sorry i don't use discord but good to hear the issue is just with my prompt then & ill start using your workflow instead! 👍🤭
@baconmessenger do you have any skype? Or email? Here is a prompt that made a really good video of what you wanted
A low-angle shot opens on a vintage billboard reading "BUS STOP PLEASE," set against a vivid blue Detroit sky. The camera slowly tilts down, revealing a wide intersection below. Suddenly, from the left, a white 1970s Chevy blasts into frame, soaring airborne. As it crashes into the billboard, the structure buckles and collapses, scattering shards and debris in slow motion. Sunlight casts hard shadows, highlighting every ripple in the car’s crumpled frame. Dust and particles fill the air in a gritty, cinematic explosion.
@vrgamedevgirl Thanks ill try that prompt :)
i2v Any tips for reducing contrast? Particularly noticeable with skin tones. Over the course of the first 3 or 4 frames the skin morphs from the reference to a much more intense contrast. Tried different samplers, but didn't see much difference.
Are you using the WF I shared or a different one? I would need to know your exact settings and example of whave your issue. Can you chat with me over here? Click here to join the Discord!
there are couple of nodes called Match-Image-color or something similar. they read the color palette of the initial image and applying it to the rest of the frames. helps a lot
'cfg' should be 1.0. If you have a larger value you will get that over-burned effect.
@ejsd i have notice some differential on the first frames as well, no matter that my CFG is 1. i am using the comfy nodes though, haven't tried the KJ-Wraper
@vrgamedevgirl @menegosm Testing: It was straight your WF. Tried Color Match node and it was much worse. Had read that loras can be the culprit, but I wasn't using any, so: ran with regular base 14B i2v, same workflow, but 20 steps. No color shift. Added CausvidV2 and dropped back to 10 steps. No color shift. Added AccVideo. No color shift. Couldn't figure out how to add MoviiGen, tried replacing base with it but got a lot of lora errors. Added MPS Reward Lora. EUREKA! got the color shift!!!! For testing, concentrate on a tanned face in bright sunlight.
@menegosm the prompt can change the color of the overall output to i think. Like If i mention "golden hour" int he prompt, the video will have a different overall tone to if that make sense.
@jpXerxes Ok so MPS is the culprit. That is good to know!!!
@vrgamedevgirl all of my tests were with the same seed. It's not that the color/contrast are bad, just different from the reference. Starts fine, then saturates over the first 3 or so frames
Are there more GGUF versions of I2V? Q3 is a bit low, I could use Q5_K_M. On your HF page I can only see GGUFs for T2V
Yes sorry about that I updated the description. they are here HERE
@vrgamedevgirl I think it got deleted since it's 404 now, but I guess this is the reup?
QuantStack/Wan2.1_I2V_14B_FusionX-GGUF at main
The team behind the gguf's did create new links. I'll update the links soon!
this has been updated!!! The end of the description has the new links to all the GGUF models now.
@vrgamedevgirl I tried I2V Q6, but due to its built-in LoRAs, it tends to add realistic details to anything anime-related.
@EechiZero Hey! Feel free to reach out to me on Discord and we can chat more. Just to clarify—the two Detail LORAs I added were set to very low strengths, so they shouldn't be affecting that. I did extensive testing to make sure that wasn’t the issue. There are different settings to move away from realistic look. Shift should be set higher even up to 9
Does Phantom FusionX have any ggufs by any chance?
yes, if you look at the end of the description you will find a link
Which one? t2v? i2v? can we chat on discord?
Vid 2 Vid?
I don't have a workflow for that yet but can add it to the list. I would visit the discord server. Link in description
Hey, i think you switched the parts for recommended scheduler on this part:
"Important Details for text to video:".
I can be mistaken though.
What workflow? I have done more testing and if its in the WF I shared then it works. I know for the GGUF workflow a diff one looks best but its all about testing to find your own personal one that you like.
the I2V is really incredible. is there any chance for the 480 version though?
That's actually a good idea.And I will add that to my list.You should join my discord server.The link is in the description of the model.I have a request section
several people report that it nukes the likeness of the reference image
Hey! Feel free to reach out to me on Discord and we can chat more. Just to clarify—the two Detail LORAs I added were set to very low strengths, so they shouldn't be affecting the face. I did extensive testing to make sure this would not happen.
