[Edit:
Version v5.0 works with latest comfyui (v0.15.0).
If you have any problems, please refer to the FAQ at the bottom of the page or have a look in the comments.
Many thanks to everyone who tested this workflow. Thank you very much for the many inquiries and, of course, for all the knowledge and experience you have contributed. here👍🙂
Special thanks to:
@SeoulSeeker for the "Dead Simple MMAudio" workflow wich are the basis of the audio part here,
@taek75799 for the really well working enhanced models
@Bakazaya pointing to the color issue in version v3.0 and running lots of tests,
@bluntfeather sharing latest experiances with installing Comfyui-Easy-Install,
@nitrovtx for remain persistent in matters of quality and running a lot of tests,
@Icey64 for providing the link to "Comfyui-Easy Install",
@boinobin730 for asking for a First to Last Frame option, running pre tests and responding fast as hell 🙂 and
@SnowShoes311 thank you so much again for all your buzzing 😋]
Features:
Optimized Wan 2.2 workflow, runs perfect on RTX 3060 12 GB VRAM GPU and 32 GB RAM,
"Text to Video", "Image to Video" and "First/Last Frame 2 Video" generation in one workflow, all with easy audio generation,
easy installation/model downloading, all necessary sources are specified,
easy to use workflow, clearly structured, all necessary steps are explained,
easy switches for mode selection,
easy prompt selection for fast prompt creation/testing,
easy switching between "standard" and "enhanced" models,
very fast and smoth high quality outputs up to aprox. 1440 x 960 with 60fps,
2x fast upscaler,
4x fast framerate multiplier,
MMAudio Sampler (generates sound accordingly to the video action),
Triton and Sage Attention option,
A 5 Second long high quality video generation takes about 10 - 15 minutes (see below).
Tested generation times:
As a rough guide value for RTX 3060 GPU: generating a 5 second long high quality 1440 x 960 60 fps video with 6 steps it will take:
t2v: around 10 - 12 minutes,
i2v: around 15 minutes.
Comfyui-Easy-Install with Triton + SageAttention:
This workflow should work with any latest comfyui version >v0.6.0 (Desktop, Embedded, Windows/Linux).
However, comfyui is developing rapidly, and it often happens that some of the custom nodes used are not updated quickly enough or not updated at all. Manual workarounds are sometimes necessary. Furthermore, care must be taken to ensure that there are no conflicts with other nodes.
If you're having difficulties with your existing comfyui system or if you want to run video generation on a separate (parallel) comfyui system, like I do, I would recommend you the following installer: https://github.com/Tavris1/ComfyUI-Easy-Install.
Complete installation of comfyui including manager and some pre configured custom nodes is just one click - really 🙂
Installation of Triton + SageAttention is just a second click - really 🙂 And since it's so easy now, I would definitely recommend it to you for video generation.
Cause it is an embedded version, you can install it parallel to your existing comfyui version without the risk to ruin your working system.
After installation just configure the "extra_model_paths.yaml" file to use your existing models.
After a fresh installation of Comfyui-Easy-Install you might have some issues too, but there are known workarounds - please see the FAQ below.
For testing/understanding/experimenting/changing the workflow:
Click "Toggle Link Visibility" to see the links.
click the Subgraph symbols to open the Subgraphs.
for quick testing you may lower the settings for: steps, clip lenght and video resolution,
be really carefull with modifying Groups or Subgroups (even Titel or Color) cause they are essential for switching,
feel free to try and test other models. Just give me a hint if you find models which deliver better results and fitting the 12 GB VRAM limit.
And as usual: Have Fun 🙂🙂
Short Conclusion:
This workflow is based on elements of a variety of allready published workflows. My "job" was only to put things together, optimize it for a small machine and create a most simple and hopfully user or even "beginner" friendly workflow.
I`m not an "expert" - just a user who wants to get it running on "available" hardware.
There are many things I don't really understand. If you find mistakes or better solutions please give me a hint.
And I really hope that even "beginners" have a chance to go the first steps...
Frequently Asked Questions (FAQ):
For quick and better overview I will try to merge all known issues here - step by step (please be patiant). If your issue is not listed here, please have a look in the comments first. Most issues have been allready discussed.
Comfyui Nodes 2.0:
Turn off Nodes 2.0 in comfyui (use comfyui menue). Actually not all custom nodes are supported.
Comfyui crashes after generation while vae decode, upscaling or frame rate multiplying (Rife VFI) without any error report:
This is a RAM problem (not VRAM). Increase your swap file (min. 64 to 128 GB) or set it to automatic management on a fast drive with at least 100 GB free space.
JW Nodes (JWFloatToInteger, JWIntergerDiv, JWImageResizeByLongerSide), soundfile missing:
For the workaround look here and here:
python -m pip install soundfileFresh Comfyui-Easy_Install Installation (missing soundfile and Pytorch v2.9.0 issue with SageAttention on Windows:
For full conversation look here.
Open cmd in python_embedded folder:
python -m pip install soundfile python -m pip uninstall -y torch torchvision torchaudiopython -m pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu126Slider Nodes - how can I modify the "default" values:
Right click the slider node, choose Properties and set the values you like 🙂🙃
Description
Attention:
This workflow works only with Triton and Sage Attention successfully installed! Please see my description.
Otherwise please use my previous workflow v1.1.
This workflow is exactly the same as v1.1, except for the additional Triton and Sage Attention nodes and the switches to turn it on or off.
FAQ
Comments (205)
Is it possible to add negative prompt into this workflow?
It's allready included but "hidden" 😉 To see the negative prompt node just move the covering nodes away. But as far as I know, the negative prompt has no effect on wan.
This worked perfect for me
but I followed this guide to get sage attn and Triton installed into comfy portable
https://youtu.be/-S39owjSsMo?si=dnjeHcmf-oaz9ms1&t=103
Thank you 🙂 Did you checked your link? It is just another link to the same youtube video as published here.
@arkinson oh, I've been trying a few different posts in order to try and get sage/triton installed, lol. Must have mixed up a few and thought that youtube link came from somewhere else. Kudos!
@cuntryblumpkin581 Yes I know this, 10 different links open just to get one thing running 🙄😅
Thank you for this awesome workflow, really appreciate it.
Is there a reason why my videos are overexposing to some red color? not using any additional loras, just base models (h/l) and lightx speed (h/l) indicated in workflow, had to increase to 40 block swapping because of low vram issues...
ok, so if anyone else faces this issue, its because I changed the lightx speed loras for "newer" wan 2.2 ones, just stick to old reliable wan 2.1 ones, forgot I did that...
