Hello Everyone^^ ✨
I am so incredibly excited to share this with you all! This workflow was created as a massive Thank You to celebrate reaching 1K followers here on CivitAI!
☕ Get any works I've created "Concept Request", or Order Commission here:
https://ko-fi.com/lumi_nami/commissions
Support my works on patreon^^
(some exclusive NSFW works with sound also available here)
https://www.patreon.com/LumiNami
🤖 What is this workflow?
"This is a seamless, looping video workflow designed to easily switch between multiple image references. It is highly optimized for generating Live Wallpapers and General NSFW anime style videos using the power of Wan Video 2.2 I2V 14B."
Watch the Demo Video to know how to set up this workflow^^
🛠️ Version 1.1 - The Optimization & Fixes Update! ^^
Hello everyone! Thank you so much for all your feedback on the initial release. This version focuses heavily on simplifying the workflow, removing visual bloat, and fixing the pesky custom node errors that popped up on fresh installs! ✨
What's new and improved in v1.1:
🔄 Node Replacements: Swapped out problematic/deprecated nodes causing unknown errors with simpler, stable alternatives.
🧹 Streamlined Interface: Removed MMAudio and complex visual bloat to strictly focus on single-concept generation (1st group nodes). Improved the use of "pin nodes" and subgraphs for a much cleaner look.
⚙️ Better Generation Flow: Optimized SET_GET, Fast groups, and Switcher nodes for easier usage and troubleshooting.
🐛 Bug Fixes: Resolved the known "AIO disconnected" and "save text" errors.
📚 Built-in Resources: Added helpful notes inside the workflow containing used model and custom node list.
✅ Fully Sanity Tested: Ran complete, from-scratch tests across various GPUs (Ampere (3070, A4500), Ada (4060 Ti, 4090), Blackwell (5090)). It runs smoothly from missing node downloads all the way to the final result without a single error! (Note: performance still depends on your specific system).
Read my Patreon post below for more in-depth explanation about this fixed version:
>>> Go to Patreon post <<<
✨ Key Features v1.0
Semi-Automated: Set your prompts, upload your images, adjust the light blue nodes, and hit Run!
Long Video Generation: Create (up to 15s+) videos using Stable Video Infinity (SVI 2.0 PRO).
Better Adherence: Better results and prompt adherence powered by Very Big Video Reasoning (VBVR) Wan 2.2 LoRA.
High-Quality Upscaling: Generates at 720p 24fps (by default), then upscales to a stunning 1080p 60fps final render using SEEDVR2 and RIFE VFI.
Built-in Audio: Video-to-Sound effects generated with MMAudio (with plans to upgrade to AudioX in the future!).
⚙️ How the Looping Magic Works
The core looping logic is heavily inspired by the amazing "Wan 2.2 Perfect Loops" workflow by @Caravel. Here is the simplified breakdown of how it creates a seamless loop based on my understanding:
First, it generates "Main" video which you can completely control the resolution, length, and FPS.
Next, it takes the very beginning frames and the very last frames of that Main video to generate an "In-Between" transition video. (Note: the number of frames is the half of your inputted FPS.)
Finally, it blends those Main and In-Between videos seamlessly together to create the perfect loop!
🚀 How to Use (Step-by-Step)
(1st) Set IMG & Prompts: Go to this group node and paste your prompts into the text box (check inside the nodes for extra tips!). Upload the images you want to use (up to 6!).
(2nd) Generation Settings: Select which of your uploaded images to process (1-6). Here you also set your concept type, name, resolution, length, FPS, seeds, etc on the light blue nodes.
(3rd) Main Video IMG2VID: Choose whether you want the Default or SVI workflow for your Main video. Customize any light blue nodes you want. Important: Bypass any unused group nodes before you click "Run" XD.
(4th) In-Between VID: Upload your freshly generated "Main" video into the Load Main Video node. Right below it, manually adjust the video length by frame count and tweak any light blue nodes if you wish. Run!
(5th) Video Combine, Upscale & Interpolate: Upload both your "Main" and "In-Between" videos here. You can fine-tune the final look by adjusting brightness, contrast, and saturation in the LayerColor Node. (Bypass MMAudio if you don't want sound). Run and Done!
The light blue nodes in question:
🎥 Results
The final video is a FHD 60FPS seamless looping video which the resolution, duration, and FPS depends on your settings. But, by default is:
Main Default (5s) + In-Between (4s) : 1080p, 60fps, 9s
Main SVI (2 batches of 5s) + In-Between (4s) : 1080p, 60fps, 13s
🖥️ Tested Hardware
RTX 5090: Runs very smoothly on Main & In-Between steps. (Note: You may experience an Out Of Memory error during the Upscaling step when upscaling Main SVI video with 2 or 3 batches).