It is really great work but it seems like not working well with reference face. Still trying some setup right now
Can you reach out to me on discord?The discord server link is in the description of the model
Just to clarify—the two Detail LORAs I added were set to very low strengths, so they shouldn't be affecting the face. I did extensive testing to make sure that wasn’t an issue.
fusionxPhantomv1 is unparalleled, but I encountered some problems when using the fusionxPhantomv1 model. I added prompts to control the movement of the camera, but it seems to have no effect. Please share some keywords for camera movement so that I can test it. Thank you very much!
I can say that Phantom is hard to control and adding some motion lora's might help. Have you tried my GPT prompt helper? If you provide it with the images your using in phantom and tell it what your looking for along with the camera movements it normally does a decent job. https://chatgpt.com/g/g-67c3a6d6d19c81919b3247d2bfd01d0b-wan-cinematic-video-prompt-generator
@vrgamedevgirl Thanks, I'll test it again
I haven't looked for three days, so I'm entirely out of the loop. Which one is / what does phantom do?
I2V doesn't seem to work very well with LORA. I'm using your workflow, just added the LORA node, I also tried with other workflows and I got same results. Also, as someone said, the color and saturation are increased compared to the input image.
I had no issues using Lora's. What issues are you having? Keep in mind this model does have AccVid, CausVid, and other's baked into it that can have a slight effect on the output. I would need to see all your stettings to be able to assist. Please join discord server. Link is in the description.
I was getting much better results using Lora with the native workflow than the wrapper workflow.
You've earned a place in heaven! ❤️ I can't believe the native 15GB I2V model runs smoothly on my RTX 3080 10GB—even with its limited VRAM. It generates videos in just 8 minutes with really good quality. Thanks for the model!
Your very welcome! would love to see some of your stuff! You should join the discord if you have it :) Its a great little community over there already :)
You model-merge is awesome! I absolutely love the results. I'll soon post some things for sure!
Thank you for your work :)
Your very welcome!!
Hey Is the i2v GGUF model 720p or 480p?
This is the fastest workflow that can gives me clean and good results only with 1 cfg, low steps and without using 2 samplers. *chef kiss..
However, my issue with this gguf i2v model is that the subject's facial likeness changes frequently, it struggles to maintain a consistent face across different seeds. Around 30–40% of the time, the face ends up looking different compared to source image.
Original i2v wan2.1 gguf + causevid and accvid + 2 samplers gives me more consistent face likeness imo.
It's the model itself apparently, the LoRas that are baked in affect the face in i2v. Apparently the author knows about this, but nothing to be done about it right now. https://www.reddit.com/r/comfyui/comments/1l96g4t/fusionx_version_of_wan21_vace_14b/
That appears to be misinformation. Although many people are reporting that I2V is wrecking faces, this post was made in the thread by the Author so idk:
"Creator of the merge here. The two enhancer loras were set to a very low strength and do not change faces. I tested this before merging. Many days of testing. What model and workflow are u using? This model needs specific settings to get best results. Please join my discord and i can help u https://discord.gg/NtvxDhvV
@Kiefstorm Hey there. I have been testing the i2v workflow with photos of myself and husband and so far the faces are left intact. The step count and scheduler play a big role in this though along with the prompt.
@Kiefstorm Hi, it seems the problem is the built-in MPS lora with large weights more than 0.3 - 0.5, I tested separately this lora with other diffusion models and with weights = 1 i got the same distortions of facial features. Btw, thanks a lot for you great work!
This model is very impressive. It is currently the best WAN2.1 model. Different from the original WAN2.1 model, FusionX can accurately guess elements that do not exist in the picture, and the movements of the characters are more reasonable. In the original WAN2.1, you just press Run and pray that it doesn't go crazy.🔥
can you please share your gguf workflow fastest
@kumarkishank959811 https://civitai.com/models/1309369/img-to-video-simple-workflow-wan21-or-gguf-or-lora-or-upscale-or-teacache
Here you are.
@Era1701 thank you so much i am trying
You have really accomplished something great here, given Wan2.1 a serious upgrade. Sent you some buzz in thanks. Is it possible to create a gguf version of the phantom model for those with lower specs?
Thank you so much!!! I am working on a Phantom Workflow with gguf
@vrgamedevgirl You are a goddess amongst mere mortals.
@sarashinai Your too kind :) I am also releasing a FusionX LoRA that has all the benefits of FusionX but can be used with the base wan model and skyreels and it has a speed boost and can be adjusted! Posting that ASAP hopefully by tomorrow once I get all the new Workflows done and examples video's. If you visit Discord you can see some examples. Discord link in description!
