⚡ Recommendation
You can use my custom checkpoint V2 NSFW, Lightning included:
➡️ WAN 2.2 Enhanced – Lightning Edition
for wf lightningedition long v2.1
https://civarchive.com/models/2053259?modelVersionId=2477539
Or for wf svi , no Lightning included:
➡️ SVI NSFW
https://civarchive.com/models/2053259?modelVersionId=2609141
Update: 2 or 3 KSamplers
The triple-KSampler workflow added dynamism to avoid slowdowns, but it could cause issues in certain sections, such as high brightness or other bugs. To fix this, simply enable the Double KSampler in the problematic section and change the seed.
🟣 uptade SVI LoRAs – Strengths & Weaknesses
⚡ Overview
✅ Strengths
Best solution for making long videos
Excellent transitions between video segments
Reduced degradation compared to other solutions
Strong character coherence: the model retains information from the previous video, helping maintain consistency
⚠️ Weaknesses
Weaker prompt understanding
Weaker camera understanding
Videos are less dynamic
Sometimes slow-motion effect
(can be improved with proper Lightning LoRAs and dynamic prompts)
🟣 SVI LoRAs – Download Links
⚡ Download
Note: Both LoRAs must be loaded manually in your workflow.
🟣 Suggested Lightning LoRA Combos (Optional)
⚡ Overview
You don’t have to use these Lightning LoRA combos. They are optional and allow you to fine-tune motion and degradation.
You can also use other Lightning LoRAs or assign different combos per video for more control.
🔥 Combo 1 – More Motion (Rapid Video Degradation)
High LoRA →
https://huggingface.co/Kijai/WanVideo_comfy/blob/709844db75d2e15582cf204e9a0b5e12b23a35dd/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Weight: 4
or
💜 Combo 2 – Less Image Degradation
or
or
🟢 Combo 3 – Balanced Motion / Moderate Degradation
High LoRA →
https://huggingface.co/Kijai/WanVideo_comfy/blob/709844db75d2e15582cf204e9a0b5e12b23a35dd/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Weight: 3
Low LoRA →
https://huggingface.co/Kijai/WanVideo_comfy/blob/709844db75d2e15582cf204e9a0b5e12b23a35dd/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
Weight: 1.5
If you want to create several-minute-long videos while maintaining high quality, it is possible, but it will take a very long time. You just need to avoid using Lightning LoRAs and use the full model instead.
🧠 Advanced Usage Tip
You can disable global Lightning LoRAs and assign different combos per video:
Combo 1 for Video 1, Combo 3 for Video 2, Combo 2 for Video 3.
Each combo produces different motion and degradation behavior.
If you don’t use Lightning LoRAs, degradation will be even lower, but generation time will be very long.
🧠 Dynamic Prompts – Better Control & More Motion
You can use dynamic prompts to have better control and this helps to make the video more dynamic.
You just need to give this example prompt to an LLM like ChatGPT. It will be enough for it to describe the image you have and what you want as a video while keeping the same structure of the prompt in the following example.
⚠️ This will be a NSFW prompt; ChatGPT will not accept it.
You can use GROK https://grok.com, which can make NSFW prompt modifications.
For more examples of prompts with different poses, check:
Enhanced FP8 Model.
Give these prompts to GROK with the dynamic prompt structure if you want.
Example 1
(At 0 seconds: Wide shot showing a slightly overweight man casually walking down a city street, camera fixed in front, urban environment with buildings and cars.)
(At 1 second: Suddenly, a massive shark bursts from the pavement ahead, looking terrifying at first, pavement cracking, dust and debris flying, camera from side angle.)
(At 2 seconds: Medium shot from the side, the man stumbles backward in shock, while the shark dramatically slows down and strikes a comically exaggerated sexy pose, revealing large, exaggerated shark breasts, covered by a colorful bikini.)