RTX 6000 Ada: Runs incredibly smoothly across all steps, though slightly slower than the 5090.
Other GPUs: Definitely possible! Just always make sure to adjust the resolution or video length to match your system's capabilities.
⚠️ Known Issues (v1.0)
The generated video might look contrasty on "Main SVI batch 2 & 3" and "In-Between" generated videos. Don't worry, this gets slightly fixed and smoothed out during the SEEDVR2 upscale step at the end!
On default workflow to generate (720p, 24f, 4-5s) video, it really needs a massive amount of VRAM. On 3rd & 4th step a 24GB VRAM card is recommended to run it smoothly, while the 5th step requires at least 32GB VRAM card while still need to full model offload 😨.
🌱 Future Updates & Feedback
This is just v1.0, and I plan to keep tweaking and improving this workflow! Please let me know in the comments if you run into any problems or have suggestions to make it even better ^^
🙌 Special Thanks
A massive shoutout to @Caravel for the amazing [Wan 2.2 Perfect Loops] which serves as the core of this heavily modified workflow! And a huge thank you to @Coffee_ICE for helping me out with the SEEDVR2 upscale experiments!
I also want to give a massive shoutout and credit to @NRDX for his incredible Live Wallpaper Style LoRA! This model is an absolute pioneer for this workflow, It has been a huge inspiration and foundation for my generations since Wan 2.1. The results in this workflow wouldn't be the same without it, so please go check out if you're interested! ✨
Your help made this workflow possible >w<
I'll also write the same license that applies when modify the "Wan 2.2 Perfect Loops" workflow below and share the same permission as it should be:
License Information: These workflows are licensed under the GNU Affero General Public License, version 3 (AGPLv3) and constitute the "Program" under the terms of the license. If you modify and use these workflows in a networked service, you must make your modified versions available to users interacting with that service, as required by Section 13 of the AGPLv3.
Read the full license here: https://www.gnu.org/licenses/agpl-3.0.en.html#license-text
Description
What's new and improved in v1.1:
🔄 Node Replacements: Swapped out problematic/deprecated nodes causing unknown errors with simpler, stable alternatives.
🧹 Streamlined Interface: Removed MMAudio and complex visual bloat to strictly focus on single-concept generation (1st group nodes). Improved the use of "pin nodes" and subgraphs for a much cleaner look.
⚙️ Better Generation Flow: Optimized SET_GET, Fast groups, and Switcher nodes for easier usage and troubleshooting.
🐛 Bug Fixes: Resolved the known "AIO disconnected" and "save text" errors.
📚 Built-in Resources: Added helpful notes inside the workflow containing used model and custom node list.
✅ Fully Sanity Tested: Ran complete, from-scratch tests across various GPUs (Ampere (3070, A4500), Ada (4060 Ti, 4090), Blackwell (5090)). It runs smoothly from missing node downloads all the way to the final result without a single error! (Note: performance still depends on your specific system).
Let me know if this updated version runs smoother for your setups! Have fun creating! ✨
FAQ
Comments (39)
Hi, bro, will you release new tutorial?
HI!
Yeah, sure^^
I've reuploaded it. The old upload enters an infinite loop of "Analyzing image.." progress, sorry for that XD
hi，I have been searching for some time, but I couldn't find these several lora, which seem to be hidden. Do you have any alternative solutions?
Fused_NSFW_Dynamic_HN_V1.safetensors (Civitai)
Fused_NSFW_Dynamic_LN_V1.safetensors (Civitai)
Fused_Triple_HN_Blend.safetensors (Civitai)
Fused_Triple_LN_Blend.safetensors (Civitai)
I found it through other means on hf
https://huggingface.co/Kutches/2BasedV3/tree/ed246718f499a228b747379fbf24ec404f775520
@djh515064403973 Well, that's a great alternatives to download those loras. Nice found!
Anyway, actually those loras were from wan 2.2 motion enhancer model page:
- Fused_NSFW_Dynamic_HN_V1.safetensors (FluidCore_high_v1.0)
- Fused_NSFW_Dynamic_LN_V1.safetensors (FluidCore_low_v1.0)
@LumiNami I tried it last night, and the wlsh_nodes node kept reporting errors. There was a compatibility issue. I will try again when I get back from get off work today to see if it can run normally:-(
@djh515064403973 Hmm, I actually haven't run into the wlsh_nodes error before.. either on local or RunPod. Guess it's a conflict with other custom nodes, or might be something else. Since wlsh nodes doesn't need a ton of dependencies, feel free to just DM me later if it's still giving you trouble!