@vrgamedevgirl I'm trying to get a sense of the age captioning, FusionX seems different than base wan2.1. I've tried "mature", "old", "elderly", "octogenarian" with inconsistent results. Do you happen to know how to prompt it?
@sarashinai Thanks for the question! The other LoRas like Accvid and CausVid that were fused into this may effect the age prompting but I would have to do some digging. In general, Wan2.1 can be a bit inconsistent with general age terms like “mature” or “elderly.” To get more reliable results, try using more specific and descriptive prompts—like “a wrinkled old man with gray hair and a cane” or “an 80-year-old woman sitting in a rocking chair.” Including visual details and context tends to help the model interpret age more accurately. Let me know if you'd like a few sample prompts to test!
@vrgamedevgirl That's good advice, I'm still really not putting enough details into my prompts, my failing. If/when you have time, just a few samples to set the range, "early twenties" -> "middle age" -> "holy crap, how are you not dead yet?" would be helpful!
@sarashinai are you on discord? if so you should join the fusionX community. I'm active there as much as possible and can help you way better in there. the link is in the description.
Would love to see an Anime specialized finetune of FusionX!
You can just the AniWan model :) - It would just need a good LoRa to do this which I think there are alot out there :)
@vrgamedevgirl Thanks! Yeah I tried AniWan before, seems some loras trained on native Wan2.1 has weaker effect......But yeah I should explore more on AniWan!
@LovelaceA there is someone in the discord that has been posting some amazing Anime video's using FusionX and a LoRa. They are crazy good
The i2v is impressive, works well with loras and adheres nicely. I can also squeeze out larger (and longer) videos out of my poor 3060.
Great job!
Thank you!
I'm glad its working for you! :) Happy generating!
Hello, first of all, your model is really great, both in terms of speed and image quality. Then I found some problems that I don't know how to solve. When I2V is performed, the generated video is not consistent with the reference image. What should I do? I have tested it in both KJ's workflow and comfyui's workflow, and the results are the same. I posted the comparison pictures in the return area, or you should be able to find them by checking my posting records. The latest one is here.
Hey, really appreciate the kind words! 😊 Glad you're liking the model. I’ve actually been testing I2V quite a bit and haven’t run into that issue on my end. That said, definitely join the Discord community — I’m there most of the day helping out, and others are around to assist too. If you share your examples there, we can dig in and see what’s going on.
Also, I’ve got a FusionX LoRA in the works that might help with results — should be posting that by the end of the day today. Stay tuned! 🔧🚀
I'm getting a blurry mess with native t2v workflow?
is the VAE different or anything?
please join discord and post in support channel. I'll need more details and screen shot of your settings. discord link in description.
This is best model of wan2.1 which I tested. Great!. Thank you so much
I used 4060Ti 16gb, 20 minutes/video, default settings, Wrapper Workflows
I couldnot use SageAttention although I installed it succesfully
Comfyui portable, python 3.11 , cuda 12.4
can you reach out to the discord and go to support channel please? link in description.
WanVideoSampler
Given groups=1, weight of size [5120, 16, 1, 2, 2], expected input[1, 36, 21, 72, 128] to have 16 channels, but got 36 channels instead
how to fix this?
Join discord server for help, link in description.
The GGUF I2V model works very well on my RTX4070. 4 minutes to generate a 3 seconds video at 24fps, very fast!
I've used the Q3_K_M variant from HuggingFace.
This is next level. So far it has worked with all my previous workflows and loras for both T2V and I2V, and is producing better results. I just had to change the CFG and Flow Shift.
I'm so happy you got it working with your workflows!! :) Would love to see your results :)
Best Wan 2.1 model on the market so far!
Thank you!!! I just took what was out there and merged it together. That is the magic of open source :)
This is working really well for me.
That is great to hear!!
Hi! Is there a way to make it work i2v with gguf?
I just released new workflows that include i2v gguf. If you go here https://civitai.com/models/1681541?modelVersionId=1903277
you will find the i2v native. The zip has both native and gguf workflows. The only difference is there is now a lora node for the new FusionX loRa which you can either start using or you can bypass it and just use the FusionX base in the main model loader and it will work like normal. Just don't use the new loRa with the FusionX base. Use the normal wan i2v model instead. Link in workflows.
The speed, quality, and ease of use of wan2.1 has been taken to the next level with the introduction of this model. Thank you very much!