@Rainart1989 Hi thank you for your feedback 👍 I would also have suspected a "false" Lora first. You are riight, my workflow actually uses the wan 2.1 Lora. Have you experimented with the weights on the new wan 2.2 Loras yet? I hadn't even noticed that there were already out. I will have a look at this soon....
@Rainart1989 I just tested the wan 2.2 t2v Loras from here: lightx2v/Wan2.2-Lightning at main. Seems to work perfect with Lora weight 1.0.
@arkinson thats great news! Both High and low at 1.0 or high 1.0 and low 0.5? did u only tested T2V? what about I2V?
@Rainart1989 I did just a few short tests with the t2v models (both with weight 1.0) but I would assume the i2v models should work too. I have not tested different weights yet either.
But my first impression is that I don't see any difference in quality compared to the old wan 2.1 Lora so far. If you have the capabilities to do more serious tests please let me know about your results.
Holy shit, this is amazing! I was getting destroyed by that 14b thing and started losing hope, but then I tried my best with my buddy ChatGPT helping me troubleshoot and I gotta say, the work you did is immaculate. I never thought I’d be able to generate these videos in just 5–10 minutes
By the way, the videos you’re making are you really getting those results just with the prompt ‘beautiful woman,’ or do you add something extra?
Thank you so much for your feedback. I`m glad you got it running 🙂
No - this prompt is just a "fake" of course. With real prompts you would have a lot more trouble with getting blocked. If you are looking for inspirations/real promts please have a look at my more sfw content. Most of my "sfw" images/videos I published with real prompts.
I am loving this. Perfectly working. Thanks for sharing this workflow and documentation.
A gem, thank you. 4060 8GB laptop here and I can get up to 7-secs, but with a lower starting res image.
@craigsliststuffs19698 Hi - thank you for sharing your experiances 👍
why is high pass being used first and then low pass? isn't low supposed to run first?
I don't know the background, but according to the official workflow high runs first then low.
excellent workflow!
I do sometimes get deformations/jaggedness of the faces/hair, do you know what causes this?
No, not really, But you can try to lower the Lora weights. Or you can disable the Loras for testing - but keep in mind, you have to adjust the KSamplers in this case.
I have a 12GB 3060 GPU, and I'm very pleased with this workflow. The experience is fantastic, more than twice as fast as before. However, there's a small issue. After selecting "Image to Video," the input image only appears briefly in the first frame; the rest is ITV video. I noticed two preview images appear in the K-sampler: one for my input image and one for what should be the ITV video. I'm wondering if anyone else has experienced this.
Hi - thank you for sharing your experiance 🙂
Unfortunately, I don't understand your issue. What do you mean with: "the input image only appears briefly in the first frame; the rest is ITV video"? Have you checked that in "Image 2 Video" -> "Image or Batch" switch only one option is checked?
@arkinson Thanks for your reply. The first frame of the generated video is the image I input, which flashes by quickly. The rest of the video is normal ITV. I only turned on one of the image-to-video and text-to-video switches. As for batch switching, I even deleted the link.
@moshi Sorry - I did not get what you mean. OK let`s go step by step:
1. You use "Image 2 Video" (Text 2 Video is switched off)
2. You use "Load Image" ("Load Image Batch" is switched off)
3. You loaded or drag-and-dropped your start image to the "Load Image" node.
After starting the generation process the "Preview start frame" node should display your rezised start image. Both KSamplers I used have no preview funktion, so there should be no prevew images 🙄 After generation has finished you only get the "Save last frame" image and the two videos in the "Video combine" nodes. Just a wild guess: maybe a browser cash issue? Did you try to restart your comfyui/computer??
@moshi Sorry - I did not get what you mean. OK let`s go step by step:
1. You use "Image 2 Video" (Text 2 Video is switched off)
2. You use "Load Image" ("Load Image Batch" is switched off)
3. You loaded or drag-and-dropped your start image to the "Load Image" node.
After starting the generation process the "Preview start frame" node should display your rezised start image. Both KSamplers I used have no preview funktion, so there should be no prevew images 🙄 After generation has finished you only get the "Save last frame" image and the two videos in the "Video combine" nodes. Just a wild guess: maybe a browser cash issue? Did you try to restart your comfyui/computer??
@arkinson I tested this with a different workflow and got the same results. It's not a problem with your workflow. Thank you for your patience. I'll try to find the problem myself.
@moshi Thank you for your reply and good luck 🙂
@arkinson After more than 70 video generation tests, I found that after adjusting the CFG value of the K sampler to around 0.4, the video began to look closer to my input image and more coherent.
@moshi I´m really not sure if you use "my" workflow 1 : 1. Without any "tweaking" the very first frames of the gereated i2v video should similar/equal to your starting image.
I can get this to generate a 5 second video. Wonderfully.
I was expecting that I could T2V the first video ("a cute little grey cat is cuddling in her basket.") and then I2V a second video (seeded with the 1st's last frame - "the little cat roles over") and that would result in a 10 second video of the two stitched together.
I'm pretty sure my expectation is wrong. How would I combine these vids to get a longer one?
(I'm a bit of a newbie)
@schnee_actual Yes it is quite easy and you are on the right way. But keep in mind: you have to do all steps manually. Generate 2 or more videos (allways use the saved last frame image as the input image for the next i2v generation). Finally you have to merge the individual videos manually with some simple software (there are even some online tools).
I am getting this error message when trying the run the WF "KSamplerAdvanced
Given groups=1, weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 29, 90, 90] to have 36 channels, but got 32 channels instead"
I have a 3060 12gb, so i`m not sure what is causing this. Any help would be appreciated
Probably a wrong model was loaded. Did you use the right VAE model? It must be wan 2.1! Please check all downloaded models twice. Then go to each loader node and manually select the model from your local path just to ensure that comfyui really uses the right models.
And have a look here too: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?commentId=907658&dialog=commentThread
@arkinson I made sure to download each model you had stated in the notes, and have now tried making sure to select them manually from the dropdowns (noticed the workflow had the lighting loras selected from a sub-folder called video, so is now linking to the correct location), so all the models are correct, but the issue persists. Checked the link you posted and followed it through to the github but still can`t find a solution.
The only thing I have noticed that doesn't seem right is the image/batch selector node always showing batch as being On. When I select image, it turns off for a second, then comes straight back on, even though the batch node itself is then colored out
@Grim0r Your comfyui is up to date? You can google the error by yourself - most hints going in this direction: model mismatch/outdated models/outdated comfyui....
"When I select image, it turns off for a second, then comes straight back on" Here I would guess a general comfyui problem on your installation, maybe conflicting nodes etc...