@LumiNami Hello! ^o^I finally finished my work recently and purchased a new graphics card and 64g memory stick. After debugging tonight, I finally successfully ran this workflow! ^o^
I currently have a question I would like to ask you. How should I write prompt words that have the effect of the dynamic wallpaper you made? My prompt words are poorly written, and the generated video effect is quite different from yours (＞﹏＜)
hi how did you make pictures like your style? any tip? plz
Hi! ^^ Honestly, the OneObsession model does a lot of the heavy lifting for my style! XD
Plus a few LoRAs to really push the lighting and looks I like. Maybe you can try using the exact same models and values first to experiment with it! ✌️
I'm still experimenting with this, but...the in-between grey frames, are like 90%+ of the video. So, if you drop in a 24fps 10s video, you are regenerating more than 9s of it which seems extreme.
Essentially, the math behind start frames should account for video length at least somewhat. This workflow must be good for 2-5s clips, that need a huge re-done portion to maintain the loop. But longer videos, why regenerate the whole video for matching 1-2 worth of start/end frames?
Well, that's the main intention of 4th (In-Between) Group Node in this workflow. It's to generate another 3-5s continuation clip from the generated main clip as the looping part, since every generation from main (either basic I2V or SVI have quality degradation overtime). So as the results, In-Between could generate a short clip with high consistency, more frames for loop, and saving vram usage.
And generating (let's say 10s clip) on In-Between may need a MASSIVE amount of vram, assuming the start/end frames changed to 3s (73 frames @24FPS) each so we're only generates around 4s as the loop. Even just generating at 4s 720p 24fps, I already need to turn on block swap for a 32GB gpu. This is why I'm not recommend generating long videos in In-Between XD
But, I think your idea is more towards SVI video rather than the basic. Cuz, with more start/end frames in In-Between means better motion quality (wan22 could know the context better) while also soften the visual quality shifts from start frames (original) to end frames (degraded). [Based on my smol experience]^^
So, um.. are we on the same page here?
@LumiNami hmm, well I didn't try the 4th section. I did try a different workflow, https://github.com/stuttlepress/ComfyUI-Wan-VACE-Video-Joiner which works better. It allows you to set the number of frames and works with like 8 frames. 2x 8 frames is 1s, and it keeps the rest of your original video. Whereas here it's still keeping only 10% of the original video and regenerating 90% of it. If you work hard on the original, with svi or what have you, it still doesn't make sense to me that you would regen 90% of it. You could just generate 2s of video and throw in blank frames for 8s and it would do the exact same thing as using a 10s video in this workflow. To me it doesn't make sense. I think maybe you could try looking into how the other guy does it if you want to improve your workflow further. I used 2x the same video (10s x2, total 20s), used the loop function, and 1s of inbetween frames. Yea, because these dialogue/wallpaper vids like in your own examples have low motion, again...idk 90% regen is not only taking a lot of time, but vace results not as good as original either, especially if you need to lower the resolution or other. Just my 2c, can ignore it if you like.
I'm the newbie in comfy, but I installed all models, generated 1st step photo, but idk what to do next cause if I just generate video with some params it's generation still photo, how I should exactly use this workflow?
Hi there!^^
Um, have you watched the full guide tutorial on how to use this workflow yet? If haven't, then I recommend to watch these videos first:
Wan 2.2 Reasoning Loops v1.1 Full Guide - Part 1
Wan 2.2 Reasoning Loops v1.1 Full Guide - Part 2
Speaking of generation still photo instead a video, how do you set up those parameters exactly?
@LumiNami Why can I only see a picture with music playing?
@CCCino Um, that's kinda strange. It might be a playback compatibility issue, cuz I used H.265 10-bit for that video XD.
Anyway, If you're still having trouble, I uploaded an H.264 version to my GDrive here instead, this one should play correctly: Wan 2.2 Reasoning Loops v1.1 Full Guide - Full
@LumiNami Thank you, this can be watched
@LumiNami Hey im still getting a problem on generate final video group, even with the right lora im still getting a vommit transition when it hits the low model sampler, any idea how to fix it?
Your workflow is very good. One thing that's sad is that several people and creators of this type of content use my LoRA, which is basically the pioneer in this style, but none of them reference my model. This saddens me a little because, imagine, my model has a great influence on what is done here. Referencing it ensures that the model continues to be updated by me; if it's forgotten, improvements never happen. Anyway, consider this type of reference for your next posts. I know this one is older, etc., but it would be nice, and I'm almost certain that DaSiWa also uses my model, but there's no reference to it. Please don't misunderstand me.