Your very welcome!
For the life of me I can not get the background to move when my subject is walking forward or backwards with I2V. I've tried dolly shot, panning out, and anything else to trick it to do so. I've also tried a few loras it just refuses to work.
This is just a wan i2v downside. I personally would try phantom and use your ref image in there.
@vrgamedevgirl seems anything over 3.5-4sec it resets to 0 weird. I just set to 22fps for 109 ish frames seems to do the job for now. Haven't messed with the new model much this one is really good. One day we will be able to do 20-30 sec clips :)
This sounds too good... 🙂↔️ Tsk... Another Sunday afternoon "lost"... 🤣
EDIT: amazing, bruh. You've done magic here. Ngl, I'm shocked with the improvement in the generation speed, and also the general quality of the videos.
Just one thing - a personal one... with the base wan, I managed with prompting to my girls keep their black sclera eyes, but so far, with your model I can't get my girls keep them. The model insists, from the very first frames, to turn them into white.
Pro tip, if you want even better character likeness in Phantom try to add multiple references of the same character from different angles and expressions. It works best if they have the same distance to the camera, like a mugshot.
This model is sooo good! I havent even got started on the new stuff. I've spent most of my weekend making clips! Old dog new tricks!
Thank u! But I can't take all the credit.I just merged some good Loras into the base wan to create this😀
What tool did you use to merge the loras??
I did not make them someone else did and provided them to me. The process is a bit complicated and i'm trying to learn.
Unless you mean the merge model? There are Loras so i'm not sure which you are asking about.
@vrgamedevgirl yes merge the model/checkpoint with the loras.
@fronyax There are nodes in comfyUI that let you merge models.
The male's "member" looks alian in nature. How do people make nsfw content with this? Does it Require loras to do so?
I never have made NSFW so I would not know. But I do think you need custom loras.
dude , get a lora
why is the basic i2v model on all models?
I have no idea what your asking. You need to pick the correct model to use in the "load Model" node. There are also links to the correct model in the workflows.
Please upload BF16 Q8 GGUF
You need to be way more specific as there are many models and already GGUF's of them all. If scroll to the end of the description page you will see the links. Also the GGUF workflows have links to the GGUF files.
Didn't notice all the GGUF links at the bottom until you mentioned it.
Great Work Thanks!
absolutely amazing models, this has changed my whole WAN experience for the better, from a 20+ minute wait for a 5 second vid in all previous Wan models I now can do the same with FusionX, and higher quality in just 4 minutes thankyou so much
Your very welcome! And credit also goes to the the loras that were used to make FusionX. I only merged them into one model to simplify things for people who don't want to stack them. :)
Why the size of the t2v model downloaded here is 13G, and huggingface.co is 28G ？
Is the hugging face fp16 by chance?
The workflow is very good. However, my GPU does not handle FP8, so I have to use the GGUF version. Unfortunately, I cannot use GGUF with this workflow because the node doesn't allow it — or maybe there's a way to use GGUF that I don't know about. I want to use gguf in Wan_FusionX_i2v_Wrapper workflow.
Gguf is only supported it native
There is a gguf workflow
It is better if wan videowrapper nodes supports gguf. Actually i want to use gguf in wrapper workflow it supports both i2v and start and end frame.
@muzicworld456509 I did not create the wrapper so im not sure what your asking me to do. U would have to reach out to the creator of the wrapper and request they make a gguf node.
@vrgamedevgirl Thanks for your support. I definitely reach creator of wrapper.
This reminds me of the lightning model from the painting era, where the same acceleration does not compromise quality, it's so perfect! Thank you
the new workflows link doesnt work
Try now
Does this model support 8gb vram? I've tried like every model including gguf and I keep getting a torch.outofmemory error. Can anyone with 8gb vram relate?
Did you try block swapping 40 blocks? 8 gb is very hard to get this working on and if you did it would take a very long time to create a video.
I can run this model at maximum 636 * 360, 81 frames, takes about 7 mins per video. I use a RTX 2060 of only 6GB of VRAM, using SWARMUI
@vrgamedevgirl sorry for late reply. I don't use the wan wrapper workflow from kijai as no matter what it seems to run out of memory. Native workflow seems to work just fine for me but I would like to try the wan wrapper. Even switching to 40 blocks, it still crashes. I used all the models with the lowest memory usage, but I can't really deduce which model is making my entire computer go ballistic and explode. I would really like some help with using which models that can balance speed and accuracy the best. Currently I use the 480p Q4 gguf model, fp8 scaled clip model, clip vision h, and wan 2.1 vae. As of now the native workflow works just fine taking about 15 minutes.