(At 3 seconds: Close-up on the man’s face, eyes wide in disbelief, as he turns to look at the shark, small cartoon-style hearts floating above his head to emphasize his amazement, camera slightly low-angle.)
(At 4 seconds: Dynamic travelling shot showing the man frozen in the street, the shark maintaining its sexy pose, water splashes and debris still moving realistically, urban chaos around.)
(At 5 seconds: Wide cinematic shot pulling back, showing the man standing in the street, staring at the bikini-wearing shark with hearts above his head, epic perspective highlighting absurdity and humor.)
Example 2 – Anime NSFW
(At 0 seconds: The couple in a cozy bedroom, anime style, soft lighting highlighting their intimate embrace, her back arched slightly as he positions himself.)
(At 1 second: The man’s hips moving rhythmically, the head of his penis sliding effortlessly into her vagina, her body responding with a gentle, fluid motion, anime-style motion lines emphasizing the smooth penetration.)
(At 2 seconds: Her back arching deeply against him to intensify the pleasure, hips swaying with each thrust, breasts bouncing subtly, small hearts floating around them to capture the erotic energy.)
(At 3 seconds: Her face, eyes closed in bliss, a soft moan escaping, hands resting behind her head, anime-style blush on her cheeks, the air filled with a seductive aura.)
(At 4 seconds: The man penetrating her deeply, her body moving in sync with his, the bed sheets slightly rumpled, the room’s warm lighting enhancing the intimate, lustful atmosphere.)
(At 5 seconds: The couple locked in a passionate embrace, the scene exuding vibrant, seductive energy, anime style with smooth lines and soft shadows.)
findperfectresolution
https://github.com/ashtar1984/comfyui-find-perfect-resolution
rgtree
https://github.com/rgthree/rgthree-comfy
kj-node
https://github.com/kijai/ComfyUI-KJNodes
iamccs
https://github.com/IAMCCS/IAMCCS-nodes
🔧 What’s New – V2.1
The workflow is now much simpler than before.
You can create videos up to 20 seconds.
New sections can be added very easily by copying and pasting an existing section.
🎨 New Node: PainterI2VAdvanced
🔗 https://github.com/princepainter/ComfyUI-PainterI2Vadvanced
Main Parameters
color_protect
Enables or disables color drift correction.
Recommended: True.
Prevents green tint, darkening, and color distortion.
correct_strength
Controls the strength of color correction.
Higher value = stronger color correction
findperfectresolution
https://github.com/ashtar1984/comfyui-find-perfect-resolution
_______________________________________________________________________________________________
New FunVace workflow: it's extremely easy to use. If you have separate videos and want to join them with a seamless transition, use this! Thanks to Bob once again, excellent work: https://civarchive.com/models/2277993?modelVersionId=2563927
Workflow Update – Lightning Edition V2
📌 New Version Available
The Lightning Edition V2 is the latest update of my workflow.
This version is cleaner, better organized, and scene transitions are improved for smoother results.
🔧 Fixes & Improvements – Lightning Edition V2
Workflow fully reorganized and clarified for better readability.
Scene transitions improved for smoother results.
Optimized nodes and settings for easier use.
Confirmed compatibility with the Lightning Edition V2 model.
🔹 V1
Version without Super Power LoRA
A version without Super Power LoRA has been added for those who are more cautious.
If you want to try the Super Power LoRA without using the folder included in the ZIP, check the GitHub page:
➡️ GitHub – rgthree workflow
Tutorial
The workflow is divided into 6 sections:
Section 1: generates a 5-second video.
Section 5: handles transitions and interpolation.
Important notes:
Running only sections 1 + 2 will not merge the video.
To create a 15-second video, run sections 1 + 2 + 5.
Interpolation can also be enabled in section 1 if you prefer.
You can use different prompts and LoRAs for each section.
Resolution is automatically adjusted to avoid cropping.