@arkinson I am not the most technically minded person tbf, but i'm convinced that my Comfyui installation is not right. For instance when I run the update_comfyui.bat I just get a message saying "The system cannot find the path specified.". I think at this point i'm best off just trying to do a fresh install, since this seemed to fix it for other people. I`ll try the portable version or something and get back to you
@Grim0r Thank you for your reply. I would suggest you the portable version as a test invironment. Just have a look at my latest workflow desription ‘Installing Triton and Sage Attention,’ at the bottom of the page for some help and information (of course - you don`t need to install the Triton and Sage Attention part).
@arkinson Thanks for your prompt replies. I bet you didn't realise you'd have to become some technical support guru when simply posting a workflow :P
@Grim0r Definitely not 😅😆 👍 But on the other hand, it doesn't make you any less intelligent either 🙂
Have u found a solution of this problem?
@ThaiAI0 Not yet as i've not got round to doing a fresh install (which i`m sure will fix the problem), but ive seen comments saying that its most commonly your comfyui being out-of-date, or a conflict with some custom node, most common one seems to be flow2-wan-video
@Grim0r Yes, I have already reinstalled my ComfyIU and it works for me
@ThaiAI0 Thank you for confirming that. 👍
@ThaiAI0 ... and thank you for buzzing 😋
Sorry for not getting back sooner, but i`ve now done a fresh install of Comfyui portable, and can confirm, it all works. Thanks for the help guys! ;)
@Grim0r Thank you for your reply 👍And have fun! 🙂
Hello! Thank you for this workflow, I'm new to img2vid generation but still managed to make it work because of you. However, there's a problem with upscaling. It takes too long and sometimes causes out of memory error. I have checked the console and basically what is happening from my understanding, is that it upscales x4 instead of x2 for some reason. For example in video that was generated from 643x720 image the final resolution was 2560*2880. Is there any way to control it and force to do x2 max? I am using remacri foolhardy since it never alters the initial image in bad way.
Is it perhaps tied to RIFE VFI node and its x4 multiplier? If it crashes by "out of memory" it usually happens at that step
Hi - thank you so much for sharing your experiances. I`m glad you got this stuff running 🙂
"4x_foolhardy_Remacri is a specialized image upscaling model" that`s your "problem". If you use a 4x upscaler instead of a 2x upscaler the out of memory "issue" and longer upscale times are normal. You have two options: The simplest is: use the 2x upscaler from my workflow, or any other 2x upscaler. The other one is: you can try to use any 4x upscaler in combination with a resizer node - but you have to manage the out of memory problem (increasing blockswapping, reducing start resolution,....) and of course you will need aprox. the double upscale time.
@arkinson Thank you for answering. So basically the resizer node needs to be connected to upscale node? Or is there a different upscale node with multi option? I'm just too used to usual image generation stable diffusion stuff, where you can use any upscaler and just set the multiplier
@pepegamega Please keep it simple. Just use the 2x upscaler model of my worflow. This one I have tested for optimal speed and quality.
If you really want to test a 4x upscale model on a small machine, simply add a "Upscale Image by" node. But as mentioned, you have to test and "tweak" some parameters more.
I use the RealESRGAN_x2plus model, a 6-step, 6-second video rescales from 480x720 to 960x1440 in approximately 510 to 550 seconds. What node could I use to control a smaller size? I would like the output to be 720x1080. Thanks for this great workflow.
@FluffyIG As mentioned, use the "Upscale Image by" node and append it just after the upscaler node.
@arkinson OK thanks.
@metagenesis_studio Hi - thank you for buzzing 😋
@SnowShoes311 Thank you so much for buzzing 😋🙂 Yes indeed, I never thought that a simple workflow with a few explanations would obviously help so many people 🙄
The T2V part of this workflow is great but the I2V part gives me the error "MathExpression|pysssss Compex types (LATENT/IMAGE) need to reference their width/height, e.g. a.width" and I have no clue how to fix it. I thought it had something to do with the "Image Resize By Longer Side" node but switching that out for a basic image resize node doesn't change anything.
@stimuli_8864 Do you use version v1.0? Please use version v1.1/1.2 or check if the width/hight connections set properly:
Move the "Image Rezise by Longer Side" node to see the covered nodes. expand these nodes and check/set the width/height connections between these two nodes: "Get Image Size & Count" -> "Wan Image To Video". This should solve your issue.
update I got it to work.......but i don't understand why or how. For the 4th time a dropped a fresh v1.1 into comfy and (seemingly) did the same steps I listed below but now no error. I guess just assume I messed something up in the previous attempts but didn't recognize it. =/
Thank you for your reply. I am using v1.1 for now (because installing triton and SageAttention seems really complicated). Yes, the height/width are connected between "Get Image Size & Count" and "Wan Image To Video" along with image -> start image but the error continues. I even re-downloaded v1.1 and freshly dropped it into Comfy just to rule out if I had messed something up somewhere. The only thing I changed was disabling the T2V nodes and enabling the I2V nodes (and adding a start image and new prompt). Still get the same error. I really want this to work because the workflow otherwise is fantastic.
@stimuli_8864 Thank you for your reply. I`m glad it works now 🙂
The error message indeed is a little bit confusing, cause I do not use a "MathExpression"node in the workflow itself. Just a guess into the blue: You may have forgotten to load a start image at your first attempts?
thank you for sharing this, working great! :D
Hi - thank you. I´m glad it works well 🙂
I can just barely squeeze out a 720x720 vid with the non sage attention workflow if I use custom nodes to clear ram and vram before the RIFE VFI node. Without them my system freezes from excessive page file usage, using a 5060 ti 16gb but it will likely help even on a 3060 if your system hangs at that particular point
Hi - thank you for your hint. OK 720 x 720 is really heavy load. I have never tested this square resolution on a RTX 3060. Do you use the "Unload Model/Allmodels" nodes to free the ram/vram?
@arkinson I'm using the comfyui-memory_cleanup custom nodes, both the vram-cleanup and ram-cleanup ones, they are set to their most aggressive settings to unload models and clear the cache. Works fine with them, it's mostly a 32GB system memory issue rather than vram as it seems RIFE is not GPU accelerated
@FloatsYourStoat Ahh - I see - the "LAOGOU-666" nodes. I didn't know them yet but will have to try them out 🙂 Thank you so much for your explanations 👍
@arkinson Yep those are the ones! Hopefully someone else gets some use out of em too, they kind of saved my bacon.
I only get a fraction of a second of video for t2v. I see a clip length parameter, but only for i2v. Where can I set the length for t2v?
@yahoo456 Just use the "clip lenght (in seconds)" node in T2V 😉 If you can't see it, you may have accidentally moved or deleted it. Simplest solution: redownload the workflow.