Does the FP16 version improve video quality? Is it more accurate in interpreting the prompt?
No. I have compared them both and there is really no difference. Its so small that you really need to look close to see it. The only difference is like a 1% quality boost in just detail. Not motion or anything else. Also, the LoRa was created out of the fp16 version so you could try the base fp8 model and then use the LoRa.
Which sampler and scheduler (and how many steps) can I use for good speed and quality (native nodes). I don't have dpm++ at native node - I barely manage native so I need stay at it...
The default settings in the wf i shared would be best.
Is this T2V gguf_Q3_K_S different from HuggingFace/QuantStack's Q3_K_S(link is below) ? The file size is differnet, 6G vs 7.83G !
I got it from quanstack so it should be the she one
I2v version has also different size (smaller) then other Q3_K_S at hugginface but that it is even better.
I must admit, this model is just amazing. I'm on a mid-to-low range VRAM of 8GB, but the videos are simply stunning. Great kudos!!!
To run this model, you only need your VRAM 8Gb?
@Maratek Use gguf version, at hugging face you have even smaller Q2K version 5,6GB (text encoder fp8 scaled should be change to CPU offload)
@Maratek exactly, I tried several clips, and they came out fantastic
Even the fp8 version runs on 8GB of VRAM!
You then need at least 40 GB of RAM, max. frames 81 and max resolution about 640X400. I did a lot of tests. You can compensate VRAM with RAM a bit and the GGUF versions are not much better. There is finally a limit you cannot overcome. even using the smallest GGUF version. This thing runs for sure on 8GB. The question is how high can you go with frame count and resolution.
I also figured that at low (8GB) VRAM the native nodes combined with for instance KJ's extra nodes ModelPatchTorch settings , or patch sageattention perform better than his wrapper nodes for whatever reason.
que parametros debo usar si tengo una 4060ti?
adema stengo rpblemas con el flujo de trabajo
t2v is for text to video, and i2v is for image to video, what is phantom for?
Phantom is made to load up to 4 reference (character) images to be combined (using WANVideo Phantom Embeds) as image2vid... well it's a sort of subject2video.
I have tested 2 images and it works quite well IF you get the right seed. For instance using an image of yourself and an AI character and then make a 5 second scene, (it's limited to 121 frames and trained with 24FPS). This was my favorite until VACE has been released.
@nirodha07211 any good vace workflows around? there are a ton but so far i couldn't find any anything better working than fusionx ones I already use, just started experimenting and it's a jungle out there on models and WF's and thats just for wan
This is so much faster for me. But all my old LoRas no longer seem to work as well as they did before. Any ideas?
are you using text to video? They should work. I would say to try the ingredients workflow and replace the two detail lora's with your own. But if you add a lora it should work. Let me know what type of lora and what wf your using. Ping me on discord. server link in description.
Not happy with stability and NSFW. Anatomically not good.
You would need a lora to make NSFW content. Keep in mind this model is Just WAN 2.1 with Lora's merged into it that speed it up and help with quality. It was not trained for NSFW content.
I have an issue with i2v generation for anime characters. The character's mouth keeps moving non-stop, like it's constantly talking/simulating speech. Does anyone know why this happens and how to fix it?
请问有人知道使用图生视频，二次元角色的嘴巴会一直动个不停，就像一直在说话一样，这种情况是怎么回事，要如何解决？
It's a constant problem with Wan 2.1 in general.
Try putting "talking, teeth, open mouth, lips" in the negative prompt. If you are using 1 CFG then be sure to use the new "NAG" node as well, or the negative prompt has no effect.
That cuts it down by about 75% for me.
@Zergkool It works,thank u!
WAN Cinematic Video Prompt Generator
↑ Could the parameter settings of GPT be made public? Then we would be able to use it in Google AI studio without being subject to the censorship and rate limits imposed by OpenAI.
I think this is extremely important. A good prompt can maximize the capabilities of Wan2.1. Please pay attention to this issue.
I'm sorry, I'm not seeing what the issue is? You can easily find the WAN documentation. I made that GPT for myself back when wan just came out and just shared it since it works well. There are no "Parameters" just the document that is public.