💥 Super Power LoRA – V1
Requires the custom rgthree node:
➡️ rgthree ComfyUIInstall via git clone or the Manager, then replace the folder with the one included in the ZIP.
Allows you to save and load templates, very useful when working with multiple LoRAs, especially for WAN 2.2.
A version without Super Power LoRA is also available for those who prefer.
For the version with Super Power LoRA, see: GitHub – rgthree workflow
✅ Tips – V1
Keep the default settings for fast results.
Increase the steps if you want better quality.
Description
FAQ
Comments (114)
I have an issue where the video loses contrast and saturation after like 3 frames. Starts as a vibrant shot and ends dull and grainy.
Same problem here. But an even bigger problem is that the information is an absolute mess. Too much information that deals with too many things all at once. There should be like a baseline starting point for every workflow with clear labels and names.
Hello! Download one of the preview videos to your desktop, for example, and drag and drop it onto your ComfyUI interface. This will load the exact parameters used to create it, even if the default ones are the same.
@taek75799 I unfortunately also had issues with that. The Comfy Node Manager can not find the "Super power Lora loader" at all or whatever it is called. I'm not sure if it's because it's deprecated or something.
I am having the same issue, I used the lightning workflow and it worked for me perfectly, but trying the SVI and the video loses everything after 2-3 frames.
I think it has something to do with the lightning loras, the names that I see in your examples are very different from the names I find in the links, I think maybe the links are not matching the right loras.
Trying the 4th different attempt of lightning lora combinations, not sure if it will work but we will see
@OhNoMyChickens Hello! The Lightning LoRA names are indeed different because I rename them. Most of my examples are made using Combo 2. Make sure to place them correctly, along with the SVI models. Otherwise, just generate your video (even in low quality), save the workflow (wf), and send it to me so I can check it: https://www.swisstransfer.com/fr-fr"@Derpxten
@taek75799 I'm at work but I'll resume my experimentation with this once I'm home. I'll check out combo 2 and whatever else I may have missed. And if I can't figure it out I'll drop a video. Thanks!
@OhNoMyChickens Feel free to send me the workflow using the link I posted; it's the easiest way for me to check if everything is correct.
@lorenzo881743 @OhNoMyChickens Have you tried this?
Color correction: 0.2000
Ksampler Low:
Steps: 12
cfg: 2.0
sample name: euler
scheduler: simple
Start step: 0
End step 6
Ksampler high:
Steps: 12
cfg: 1.0
sample name: euler_ancestral
scheduler: simple
Start step: 0
End step: 12
That managed to fix the grainy and lighting visual issues for me. Though that's for me using some random anime image as an inputted image. Also I'm still generating 5 second gifs, so I can't speak for anything past the 5 second gifs.
I got this error
"Prompt outputs failed validation: Required input is missing: images VHS_VideoCombine: - Required input is missing: images"
when trying to use the Wan lightning V2.1 wf. There's no missing nodes or anything
It looks like the error is due to some nodes being bypassed. Try this: move all Get and Set nodes out of their groups so they remain active at all times. To enable or disable nodes, simply click on them and press Ctrl + B.
@taek75799 I'm sorry but i notice that's because the 16fps node image somehow made that error but now after i bypass that node the error only Prompt has no outputs:. I make zero change tbh from the file that you provided.
Is there other workflow that maybe you recommend me to use?
@indonesiangirlonly I mainly use these nodes: https://github.com/princepainter/ComfyUI-PainterLongVideo. I also use PainterAdvanced for color correction from the same author. You just need to replace PainterI2V with PainterAdvanced."
@indonesiangirlonly There is a workflow using PainterLongVideo."
@taek75799 @indonesiangirlonly I found that the node causing the problem is the Get_16FPS node in the UPSCALE group. When I moved that node out and unbypassed it, everything started working.
@AngelOfTheGrove I did by pass that node but now the error change to something about no prompt input. Do you by chance can share the workflow link?