This is my new go to workflow, thank you for your amazing work. I have two questions. The node groups you created. Some are behind other groups. Is there way to bring node groups forward to the front? Would there be any benefit adding TeaCache?
Hi - thank you 🙂
I`m not really shure what your question is about "node groups". There are many "hidden" nodes covered under the overlaying nodes. To see them, just move the covering nodes away.
Please don`t touch the groups, cause most of them are essential for switching.
I don't think that TeaCache will help much, cause we allready working with only 4 - 6 steps at all and I`m really not shure if it will work together with the lightx Lora. Perhaps you would like to test it 🙂
@arkinson Thank you. I understand what you are saying about Teacach and the number of steps. Thanks. Regarding the node groups. You have the workflow sectioned off into boxes. For example '02 Generation Settings' and within that section you have '03 Image 2 Video' and '03 Text to Video'. Those inner sections/boxes 03 are behind or underneath the main section/box '02 Generation Settings'. If I grab the title bar of the '02 Generation Settings' and move it around, the boxed behind don't move, they are on a different layer. Is there a way to move those inner sections on top?
@dreadfulpirate Ahh - I got what you mean 👍I am struggling with this problem myself. That seems to be a comfyui issue, just look here and: here. If I understand the fix correctly - obviously, this is not easy to resolve retrospectively without recreating all groups manually☹️I only have this poor workaround:
1. delete the overlaying group '02 Generation Settings'
2. now you can manipulate the underlying groups, like '03 Image 2 Video'
3 recreate the group '02 Generation Settings'
I know that`s really frustrating. Maybe there is someone with some more knowledge to fix this in a finished workflow....
simple, fast and just works, also nice to have the extra prompts
When I use I2v, the video seems to be looping, or at least trying to. Is there a way to stop that? It effectively makes the video shorter, since so much time is spent undoing what was done to return to the first frame.
Wich clip lenght did you set? You mean "real" looping or just the preview in the combine nodes?
@arkinson I'm doing a clip length of 7 seconds. I mean "real" looping, for example if I try to have a woman take off her shirt, she lifts it up and puts it back down instead. If the character starts out looking away, and I prompt her to turn to face the camera, she always turns back around by the end of the clip. I can't get the video to end at a point where things are significantly different from the starting image.
@capnreynolds999504 Ah - ok. I belive this is "normal". The same thing often happens to me. Sometimes you can enforce a better behaviour by prompting appropriately.
The only thing I could imagine would be to change the "shift value" in the ModelSamplingSD3 node (move the KSampler to see this node). But I have no experiances with this, I just used the values from the official workflow. You might have a look her: ModelSamplingSD3 Node Documentation (ComfyUI). Would be cool if you have the capacity to run some tests 🙂
Is anyone else getting a lot of 2d anime line art? I didn't even add my loras yet, it's just the template and it's giving me line art.
@Nightingale13 I would guess that's just a question of prompting.
I try the 1.2 but keep stuck on here
KSamplerAdvanced
mat1 and mat2 shapes cannot be multiplied (231x768 and 4096x5120)
I have no clue what cause it
I'm doing the I2V
@CECIL2000 "I try the 1.2" <- Triton + SageAttention is installed and working?
@arkinson I've checked and found that I have sageattention but without the triton installed. I guess I just turn off the sage feature without messing with the comfyUI. Thanks for the help !
@CECIL2000 Thank you for your feedback 👍Yes - you can switch it off in v1.2 or just use v1.1. On the other hand the error message is strange. Do you use any "exotic" aspect ratio for the input image?
@CECIL2000 I googled your error message:
1. Please check that your comfyui installation is up to date and you don`t get error messages while updating. 2. Check twice that you definitely run the right models. Look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=918029
How can I solve wanblockswap 'nonetype' object has no attribute 'clone'
@rdfssqa Your comfyui is up to date?
@arkinson yes,it's v0.3.55 now
@rdfssqa I would guess there is something wrong with your comfyui installation, maybe conflicting custom nodes? What happens if you bypass the block swap nodes?
@arkinson I still ended up running into something like ModelSamplingSD3: 'NoneType' object has no attribute 'clone'. Should I reinstall ComfyUI completely
@rdfssqa A fresh install could be the simplest solution if you can`t google the errors by yourself. In this case I would recommend you the portable version as a test environment. You may have a look here too: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=918029
@arkinson same issue here. did a fresh reinstall of CUI and the error persists. not finding much on google. I'll keep digging and let you know if I find anyting
if you have any more insight lmk
@CyberCurves Unfortunately @rdfssqa did not respond anymore, but anyway.
OK, you are on a fresh install. The error 'nonetype' object has no attribute simply means, the specific node gets no input. A common user error is, not to select the right models in the model loader nodes. OK, to make it simple. please use my latest workflow version v2.1 and run a simple test with T2V only (disable Triton + Sage Attention and all other options) and make sure you have installed and selected the right models. This should run out of the box.
@arkinson works like a charm. thank you for replying here. I basically uninstalled all models, re-installed all models, and double checked everything was in the right directories. no issues any more! I think I was getting lost in the sauce with a model mismatch or a spaghetti pipe getting lost somewhere. thank you
@CyberCurves Thank you for your reply. I`m glad you got it running 🙂
@arkinson sure thing, it's a great workflow. I can generate a 6 second video on my rtx 4070 (12 gb vram) in about 6 minutes. I'll post some of my stuff to the page here soon
Love the Workflow!
4070 12GB here and works as advertised.
For those of you who are using the default settings, I found that they produced images to "bright" or washed out for my liking.
Settings that toned down brightness/baked look:
Lightning High str: 0.7
Lightning Low str: 1.0
High CFG: 2.5
Low CFG: 1.0
Sampler (Both): Euler Simple
Of course, your mileage may vary, so feel free to play, but this is what got it going for my asthetic.
@Valomar Hi - thank you so much for your feedback 🙂Looks interesting to play with the settings. I would never have thought to reduce the Lightning High so much and increase the CFG. Would be cool if you could publish some of your videos here.
@arkinson Definitely will. Yours is the only workflow that has been able to give me any reliable results. Right now, nothing is ready for me to share, but once I get it to my liking, you can count on me tagging your workflow!
@Valomar Thank you 👍 Yes I know - tested for myself a lot of barely documented workflows, just to see that they don`t work or never will run on my machine. Unfortunately, there is a lot of confusion about the different models, Lightning Loras, Sage Attention etc 😕 I never imagined that this workflow would be downloaded more then 3K times 🙄
I´m really keen to see your "experiments". Just leave a short comment on it on my adress - cause sometimes I´m loosing the overview here 😅
@xeper Hi - thank you so much for buzzing 😋🙂
@00abad1dea368 Thank you for buzzing 🙂
Great workflow, though my PC (3060 32gb) isn't so great. It keeps crapping itself at I2V with only 520p and three seconds :p
It strangely works fine sometimes (around 200s), but usually it uses all my random access memory and takes longer to finish (around 400s or more).