@vrgamedevgirl Thank you for your reply. I have solved this problem through some reverse work. However, my viewpoint remains valid. A good hint is very important. Maybe you could add the document link of WAN to the page?
@Era1701 I created the gpt months ago and don't even remember. Please join the discord server and maybe i can further assist you there.
So I added an i2v GGUF variant as the base model in one of my existing wan 2.1 i2v workflows. tried using the regular wan 2.1 loras, and included the fusion lora with 0.25 strength and even without it. tried 8 steps as recomended. the output seems to change the faces of the people in the provided image.
what's wrong here?
In some T2V generations, there's a ghosting effect—like an overlay layer with low opacity. Is there a way to fix it?
It may caused by teacache try bypassing it.
can someone tell what`s the difference between this and vace?Now I use vace , 5 steps completed in 2 min by 4090,and the GPU temp no more than 70 degrees
Tried just now and seems to work well, thank you! How could I add video interpolation to this for more than 16fps please?
you can use RIFE frame interpolation. Otherwise set to 121 frame length and 24fps in the video combine node.
Thanks for the great model! Fast generation speed indeed with the suggested default settings. EulerA sampler works best for me.
wow
This checkpoint is the only reason I can use WAN now. Had been struggling along with Hunyuan for months because WAN was just too slow and low quality before I started using this. So nice to have such high prompt adherence most of the time compared to the mutated nightmares that Hunyuan produces half the time as I mined for adherent gens.
I definitely have to follow up generations with your upscale workflow to get closer to Hunyuan clarity, though.
Can you join the discord and we can chat? you should be getting good clarity, might be some settings. I can help you troubleshoot if you reach out. Link to discord server in description. my user name is vrgamedevgirl
@vrgamedevgirl I actually realized that the loras I was using were causing the clarity issues. Turning down the weights brought gens to considerably higher clarity.
I have slow process problem. I am using the same i2v workflow. input image: 576x768 Output video: 576x768. It takes to complete in (8-12 mins - 10 steps)
4060TI 16GB x 32 RAM is this normal ? Should i change some settings ?
I would need to see all your settings. Was that the first run? The first time takes longer because it has to load models.
Hi there..any chance to update this workflows with controlnet? fusion x t2v and fusion x i2v are the ones that work great for me. thanks for sharing
Go to the fusionx lora workflow page and there is a controlnet workflow
Can i use this t2v model to train WAN loras? Or do I use the original WAN 2.1 14b 720p FP8 model?
You would want to use the org base wan model to train with.
Brother, you're amazing, to be honest, without your workflow, the original wan2.1 basic model's ability to follow prompt words in the image video is basically a waste, and your workflow at least makes the prompt word following ability go up a big step! The only drawback of the fusionX model is that it will automatically change the appearance of the characters to the style of European and American characters, and it cannot maintain the consistency of the original characters (I tested the animated characters and found this problem)
I just released a lora that fixes this exact issue!
Its meant to be used when using the FusionX base model or the FusionX Lora.
Just add this lora, play with the strength. 1.00 should do the trick. I like 0.7 as it still leaves some nice details. Please let me know if it works for you!
https://civitai.com/models/1755105/wanfusionxfacenaturalizer
This model is so much fun! Combined with VACE you can control using one or 2 controlnets an entire 25 seconds clip. (401 frames worked at 480p, had to rent a 48GB GPU, but it's all worth the money.)
How do you make 25 second videos? Any tips?
You'll need 'just' 48GB VRAM. No special 'skills'. The 'trick' is that you provide that many frames to the controlnet and a reference image. My workflows are within the metadata of my videos. Just download one that has 20 sec. + I have used kijay's custom nodes. This only works using VACE since the controlnet guides that much frames frame by frame. I usually use DWpose and at complex movements additionally cany. If you need a cheap online service I can recommend runninghub personal plus. Keep in mind they do not allow NSFW, though. If setting the output video for the controlnet to 16 FPS you need 400+1 frames. Then you can frame interpolate times 2 and get an 32 FPS output. I am using 835*480 resolution. If you still get an OOM error raise the block swap to 30 / 40. It mostly works at 20, ATM I am running one with 349 frames block swap 20 without OOM. Additionally you could stretch the duration by lowering the result FPS, For instance your input has 16FPS, then you interpolate your result to 32FPS, but the destination FPS are set to 25FPS. (When finally saving the result). By that the video runs slower, but longer.