@indonesiangirlonly Redownload the WF, I've updated it. You shouldn't bypass it; instead, move it outside of the group. I think it's best not to put Get and Set nodes inside a group that needs to be bypassed, as it can cause issues depending on your ComfyUI version or other factors, I'm not sure.
@taek75799 Hey man thanks it works now, but somehow the color degrade. I didnt make any change and try several resolution the color still degraded or more like there's lot of noise.
I don't know maybe something wrong in my set up but I guess i just use the lightning long edition v2 for now.
@indonesiangirlonly After generating the video, save the WF and send it to me here: https://www.swisstransfer.com/fr-fr
@taek75799 sorry it seems my country can't access the link you provided. I can try to upload it on google drive if it's ok.
Hi just wanna ask where can I download this model cant find it thanks
@taek75799 sorry really, there is 2 load diffusion model node here one is high and the other is low should i load that link you give to those two?
Also in New subgraph node i cant find this exactly model in the workflow:
Wan_2_2_I2V_A14B_HIGH_test-derniere-version-selforcing.safetensors
wan2.2_i2v_A14b_low_selforcing_last-version_rank64_lightx2v_4step_1022..safetensors
should i put there the combo lighting you mention?
also last one I also cant find the one i should lode in Unet Loader GGUF high and low
Thanks in advance and sorry for very noob question.
Yes, you need to load both the Low and high models. You don't need the UNET Loaders; those are for loading GGUF models, which aren't available here for SVI. Regarding the Lightning LoRAs, ignore the model names in my workflow as they have been renamed. Just download Combo 2 along with the SVI LoRAs.@lazysonnie
@taek75799 Thanks i was able to generate a 5 sec videos, just wanna ask again this time about prompting. The example you give like the Example 1 should i put that on section 5 then the description of the prompt is at section 1 and 2? What i did is just prompt in section 1 to generate 5 sec vid. Thanks again!
@taek75799 also other question will my 3060ti with 32gb ram pair with i512600k good enought for this?
@lazysonnie I believe the 3060 Ti has 8GB of VRAM. Try the Q8 version first, as it's the best. If that doesn't work or if it's too slow, try Q6_K, and if it's still the same, go with Q4_K_M, it will be the least resource-intensive
@taek75799 Hi its me again, just wanna ask if i can put other loras here and what kind of lora can i use and where to put it is it in the top of each scene? Thanks again
@lazysonnie Apply the desired LoRA to the scene of your choice by setting both the LoRA High and Low. You just need to lower the LoRA strength if it degrades the video quality.
@taek75799 hi! Thatlink takes me to a normal fp8 gguf model, not the "consistent_face" version. Confirming they're the same thing? In other words, this model is being used to keep the face consistent? and "Wan2_2-I2V-A14B-HIGH_SVI_consistent_face_nsfw_fp8.safetensors" isn't some dedicated model that we can't find?
hi again,
do you set noise_seed to random for sampler 1 and sampler 2 before generating, and only lock the seed on the final sampler? Also, is this now your personal go to workflow, or mainly an example setup you shared?
To make sure I understand it correctly, is the flow basically: you VAE encode the initial image, feed that latent into the first KSampler subgraph to generate the first 5 second clip, and then reuse the latent output from that first KSampler subgraph as the anchor sample for all subsequent sequences/extensions? If that is correct, do you deliberately keep the initial clip more still or low motion to make a cleaner anchor? If I am not mistaken, KJ and others often keep using the initial image VAE latent throughout the whole video, did you test that approach vs your method and find yours more consistent?
One more: why did you decide against using IAMCCS WanImageMotion SVI? What did your tests with it show, and what results made you drop it (or keep it out of your main pipeline)?
Anyone is free to answer. Thank you for the time you have invested into your models and workflows.
I abandoned IAMCCS because some people noticed contrast shifting issues with every new video. In my opinion, the motion it provided wasn't good enough. I preferred the 3 KSampler method, which completely removes the slow-motion effect, though it introduces other issues like excessive speed or artifacts. To fix this, you just need to change the seed.