@xeper "my PC (3060 32gb)" - just to be sure: your RTX 3060 has 12 GB VRAM? or only 8 GB?
@arkinson 12gb :]
@xeper Ok - did you tried to set BlocSwap at 40?, Any additional Loras in use? Any other "heavy" background tasks? And please check twice you definately use the right models, see here: https://civitai.com/models/1852904?commentId=918029&dialog=commentThread T2V runs well?
@arkinson The Q4 works much better than the Q5 I had before, so I think the issue's solved. I also had a weird 2.1 vase that was too big. Thanks :D
@xeper You used your "own" models and asked me why my workflow not works - oh boy, that`s crazy 🙄🤣
If anyone else has this problem, try adding --cache-none to your command arguments or downgrade Pytorch to 2.7.1. Don't be fooled by my false hope, using a even a Q3 or the other solutions doesn't fix the problem. People on GitHub are saying Comfy's memory leaking, so hopefully it's a temporary solution.
@cmp1000426 Hi - thank you so much for your feedback and buzzing 😋🙂
UPD2: Problem solved!
Interesting. I have 48G RAM - workflow riuning at frame interpolation stage. ~40G+ of RAM used before this step and I see OutOfMemory error...
UPD: Well, I see where problem - lowered start size of image in I2V to 512 - 48G RAM is enough. After generating initial video too low free RAM. What about unloading WAN models from RAM before upscaling and frame interpolation? How I can make it?
Hi - sorry I didn`t got what you mean. You are talking about RAM and OOM errors? If you have 12 GB VRAM and 32 GB RAM as specified, my workflow should run with start resolutions of 720 x 480 without errors. If you get OOM errors that is primarily a VRAM and not a RAM problem. Please explain exactly what you are doing.
@arkinson Oh! Found ny problem! It was in upscaler from another workflow - x4 instead x2. So - RAM not enough. With correct upscaler - 720x720 6s video works fine!
@Inquis_Eisenhorn Thank you for your reply 👍 and I`m glad you got it running 🙂
I have a 5070 Ti with 16 GB VRAM and was struggling with other workflows on this site.
This workflow works perfectly for me. The results are very high quality, it is the fastest workflow I've tried so far and I've never had it crash on me in 100+ creations. Running it without Sage Attenion btw.
The only thing I'm missing is Florence2 integration for better automated prompts.
Thanks for sharing!
Hi - thank you so much for your feedback 👍 I believe with your GPU there should be plenty of room for improvements 🙂
I have nearly no experiances with Florence2, but I´m afraid it will use some VRAM too and could "kill" the process on small machines (yes I know there are a lot of ways to tweak it). But as I see from the comments, a lot of users are struggling hard to get just "this" workflow running 🙄
Implementing Florence2 should be no problem and sounds interesting. Maybe I will do some tests with my image generation workflows soon.... If you have the opportunity to test it with wan I would be glad If you could publish your experiances.
When i hit a run i get this error KSamplerAdvanced
'NoneType' object is not subscriptable. Any tips for the solution?
Hi - please use google. The error generally means there is a missing input "somewhere". Comfyui should highlight/output the relevant node.
And please check twice you have selected the right models: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=918029
@arkinson yes the image input is missing from i2v image resize and i don't know where have to connect. I sent you pics for the error on private message in here
@markkuczko272 Sorry, I don`t use the direct message system. Just redownload my workflow. Version 1.1 works properly.
@arkinson No problem the problem is solved brother :)
I have question I generated an 5sec i2v video 720 and i have a RTX 7700XT 12gb 32gb ram and my generating it took me 30 minutes to generate the video, why is it so slow?
Comfyui is optimized for Cuda graphics cards(Nvidia) Is it utilizing your gpu or cpu?
@SnowShoes311 It uses a GPU for generation and when I compared which GPU is more powerful, the 3060 or the 7700, the 7700XT was written as much more powerful.
@SnowShoes311 But it generates beautiful and quality results too slowly anyway
@markkuczko272 Hi - I don`t know your gpu.
Short question: you generate i2v 720 x 480?? Keep in mind, if your start image is in square format you will generate automatically 720 x 720 without reducing the resolution this could lead to OOM errors and long generation times. Did you get OOM VRAM errors? To compare please do a simple t2v test with 720 x 480 and 5 seconds without additional Loras. This should run in about 10 - 12 minutes.
First, the creator of this Workflow is very responsive and helpful. Thanks for that. He helped me clear up an issue.
Second, this workflow works like a freaking champ. I did switch out the Q4 gguf files for Q8, for better results. But this workflow is excellent. I've only got 16GB vram, so it was a breath of fresh air to find something that works so well!
Good job, and thanks for the helping hand.
@hdean Hi - seems there is something wrong with your comfyui. Please look here: https://civitai.com/models/1852904/wan-22-workflow-optimized-for-rtx-3060-12-gb-vram-gpu?dialog=commentThread&commentId=918029
I will check it on another version of Comfy. I recently updated. However, with only one exception - 2.2 gives me that error every time. I will get on that this weekend and send you a report with my findings. Thanks for the response.
@hdean Thank you for your feedback 👍
"2.2 gives me that error every time" Do you mean Wan 2.2 - that`s my workflow made for 🙄 or do you mean my workflow version 1.2?
@arkinson This is the wan 2.2 workflow.
You were on the money. I grabbed a fresh Comfy and updated it. Then ran the workflow. I have a nice 7 second video now. Took 15 minutes, but that is not a complaint. Thanks for the assist. Hey, you suggest loading 2.1 LoRas would work. Do they get loaded in high and low or just one of the lora loaders?
@hdean I`m glad you got it running 🙂
I`m often use wan 2.1 Loras. It works very well. You should use them in both - High and Low Pass, but you have to experiment with the weights. A a good starting point is: High Pass weight around 2.0 - 3.0 and Low Pass aprox. half size: 1.0 - 1.5.
So, I am able to make some pretty good vids. Oddly, the lora loader is having issues with the lightx lora. Lora key is never loaded - it seemed to be loading it before, or I failed to notice. Very weird. I did, of course, have to tell it where the lora is on my system, but it's decided it doesn't want to load it.
Either way, it is working.
@hdean Using wan 2.1 Lora produces many errors in the logs. This is "normal". Don`t worry about it. It works anyway. The "trick" is just to use the higher Lora weights.
@arkinson Very cool. And thanks much for being so responsive.