@nirodha07211 Thanks. I'm using mimicpc for cloud based comfyui. The problem with vace I have is that the output doesn't look like the person in my reference image. I have a high quality ref. image.
@nirodha07211 too bad vace doesn't work with custom WAN LoRa's. My reference image is a hd picture. But the output looks nothing like my reference image.
@zoroofcalls378 You can try that new lora I created that helps with faces when using FusionX. https://civitai.com/models/1755105
this might help. Just set it to 1.00
@vrgamedevgirl but does that work with VACE?
@zoroofcalls378 of course it does. It works with all wan2.1 14B models.
Recently I have experimented a lot with FusionX and VACE and long videos. I want to share my experiences.
To retain most of the reference image's character (the background is no problem) describe unique attributes. Hair, clothing, tattoos, necklace, origin.....
If you want a movement at background, choose one that could have movements. Waterfall, waves, wind, cars...and prompt for that. Add" camera tracks the subject, cinematic movie".
Up to 20-22 seconds at 480p are possible on a 4090 (24 GB). Block swap 30 (VACE blocks: 10, 333 frames , original 16 FPS, result frame interpolated 32 FPS)). DWPose is most effective. Additionally Depth anything or Cany, (You might want to remove the background first, at DWPose it's not necessary,).
For a better match when creating the reference image on another AI, you can use img2img. Take the first frame from the video you then take for the controlnet. Adjust the denoise so the freedom of the AI is high and only the position of the subject and a similar pose retain. (doesn't have to be exact, but similar).
@vrgamedevgirl Firstly I have to thank you for your great FusionX model, It is so much fun to me to play with it! Best video generation model out there!
Question: Next time I want to try your mentioned LORA. I don't have problems with faces generally, but I still have slightly changes at the reference image, For some reason FusionX + VACE adds always a little own to the character (face) of the reference image, it is not that consistent as when I use img2video alone where the image has to go through the clipvision instead of the WanVace encode. Would it help to use it and should I try 0.7 strength?
Thank you again.
@nirodha07211 it should really help with faces. Try .7 or 1 depending on if .7 is strong enough
I've been using GGUF models with my default workflow without issues, but after updating to ComfyUI v0.3.44, the workflow fails to run. The KSampler now throws this error:
RuntimeError: mat1 and mat2 shapes cannot be multiplied (77x768 and 4096x5120)
Could anyone advise how to resolve this? Thanks.
我使用GGUF模型和默认工作流，之前都一直可以正常运行。但是刚更新了comfyUIV0.3.44版本后，工作流就无法正常运行了。k采样器报错，“RuntimeError: mat1 and mat2 shapes cannot be multiplied (77x768 and 4096x5120)”，请问这个问题要如何解决，谢谢。
Errors can happen for different reasons? If you share your workflow someone might be able to help. Or find/raise issue on comfyui github repo.
can't pinpoint the issue without a workflow but you can try to downscale your image by 0.5, then upscale your resultant images before combining them into a video.
Sounds like a resolution issue
i think u r using kijai clip in native workflow or ,scaled clip in wrapper workflow
If i recall correctly its about miss match of encoder model, do use/have tried umt5 xxl encoder
Where can I get the custom nodes for the kijai's wrapper?
phantom gguf plz ? Q4 K S its possible ?
[Found them, my bad, THANKS] .....I have some good results using GGUF Q3 T2V version as a second pass refiner, any chance of a Q4-Q5 (as long as it's less than 10GB I take) well done
Can this model generate static images?
Use this for images https://civitai.com/models/1742500/wan-and-skyreels-text-to-image-workflows
Yes. And it is GREAT! Watch this https://www.youtube.com/watch?v=oOGiYy7cTFw
Very good speed (Native version), but the result looks like a fast-forward, motions are janky. Any one has any tips??
Interesting. Can u join discord server , go to support and ping me? Link in description
I have the same issues, default non-lighting t2v workflows. Just type "girl dancing" or something and half of them seem a bit fast forward and janky.
breaking18359 If that is your full prompt than that could be your issues. you need a much more detailed prompt. I have my GPT in the description that helps create prompts.
This is the game changer. with NAG (Normalized Attention Guidance) node, can apply negative with CFG=1. it's MAGIC!
Hi, I wonder if by any chance you remember the proportion of your mix for causvid, accvid, mps, and moviegen?
If you drag the model into comfyUI it has the meta data and will bring up all the details you need.