SVI allows for great consistency: it keeps the initial sample and references the previous one. So, if a face disappears on the last frame of a video, it won't cause any issues.
Personally, I use three workflows: 2.1 is ideal for 5-second videos. If I want to make longer videos quickly, I like using the 3 KSampler. If I want to create something truly high-quality, I use workflow 2.1, generate only 5-second clips, and then assemble them using FunVace.
@taek75799
Thanks, that clarifies a lot. Two quick follow ups I am still unsure about:
1. On the anchor mechanics, is the flow actually: VAE encode the initial image, feed that latent into the first KSampler subgraph to generate the first 5 second clip, then reuse the latent output from that first KSampler subgraph as the anchor for all subsequent extensions?
2. If yes, do you deliberately keep the initial 5 second clip more still or low motion to create a cleaner anchor?
Also, did you compare that against the alternative approach some use (keeping the initial image VAE latent as the anchor throughout the whole video), and if so which one gave you better identity stability and fewer drift issues?
@juliusmartin 1 and 2, yes, that's basically it. But I didn't invent it; that's just how SVI works. I've tried every method, and SVI is by far the best one out there. With the right resolution and without using no Lightning, you can produce videos several minutes long without any real degradation, but most people won't bother doing it.
I have been using lightning 2.1 workflow previously and it was working fine. Recently I re-downloaded the same workflow and the videos seems to have loss of clarity/contrast from initial frames. It still works fine with the previous lightning 2 workflow.
What could have changed? Only thing I did was unbypass the Get_16fps node
Hello! Did you increase the motion? Try lowering it to 0.
Yes I changed it to 0 and I am still getting the same issue
@rahulr1199 there are so many issues with ComfyUI that I'm not really sure. Maybe try installing a fresh version of ComfyUI without deleting the previous one Version 2.1 shouldn't normally have any contrast issues or anything like that it was specifically designed to be very good in that regard. Otherwise, try v 2 if it works, the only difference is just a single node.
thanks for remembering about the get/set nodes <3
I saw the wan first-middle-last frame package features an extract last images node - set it to one, works pretty well to just continue into another workflow;
otherwise I'm impressed that for seven steps, it's not that much slower (around 38mn for 18 seconds at 720p on a 5060ti)
What files would you download to properly run this workflow with just 8gb of VRAM? When there are 100 different versions to download and options, it's confusing which ones I should download, and which ones I need. I was previously using Q3_K_M.gguf.
I also already have rgthree installed and updated, but it's saying I'm still missing rgthree components (I think just Super Power Loader LORA)
Hello! For long videos, the SVI version will work best with either the SVI WF or the SVI 3 KSampler WF. Otherwise, the NSFW V2 version is recommended for shorter videos using WF 2.1. However, I think the NSFW V2 version would be better for you since it’s available in Q4_K_M. WF 2.1 doesn't use 'Super Power LoRA', so there shouldn't be any issues."
@taek75799 thank you! Just to be certain, which workflow should I be downloading? I see 2.1 workflows, but they also say for long videos when I'll likely just be making short videos for now. I'll try wan22I2VNativeEnhanced_withoutSuperPower and wan22I2VNativeEnhanced_painterLightning. For the models I got wan22EnhancedNSFWSVICamera_nsfwV2Q4KMHIGH.gguf and wan22EnhancedNSFWSVICamera_nsfwV2Q4KMLOW.gguf
@heyheyhoho Everything looks good, good luck!