Got a quick question for you. THe flow works really well. But I would like to get a more detailed video. What suggestions do you have?
@hdean Right prompting 🙄😉 Loras of your choice🧐 or just use/create a detailed start image for i2v 🙂
Yes, seriously now, you may have a look for "detailer" Loras. I actually have no experiences with it, but there should be some on the market already.
@arkinson Thanks. I have the detialer loras and I do not know why I didn't think of using them. Some days the brain just doesn't engage!
@hdean Yeah, yeah - we are all getting older day by day 🤣 Anyway, good luck with testing 🙂
@arkinson So you know, I updated my comment on this flow. It works great. I am using the q8 ggufs and have no issues unless I go over 5 seconds. But that works beautifully. Great workflow. Thanks for releasing it and thanks for being so responsive and helpful.
@hdean Thank you so much👍 You use the Q8 models with a RTX3060??? This would be very interesting. I had never tested it. What`s about the generation time? Is there a significant difference in quality?
I`m just "working" on an extended workflow for integration of first/last frame to video. If you like you may have a look here: https://civitai.com/articles/19794/test-workflow
Not the rtx 3036 - I am on a 5060ti. Got 64GB ram. Takes about 10 minutes with the Q8. Quality is solid - not photographs, but very good. And, man, I would be very interested in the FLF for this. Sadly, no time to test as I have work. But, I will be checking that WF.
@hdean Ok - a 5060ti with 16 GB VRAM? Uses the Q8 model really so much RAM? 🙄Seems I have to try it for myself shortly 🙂 Btw. I have allready published the new workflow v2.0 with FLF generation here on the model page.
@arkinson I have it on 30 for block swap. I think I tried at 25 for the high and it worked. As I recall, it failed at 7 second vids, but has no issues with 5 second vids. It generally puts out a good vid at about 10 minutes.
I am making good use of the Speed LoRa, too. Otherwise, some movements are kind of slow. Secret Sauce works well, too. Gotta keep the strength low sometimes or it sort of takes over. The only thing it isn't doing well is Face LoRas. But, I haven't experimented with those properly.
I am currently in the midst of a rather long movie (I have a DA account) and will probably get to your FLF model shortly. I have some specific intent in mind for that bad boy. Though, come to think of it, my current movie might actually have a use for it. I might be looking into it sooner than i thought.
I will get some vids uploaded, too. That way folks can see the good work your WF can do.
Okay. You gotta try the lightx2v for wan 2.2...but only use it in the low not the high.
@hdean Hi - sorry I´m a liltle bit out, cause we tested a lot different things the last days. Finally I decided to publish the old Wan 2.1 Lora in my last workflow version v2.1 cause I belive it will suit the most use cases and is very simple to use: just one single Lora for hig/low/T2V/I2V.
Wich wan2.2 Lora do you mean - the "official" or the new one linked by boinobin730? What are your experiences to use it in the Low pass only? Wich weight do you use? Did you made some tests in comparison to the Wan 2.1 Lora? Do you get better results in generally? Yes I know, lot of questions, but I wouldn`t like to start new test challenges without some specific information 😉😅🙂
[edit: I had done some tests with the Q8 model too. Yes it works, but visually I couln`t see much differences. Maybe the max start resolution of 720 x 480 I can run on the RTX 3060 is too low for a seriouse comparison.]
I am away from my comoputer for the moment. But I will answer those questions when I get the chance and can be right in front of my system.
Quick question: Can negative prompts be installed on this WF or will it screw shit up?
@hdean Keep calm boy😅🙂Last working version is published, everything has time now. I just wanted to see if I have missed something.
I read somewhere that the negativ prompt has no influence for wan 2.2 video generation. But I got some requests for "having" it. So I made the negative prompt visible in version v2.1 again. For myself I never used it.
@arkinson You sounded like Foghorn Leghorn. I say, I say, boy! I read that the negative prompts don't work, too. Figured you would know better than I.
The Lightx version I am using is the pre-official release. Don't remember where I found it. But I am only using the low. It seems to give it a little more oomf, so to say.
@hdean Foghorn Leghorn - oh my: "You are so damn right boy!" 😂🤣🙂motivated me to look some episodes with a lot of fun 🙂 Arg - I don`t know anything better - just hear some rumors here and there 😅I believe with this itsy bitsy AI even the "experts" know nothing and every one runs on try and error. As you can see in the other comment we did a lot of tests with some simple combinations of Loras and KSampler settings and the final "scientific" result is: "taste matters" 🙄 Anyway, I will try your idea to run the lighning Lora only in the low pass soon. On the other hand I did allready some 3 stage tests (first path without Lora, second pass high and third pass low), but the results were mostly unfavourable....
@arkinson I've done some side by side comparisons and you are correct. the 2.2 version gives some sloppy looking results compared to the original.
This workflow was working great. But I think after an update I'm getting this message.
Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = True
0%| | 0/3 [00:04<?, ?it/s]
Restoring initial comfy attention
Patching torch settings: torch.backends.cuda.matmul.allow_fp16_accumulation = False
!!! Exception during processing !!! BaseLoaderKJ._patch_modules.<locals>.attention_sage() got an unexpected keyword argument 'transformer_options'
Traceback (most recent call last):
File "H:\Stable_Diffusion-STABILITY\Packages\ComfyUI\execution.py", line 496, in execute
Does anyone have any ideas?
Thanks.
@boinobin730 "....attention_sage() got an unexpected keyword argument ..." Do you use my version 1.2 for Triton + Sage Attention?
@arkinson Thanks for the reply. It seems to be a sage attention issue more than anything else. Your WAN 1.2 workflow without sageattention works well. The WAN 1.2 with sage attention used to work so well. So I will spend some time trying to fix sageattention. Since I am using stability matrix, it makes troubleshooting a little more complicated. Installation of sage attention previously was very easy with stability matrix. I will report back if I get a breakthrough. Without sage attention isn't too bad with 5s videos anyway.
Edit:-
I found something. it seems that a lot of people are getting the same issue now.
https://github.com/comfyanonymous/ComfyUI/issues/9841
hopefully there will be a fix soon. in the mean time I will use the normal workflow without it. I don't really want to try to revert back yet, I'm hoping there will be an update in the next few days that will fix it.
@boinobin730 Thank you for your feedback. My first guess was a Sage Attention problem too. Ok, for myself I haven`t updated my portable version jet "fortunately" 😉 That's why I can't say anything about this.
stability matrix: What are your experiances? Is there any advantage to use it with comfyui only? I just had a short look at it. If I got it right it is possible to run/install different systems like Automatik1111 in this "invironment" and it seems to be portable. Would be nice if you could give me a short "overview".