Did you see the new acceleration lora from lightx2v?, is there any possibility for you to update your models by adding this lora instead of caus vid, I have tested it and its insane, completely solves the problem of the old acceleration loras of making "slow movements videos", this lora adds much more movement and quality to the final video with only 6 steps and lcm simple/bets. check it out
You can simply use my LightX ingredients Workflow and swap out the old LightX lora with the new one. https://civitai.com/models/1736052?modelVersionId=1964792
I won't be creating anymore "merge" models as its VERY time consuming as I would have to create GGUF's brand new workflows etc. I have a full time job so Just don't have the time. That is why I created the Ingredients workflows.
@zenpand can you provide a link? The only recent lora I found is for the 480p model.
I see mentions of ways to speed up the video animations but my issue is the opposite. A lot of my t2v renders have too fast of animation. It's not FPS, they just move too fast. Any solutions?
it is not the FPS on the output save node? it is at 16 FPS? this is the only thing that comes to mind.
I treat that as a win for me. I would add interpolation frames and stretch a speedy 3 second clip out to as much as 5-6 seconds ;) Then it's smooth with extra time lol.
Topaz Video AI is great for upscaling and framerate conversions
try "slow motion" in prompt
For some reason, the starting frame has these weird grid-like lines over it with this model for me. Running the same lora mix with the full fp16 720p i2v model doesn't produce the grid lines.
how many frames? I would move over to the lighting workflow for i2v as its much better. If you use more than 81 frames, that can create that strange first frame issue.
vrgamedevgirl pretty much any amount that works, somehow. I made my own mix for i2v and it still has some of the effect but much less pronounced. But the new wan 2.2 just seems to solve all of these issues
Will 12 GB of video memory be enough for this model or is 16 GB+ needed?
I use this model with 8GB 4060 mobile.
For ComfyUI is it suitable?
12 gb works but i would recommend the Lightning workflow.
Greetings, the new WAN2.2 model has been released. will there be an update?
I own a 4060 Ti 16 GB. While the gen speed is 15 min (which is a bit better than what I get with the generic i2v setup), I'm somehow able to gen 1024x576 using 40 block swap, which is very surprising. The results are also quite impressive visually.
Still, if you have any idea how to speed it up on such a setup, I'm listening. I haven't been able to try sage attention, perhaps that could help.
Looking forward to what you can cook for Wan 2.2.
3060ti 8G能生成动态图片吗
Thank you for your work and sharing.
Also, WAN 2.2 has been released. Do you have any plans for an update?
What is the best way to apply loras with this setup? I'm using WanVideo Lora Select Multi, but I'm not sure whether I should enable merge_loras and whether or not to use the WanVideo Lora Block with all blocks to true?
This is the first time I've created a Lora for a video. I used your FusionX model. When the generations were generated, I was overjoyed. Thank you. ♥♥♥
Does Phantom Lora work with Wan2.2
I use this model t2v+first frame to generate anime videos, but there are often real-life scenes. What is the problem
Thanks for the effort.
i2v, the woman seems to be doing a stupid tiktok dance with her legs and hips, and the guy behind her is trying to keep in time with her. Have dance and dancing in the negatives. I can see in the posts from other people, the woman are all jiggling. Any suggestions on stronger negatives?
nice generator
I get an error in the node WanVideoSampler
Cannot find a working triton installation. Either the package is not installed or it is too old. More information on installing Triton can be found at: https://github.com/triton-lang/triton 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"
which workflow to use img2video with gguf?
For anyone wondering I have been using both the T2V and I2V FusionX loras for WAN2.2 and it has still been working great, as in it brings the same vast improvements in both quality and speed to my generations. All of my latest vids are made that way.
Does the VACE or PHANTOM version have an NSFW filter? Can I generate sexually explicit and pornographic images and videos?
Out of curiosity, are you also working on version 2.2 of FusionX? Is there a chance it will be released? Because currently with 12 GB of VRAM I only use FusionX 2.1, which is definitely faster than any Checkpoint 2.1, let alone 2.2. If there was a FusionX for 2.2 in the future, it would be fantastic.
Unfortunately no because the models that make up fusionx don't have a new version but other people say u can still use the fusionx lora on the low noise pass with 2.2
Please tell me which package the nodes in workflow belong to. Because I have WanWrapper, the download missing button shows me emptiness, the comfy himself swears at missing nodes
This is/was one of the GOAT. Hopefully everyone realizes that.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.