@taek75799 workflow has been working great, thank you 👌 - if a generation is not coming out satisfactory and I change prompt and seed around, there's not much variation. So in those cases, what could I change to get a different output? If i change steps or CFG at all, I get glitches. I suspect those being locked into place are restricting the outputs to not have much variance, but I'm not sure what I can do
I got a 5080 & 32 GB ram. It works pretty wall to generate a 5 second video, but I haven't tried longer gifs at the moment since I'm having trouble generating good gif. As in, the gif generates without much image issues, there's just weird physics, which I'm pretty sure is a LoRA weights or prompting issue, but it could also be the input image... For anyone who's achieved solid results, what was your prompts, LoRA(s), & image(s)?
In your latest SVI Triple Ksampler workflow, where would I put in "normal" Loras that help with understanding of various concepts like dancing, various NSFW stuff, etc.?
Hello! Right above each video section where the prompt is located, there are 3 LoRAs to place. The top two are the High LoRAs, and the bottom one is the Low one.
@taek75799 do you use both highs and the low? As soon as I add any Lora all I get is garbage out
I am trying to use full models without loras but all generations come out blurry, do i need to change something in the steps to make it work without? thanks
HI, which model? The SVI? If that's the case, don't forget to add the Lightning LoRAs.
@taek75799 the SVI model no lightning lora, I read I could use them without those lighting loras but the video generation is all blurry
@kikiazusa856 Lightning LoRAs are used to speed up generation, meaning a minimum of 4 steps and a CFG of 1. If you aren't using Lightning LoRAs, you need to set it to 20 steps and a CFG of 4
@taek75799 thank you so much ill try!
When using both High and Low loras in the lighting edition long V2.1 workflow, how to connect the loras model/clip to the prompt node?
Hello! Everything is already connected using the 'Set' and 'Get' nodes.
@taek75799 i am getting this error if i dont manually connect the nodes of images
Prompt outputs failed validation: Required input is missing: images VHS_VideoCombine: - Required input is missing: images
Shouldn't i not have to manually connect?
@guyanime12471 Hello! Try updating your ComfyUI nodes. The KJNodes had an issue with 'Get' and 'Set' in the previous version. Also, remember to redownload the WF: I updated it by moving the 'Get' and 'Set' nodes outside of the groups, which might help. If it still doesn't work, check Painter's repo; there is a WF available without the 'Get' and 'Set' nodes.
@taek75799 the fix that works for me is directly joining the image output of the video20 sec node to the Vhs video combine node; skipping the upscale node entirely. Then it works totally fine and i dont have to do any updates.
Hi, I just opened your workflow and got the error: "PatchSageAttentionKJ: No module named 'sageattention'".
I’ve tried installing it by manager but it still doesn't work. How can I fix this man?
Hi! I’m seeing a lot of lora key not loaded warnings, for example:
lora key not loaded: diffusion_model.blocks.0.cross_attn.k_img.diff_b lora key not loaded: diffusion_model.blocks.0.cross_attn.k_img.lora_down.weight ...The video does generate, but when I describe the video in 0–5 second segments I get strong flickering / seizure-like flashes. If I describe the video more holistically, it works much better.
This makes me think there may be a LoRA compatibility or setup issue. I’m using combo 3, but I’m not sure if I’ve configured it correctly. Workflow image: https://imgur.com/a/vLXs9AW. Any guidance would be appreciated.
Hello! The 'Lora key' message will appear all the time if you use the SVI LoRAs; it's the same with the 2.1 LoRAs. This is not an issue.
@taek75799 Thanks! Your models and workflows have been a lot of fun to use. Really nice work.
Hello, Do I have to load lora on both high lora loader node or just choose one of them? Thank you!
Hello! It depends, but sometimes the LoRA won't be properly applied if you don't set it on both LoRA High nodes.
@taek75799 Thank you!! Much appreciated!!
So, If I choose Combo 3 – Balanced Motion / Moderate Degradation, I should use the same LoRA file for both High and Low nodes, right? Thanks!
Hello, exactly!
I would like to use no lightning lora.
Where do I add the high-steps and where the low-steps? Is it even possible in this workflow?