@arkinson Yes. Pretty much. It simplifies the installation process of a lot of AI Python Packages. Basically making the difficult installation process much easier for novice Python users such as myself. The drawbacks is that it means you need to wait for changes and updates to be able to get new techniques such as Sage attention. When Sage attention first arrived I had to wait a considerable time for it to be available in Stability Matrix. I tried to install it manually but Stability Matrix would not allow it in the Stability's version of ComfyUi. I probably could have tried a portable version of ComfyUi and then Sage attention but I'm kinda lazy and didn't want to put myself through the effort. Once it was available in an update, It was really just 1 click of an option and Stability Matrix updated all the files necessary for it to run properly. It ran without a fuss immediately. Your Sage Attention workflow was so good. It's nice to look at and it works.
Ohh and another very good reason why I like Stability Matrix is because it centralises checkpoints and loras. So that their version of FORGE will point to the same LORA folders as ComfyUi. This made it easier when I was playing around with SD Forge and then playing with ComfyUI. I am starting to prefer Comfy now but I have a lot to learn and I'm still a newb.
@boinobin730 Thank you so much for your explanations. I got it now 🙂 Yes indeed, sometimes it is a little bit tricky to get the manual installations runing and mantained. I use a "central" model path for all systems but of course you have allways to manually edit the additional path yaml files.... An automatic installation option for the Sage Attention part sounds very interesting. It seems I should have a closer look at it 🙄
Is Forge usable inside of stability matrix? Are there any other applications usable except Stable Diffusion? Have you installed different systems inside stability matrix - like Automatic1111 and comfyui?
comfyui: In my opinion the most advantage is, you will get a much deeper understanding about the whole generation process step by step and you can use it finally to automate exactly what you want. But of course, sometimes it is not so easy to get and keep the overwiev....
@arkinson Stability Matrix allows me to install Onetrainer, KohyaSS and AI Toolkit. These packages allow me to train loras. I've used Onetrainer and Kohya. I prefer Onetrainer but I want to try AI toolkit as it seems to be developing a lot faster than Onetrainer and Kohya. I might have a go at making a WAN lora. But I'm not sure how long it will take on my 3060. ComfyUI and SD-Forge are my usual go to for image generation, yes they both run under stability matrix. I used to have old Automatic-1111 just by itself but that was last year. I since deleted it. As I get my head around ComfyUi I am starting to see the logic and the benefits of using it.
Also take a look at Pinokio if you are interested in other AI applications. Pinokio is similar to Stability Matrix, as it allows for easy install of newly developed tools. Such as AI voice one shot text to speech. There are so many great applications being developed in the AI sphere. The development is so rapid its hard to keep up.
@boinobin730 Thank you again for your really usefull informations 👍I just have installed Pinokio and I believe I have got the main concept now. Looks really interesting! And I am curious to see how it works in practice. The idea to have different comfyui instances for example (just one for a specific task) sounds good. I believe I need much more time to test all this stuff 🙄🙂
Btw: I use Onetrainer too for Pony training. For Flux Loras I changed to comfyui. This works pretty well. Training wan would be nice, but I believe without 24 or 48 GB VRAM would be no chance 🙄
@arkinson All good. I think you are probably right. 12Gb VRAM just doesn't cut it for LORA training on WAN. I'm saving up for 5090. Very expensive still. Anyway. Happy Generating.
@boinobin730 I have done some tests with pinokio and the results are really sobering: 2 of 4 applications stucks during installation with errors. The so called "low vram Hunyan" video application needs hours/days to generate a 5s video instead of minutes, and so on. But in my opinion the main problem is: with errors you can`t get any help because there is neither any documentation nor a large community to ask.
I`m just have a look at stabillity matrix. Could you give me a hint: is it possible to "migrate" an existing "central" model path? You said installing of Triton and Sage Attention was only "one click". In the "Launch Options" in the package card I only find the option: "use sage attention". checking it finally leads to error messages like "no module named triton" 🙄
@arkinson . Pinokio I used for Voice only. In particular Vibevoice, I also found it limiting for image gens so I don't use it for that purpose. With respect to Stability Matrix. Click on the three 3 dots in the top right corner of the Comfy-UI box. Then go to the entry Package Commands then click on install Triton and SageAttention. It should install. BUT. it still doesnt work for me on the latest comfy-ui so I will probably start fiddling around with trying a work around.
@boinobin730 Argh - the Triton and Sage Attention option was well hidden. Thank you. And yes, it leads actually into errors when I click it.
The "Checkpoint Manager" and "Model Browser" menues of StabillityMatrix are very interesting. Actually I`m using the "Lora Sidebar" in comfyui as a "full" Lora manager. With some "tweaking" it works pretty well (you can use the Loras in "Lora Power Loader" node just by drag-and-drop them from the sidebar for example. But sometimes it is also very buggy. Another problem with every model manager is of course, you have to add the metadata for all your own models manually (preview images, base model, trigger words, etc.). But I will give it a try on StabillityMatrix. Btw. I moved all my existing models into the StabillityMatrix path system, wich worked well for integration. Ok, I have to adjust the additional path files for my existing stand alone systems.
Just another question: Coud you advise any useful documentation for StabillityMatrix? On Github are more open questions than answers and their discord link seems not to work....
Btw. if you are considering the idea to have a working comfyui with Triton and Sage Attention, to install the embeddet version is really not to hard. It took me about one or two hours to get it running with all custom nodes. I use it as a system just for video generation.
@arkinson Good job on the install. I did have a separate Comfy install but it got wiped when I reinstalled windows. I think I will try again. Just too busy. There is so much I want to try. My main aim is to make a game (hentai) for fun but real life is always making me busy. lol.
I unfortunately can't help you on the documentation. It is pretty sparse, so I guess a lot of the time people are waiting on the Stability Devs to fix things. Manually doing it doesn't work as it conflicts with stability.
@boinobin730 Yeah, I know - real live is allways a problem 😆🤣🙂 and the Dev`s have it too - creating fu...ing good parts of software and never have the time to write a documentation 😂 Reinstalling windows - oh my. My last attempt took me month to get just 70% working again 🙄
I tried to "trick" Stabillity by selecting older comfyui versions, but the Triton and Sage Attention part didn`t run 🙁It seems to be just some "simple" version mismatch of python parts in the installer.... but as you said, obviously no change to fix it.
I tried to install cogstudio for some testing, but even here the "stabillity" installer stucks....
@arkinson I ended up reverting back to an older version of Comfy in Stability matrix and it works ok. Do you use a Wan 2.2 workflow with first frame and last frame node? I 'm trying to work out how to adapt your v1.3 workflow with a first and last frame node. Having a last frame as a reference will keep facial features intact. Especially the eyes. The eyes change significantly if the character closes their eyes in 1 last frame and a new generation has a totally different eye colour. Thanks.