Oh my goodness, I am so sorry! (╥﹏╥)
Please don't misunderstand me either, I have immense respect for your work and your LoRA is absolutely essential for this workflow! Well, actually I've generated so many animations using your Live Wallpaper LoRA since wan 2.1.. which is also a foundation model that keeps me motivated even now! ^^
So, what happened was from a bug when I posted the workflow at that time, I actually attached your model and others (especially DaSiWa).. but the site just didn't display anything in the suggested resources box. And.. I completely forgot about that till now ; _ ;
But, I'm updating the page and the 'Special Thanks' section right now to make sure you get the proper credit you deserve. Thank you so much for creating such an amazing model, and for bringing this to my attention!
(Edit: It should be appeared now)
@LumiNami No problem, I really know you like the model because I always see your work. I mentioned it because I never see people referencing the Live Wallpaper model, but thank you very much, LumiNami.
I haven't had time to play with it yet, but a little suggestion. Perhaps put the models the workflow requires in the description somewhere, or inside the workflow within a note? I think everyone would be really thankful! :)
Ah, sure! I'll add the required models link to this description page soon. Thanks for the suggestion @tsurupettanlolis ^^
Simply Amazing!! After trying many loop workflows it never managed to give such smooth loop for me except for this one. At first I almost give up on this cause my 12GB 4070Super can't handle the default settings. But changing to RoPE to chunked and some Values in BlockSwap I can get a 720p Loop around 15 min. Which great considering you can preview and run each steps anytime and not needing to run through whole workflow if results are bad. Thanks for the workflow (๑˃ᴗ˂) b
Hello, I have been studying this workflow, but the quality of the currently generated video is not as good as the demonstration video. I wonder if you can share some notes on how to write prompt words and video generation. If possible, can you record a complete sample video? Thank you very much:）
@djh515064403973 Hmm can you clarify a bit on what you meant when you say quality. Is the output blurry? messy or is the output animation not really fit with what you have in mind?
@MilkShook After many attempts, I couldn't get the videos generated in the first and second steps to form a complete loop like the demo video. Additionally, there's a missing image input in the“ Get Image Size & Count ”node within the“ VACE Logic ”node in the third step, preventing it from running. Although I managed to connect it myself and got it working, I'm unsure if this issue caused problems with the generated between-video or if it was due to the input prompts. Therefore, I'd like to see a detailed demo before continuing to run this workflow.
@djh515064403973 I don't experienced what you experienced. I'm not doing anything complicated either, my process: Open ComfyUI / make sure you have comfyui manager installed -> Drag & Drop this workflow to comfy -> If there are any missing models/extensions error will pop up with clear names, search and download them -> activate each and generate each steps. The author already included video tutorial btw in detail.
@MilkShook Well, I'll study by myself how to solve the problems I've encountered.
Hi, before i start let me just state that I'm a beginner at generating videos.
I've tried following your setup to the letter (Only change I've made is using Block swaps so my pc can run it).
It generates Main Previews fine, but when i attempt to create Main Final videos, the images just sort of fades out into static/smudge.
Is there any fix to this?
Is it possible to use gguf models here instead of fp8?
Um sorry for super late reply, kinda busy last month.
After a few tests with the workflow, it's possible to use gguf! but only for MAIN Video Group only, IN-BETWEEN will just throw a random error which I don't know why. You can just swap the model loader nodes from ComfyUI-GGUF by city69
I am using RTX5070 12GB and am getting an OOM error. How can I lower the memory usage?
Well, that depends on which step you get the OOM.
- On Main SVI: Since I neither utilize "chunking" nor "MultiGPU nodes" on the workflow, here's what you can do..
1) swap/change the "Diffusion Model Loader KJ" to "UNETLoaderDisTorch2MultiGPU" from comfyui-multigpu custom nodes.
2) Set the virtual vram amount depends on your free ram.
3) Connect the nodes to "Patch Sage Attention KJ" if using it, then connect both high & low model to the lora nodes respectively.
- On Main Video & In-Between: For these, you can just enable Main Block Swap.. and adjust the value until it's running smoothly. Or change the rope function to comfy_chunked on each wanvideoSampler
- On Combine due to SEEDVR2 upscale: On default, I already offloaded all of the model to ram, enable encode & decode tiling :)
Will you be making a native workflow based on this in the future?
Well, maybe! I planned to use subgraphs on all steps later XD
Interesting workflow, even if I managed to get all the necessary files it gets stuck on WanVideo Sampler x.x
Hello!
Will you make a workflow for Minimax H3 too?
Thanks!