For the SVI workflow (the one that isn't TripleK), I recommend using motion_only (prev_samples) at the WanImageMotion node in the first subgraph. It’s more stable and keeps the style from drifting too much.
Also, do you want to try something a bit 'out there'? Try turning off or reducing the weight of the Lightning (LoRA) on the high noise while keeping the steps the same. You'll be surprised by the results... give it a shot. It’s a 'make it or break it' kind of thing.
Hi guys! I replaced the Lightning Low 1022 LoRA with the 1017 version. I was having some issues with brightness shifts (sudden increases) during certain transitions with the 1022. I've updated the WFs and the descriptions.
Where are the high and low models? i only see a file ending with H
Hello! H for High and L for Low
@taek75799 Sorry. Im just not seeing where to download the models. The page only has the high model.
The only model available is: wan22EnhancedNSFWSVICamera_nolightningSVICfFp8H on https://civitai.com/models/2053259?modelVersionId=2609141
@Peleus Here is the Low model: https://civitai.com/models/2053259?modelVersionId=2609148
Hello! I apologize if the English sounds strange since I translated it from Japanese using DeepL.
I tried lightningedition longV2.1, but running it directly generated a still image and a 0-second video. As a ComfyUI beginner, it took me about 10 hours to pinpoint the cause: the Select Images node and ImageScale node connected to the output from the VIDEO3 15SEC area weren't bypassed, so only the final frame reached the Video Combine node.
Additionally, the ModelSamplingSD3 node seemed to be causing the grainy quality and dull colors. I tried it with the values almost unchanged (since I still don't fully understand what the nodes and parameters mean ^^;), but bypassing both ModelSamplingSD3 nodes allowed me to create a video that retained the input image's color and quality.
I thought it would be a shame if people who downloaded it tried it and rated it negatively because no video was created, so I wanted to report this. Thank you for the wonderful workflow.
Hello, and thank you for your feedback. I like Japan very much, just like many people all over the world.
@taek75799 Will there be a fixed version?
(ぼくのポストがちゃんと英語の人たちにも読めるように翻訳されたようで安心しています……
I'm relieved that my post seems to have been properly translated so that English speakers can read it too...)
@NRubric Hello! To be honest, I can't fix something that is already working for me and for others as well. I think ComfyUI still has many bugs to iron out. Don't hesitate to change some values in the WF my settings might not be suitable for everyone.
@masorix380 Yes, your comment was perfectly translated.
Thank you! Exact same problem solved. the ModelSamplingSD3 i change 5 on both.
Just trying to keep straight what files all go where in the workflow. There's a high/low GGUF from https://civitai.com/models/2053259?modelVersionId=2625965, a high/low diffusion model with consistentface in the name that I assume is https://civitai.com/models/2053259?modelVersionId=2609141 but I'm not sure, a high/low SVI lora which is https://huggingface.co/Kijai/WanVideo_comfy/blob/main/LoRAs/Stable-Video-Infinity/v2.0/SVI_v2_PRO_Wan2.2-I2V-A14B_HIGH_lora_rank_128_fp16.safetensors I think. But then there's a high/low self forcing which I don't see at all. Can you clarify?
The triple KSampler workflow gives me very unstable results on certain images/prompts unless I drop the CFG to 3 (or sometimes needing even lower). Sometimes it's a glowing area which occurs when a block of color moves quickly (like bright green appearing where purple was previously), or sometimes a total consistency breakdown.
This seems related to the single high first pass. Do you have any suggestions? Would an additional low pass before switching to lightning be an approach?
I also have this problem sometimes; changing the seed seems to help. To address the brightness issue, I swapped LoRA Lightning 1022 for LoRA 1017. It seems to handle this better, but unfortunately, it still happens occasionally. A perfect solution hasn't been found yet. I tried the Clownshark KSamplers, but they don't seem to work with SVI.
@taek75799 I've tried both 1017 and dozens of random seeds, but those just generate a different object/pattern in the same color in the same location.