@boinobin730 You got Sage Attention working in Stabillity Matrix by using an older comfyui version??? I tried different older comfyui versions too and even all older python versions, but no luck. Ok - I never had a working version running - my installations are all "new" attempts. Anyway, could you tell me what you have done?
First to last frame generation: I know, this concept is interessting. I had done some tests with Kling AI early. Unfortunately I haven't dealt with that in comfyui yet. But as I see, you have to use a different model. I would recommend to start with a new workflow just for this purpose. In this case you can easy simplify my workflow by deleting all the any nodes and groups. They are only used for switching between the different options. I believe for running it on a RTX 3060 the main question would be if Block swap will work with the specific model. Actually I`m testing a lot with SM and other "trifles" - but maybe, I will give it a try too...
@arkinson yes. 3 dots again. go to change version. flick the switch in the middle from branches to releases. i chose 3.58 and it works. I haven't tried to update again in comfy, it might just revert back to 3.59. i will have a play with the workflow. Steep learning curve for me. there is a very basic workflow
https://www.reddit.com/r/StableDiffusion/comments/1mm5l8s/first_last_frame_with_wan_22_video
but your workflow is the best, so I want to get the basic workflow up to your level that has almost everything. Whilst running on 12gb vram. I will have a play in the next few days.
@boinobin730 Thank you. As mentioned, I allready tested different older versions, even new installs with older python versions + older comfyui versions. I thought you may have a "magic" version number 🙄But as soon as I run the Triton + Sage Attention installer I get any kind of error messages 😬In your case you seems to have luck, cause you had it allready installed. Perhaps I will open some github issues, cause the developers seems to be active at her patreon channel at least....
My workflow is a litle bit more complex cause of the different options only. As mentioned: delete the whole groups/subgroups structure, delete all "any" nodes and delete the t2v part. For quick testing also delete the SageAttention part and all unnecessary stuff like, frame rate calculation, upscaling, framerate multiplying etc. Just keep the simple i2v part and arrange it to your likeness. Changing the models and adopting the new end-frame structure should be easy now 🙂
@boinobin730 Ok - I just created a first "quick and dirty" workflow version: https://civitai.com/models/1983241?modelVersionId=2244942 I hope you can open the link, cause I haven`t it published yet. It is based on the official workflow (see the comfyui templates). I used my optimisations so it works pretty well with RTX 3060. Actually I have run only a few quick tests. Would be nice if you can do some tests too. If anything works as it should I would publish an "official" version in the next days.
@arkinson You are so quick!!! I get a page not found from the link. But thank you for looking into it for me. When the page gets fixed I will download and give it a try. Exciting days ahead..
@boinobin730 Ok - I have published it as an article now: https://civitai.com/articles/19794/test-workflow Just download the json attachment. Please download the two new Loras. You also can try the "official" models in the (bypassed) Diffusion Model Loader. Actually I`m testing with 8 generation steps - quality seems to be much better....
Hello, the description mentions a beginner-friendly v1.1 workflow, but I can only find the v1.2 sageattention file. Could someone please share a link to v1.1? Thanks!
Just choose the v1.1 version from the menu. You will find the version buttons directly over the cat image on the top of this page 😉
Is there a spot for negative prompting in this workflow? I didn't see one and am a super-noob at adjusting comfy stuff, but I wanted to try to cut down on the mouth movements. :)
It was a good start for me. Thank you💜
@MiraKoks Thank you so much and good luck 🙂
Thank you for your work! Text to Video works for me with Loras, but Image to Video I get the "to have 36 channels, but got 32 channels instead" error. Any ideas on this?
Sorry, I did an update and it's working now. I looked first before posting, I shouldve looked harder
@kevinflynn1967330 Hi - thank you for your feedback and good luck 🙂
Incredible workflow! Best out there!
With lightning, I get slow motion videos here and there. Have you considered adding a third high noise sampler to mediate the whole slow motion issue with lightning loras?
Hi - thank you so much 🙂 I`m really not sure if a third pass would solve the problem, not to mention a third more of processing time. In my opinion the "motion" is completely "generated" in the first pass and the second pass works similar like image to image, just to increase the quality of the frame images. I would suggest you just "straighter" prompting or testing other start images in i2v. In my experiances some start images working better than others.
@arkinson awesome! Thank you for your reply. Yea, that makes sense about the processing time. Yea, it was completely random for me when some would generate in slow motion vs normal. I'll keep playing around. Thanks again!
@hydragyrum2 Good luck 🙂
The 'third' stage is used first, without a speedup LoRA in an attempt to lay down good motion before the main high-noise stage with speedup is used. It's a simple concept that can work in theory, but in practice will be vulnerable to image subject, and the spread of iterations between the three stages.
@blobby99 Thank you for the right hint 👍 I knew nothing about that issue. Just found a workflow with the 3 stages briefly explained: WAN 2.2 I2V Faster Motions 6-step Lightx2v Workflow - v0.1 | Wan Video Workflows | Civitai. Seems to be not really stupid 🙄 Maybe I wiill give it a try.
keep getting distortion and blur towards the middle and end. any solutions? using the Q6 K for high/low noise
@dkain76 Can`t say anything to the Q6 models, but anyway: too high Lora weights? wired prompt? "bad" start image in i2v? wrong models in use???
@arkinson ive got it working now, i needed the lightx2v lora.
@dkain76 Thank you for your feedback 👍 Good luck 🙂
For some reason ComfyUI didn't create the folders where the models should be placed, so I don't know what paths I need to create manually. Can someone send them to me?
"ComfyUI didn't create the folders where the models should be placed" Hi - sorry I didn`t got what you really mean. In short: in comfyui allways you have the folder "models", inside this folder you have the subfolders like "checkpoints", "vae", "clip", "lora" etc. Please look at the description notes in my workflow. For every model you find the link and the path information.
@arkinson i fixed it yesterday, thank you
Is there any way for an update for the Sage- Attention workflow. Everything was working perfectly, but now for some reason the only node im not able to use is the "Load Image Batch" node.
No amount of update or whatever fixes this.
Or if anyone can suggest what I can use as a replacement
edit: just use v1.3 now 😉
@Megasherru Yes, you are right. This node seems not to be mantained any more. Please install the "mikey" pack via the manager and use the "Load Image Based On Number" node instead. This node works much faster on large directories furthermore. I will publish a update of the workflow soon.
@arkinson Thank you for such a quick update and getting it fixed :D
@Megasherru You are welcome 😉 No, seriously, I had already noticed that and struggled for myself with it. I just was to lazy to fix it ‘officially’.