@wangsouffle I just had the same problem with a video: in the third section, a high brightness appeared. I managed to get rid of it by using only two KSamplers in that section. So, you should switch from three to two KSamplers when this happens; apparently, the dynamics remain the same.
If you want to automate the process, in the subgraph where the KSamplers are, add the WanImageToVideoPro node with the two KSamplers (using the same settings as any dual KSampler for Wan). Make all the connections (model, positive, and negative). At the latent output, connect a Lazy Switch from the KJ Nodes pack to False, and do the same for the third KSampler to True. Then, connect the switch to the latent and sample output of the VAE Decode node. After that, you will just need to choose between 2 or 3 KSamplers.
You’ll simply have to choose between two or three KSamplers. If you can't manage it, I'll share my WF, but unfortunately, it will be made with Super Power LoRA
@taek75799 btw, this approach has been working well. Possibly a future addition to your published workflows?
@wangsouffle It seems to be working well. I will probably publish the workflow on Saturday
Hey man! On longer vids with multiple clips, how do I run the upscaler without having it OOM?
Hello! How much VRAM and RAM do you have? What resolution did you set, and what value did you choose for Scale by? You could also try adding a Purge VRAM node right before the upscale.
@taek75799 VRAM at 24GB (3090 if that matters), RAM at 64GB, scale by .5, using remacri_extrasmoother (just in case that's something I changed), nearest-exact method. Initial image is 512 x 768 (sometimes 512 x 512 but that doesn't work either). It runs the upscaler okay with just one clip. I should also state I put my clips at 97 frames. I'll try putting a purge node before the upscaler, thank you!
@The_Utulek_Complex I’m always amazed by how much results can vary from one user to another. I can upscale to 1024x1024 with a scale by of 0.5 as well, using 16GB of VRAM and 64GB of RAM, even on a longer video.
@taek75799 Hmmmm... Well, I think a purge is the big trick. I'll let you know how it goes! Thank you again! :)
@taek75799 I added a VRAM purge node right after "get video" in the upscaler section. It could also be that I needed to switch out the interpolation (the original with the workflow wouldn't load). So I'm using RIFE VFI. ckpt rife47 , clear_cache_after_n_frame: 10, multiplier: 2, fast_mode: false, ensemble: true, scale_factor: 2.
@The_Utulek_Complex this may not be a RAM issue, but not having enough swap configured.
What is the LoRA
wan_2_2_i2v_A14B_HIGH_test-derniere-version-selforcing.safetensors?
Same question here...
I have a question. If I use the SVI model and the SVI loras, do I have to set the SVI loras individually in every 5-second clip and set the global loras to zero? Or is it sufficient to set them globally?
Thanks for this Workflow.
Hello! You can define it in the global LoRAs, and they will be applied across all sections. I'll be releasing a new workflow in an hour; you'll be able to choose between 2 or 3 KSamplers.
somehow my result always blurry and too slow mo. i use svi fm nodel and default setting on the wf. any advice pls?
Could you please share the workflow after generating the video so that I can check it?
@taek75799 okay, check ur DM
@taek75799 hey, thanks. it works fine now, just came across some OOM during upscale. would it affect the result if i just use lower scale_by, ie., less than 0.25?
@raveGoingDark I don't know, I've added the upscale but I never use it, it would be worth a try.
@raveGoingDark How u repair this blurry? D:
@Liqvior bump the resolution a bit. use 2nd option lora combo
Bro, i have issue with ModelSamplingSD3 HIGH and LOW nodes. It says
Prompt execution failed
Prompt outputs failed validation: ModelSamplingSD3: - Required input is missing: model ModelSamplingSD3: - Required input is missing: model
I haven't touched anything in the workflow, just uploaded models. Help please...
This is lightindedition V2.1 Workflow
Solved, i'm stupid😁
I'm also stupid, cos I'm getting the same issue. What did you change? :D