This is the SVI 2.0 PRO version.
v3.5 12-Section 2nd-Fixed - Thanks to @jasonccc's suggestion, there was an incorrect connection within the LORA area, so I've replaced the file. My apologies.
v3.5 12-section fixed - An error was discovered where the prompt that should have been entered in "Section 8" was incorrectly duplicated with the 6th prompt. This issue has been fixed, so please reinstall the workflow.
Thank you so much to @MarcanOlsson for discovering the issue!
v3.5 12-Sec - This version is based on v3.5 and enables video merging through 12 generation steps.
When actually generating the video, you will notice color shifts compared to 6 generation steps. While 6 steps generally maintain better quality in practice, the ability to specify 12 steps offers advantages depending on the application, so we decided to add this version.
The 12-Section version enlarges the workflow, so unless you need more than 7 generation steps, I recommend using the standard v3.5.
v3.5 - The subgraph specification in the model input area has been deprecated and reverted to the v2 specification. Additionally, it is now possible to generate videos for only the first section.
We received multiple reports from the community that models such as CLIP and VAE were not functioning correctly due to the subgraph, and we also received feedback that the model placement was unclear. Therefore, we decided to revert to the v2 specification.
However, the subgraph of the generation section, which includes the sampler, remains unchanged. We believe that performing the generation process within the subgraph serves to prevent a decrease in generation quality. While the model area issue is simply a layout issue, the subgraph cannot be removed because it affects the quality of the generation section. If there are issues with the subgraph itself, please avoid using this workflow.
Regarding the video generation for only the first section, given the nature of SVI, we initially omitted it, believing that a single generation was unnecessary. However, we received feedback from the community requesting that a video be generated for each section, and that videos be added gradually while reviewing the generated videos. This was a very logical approach, so we added the "first video" and modified the workflow to allow videos to be accumulated while keeping the seed value fixed.
v3.4 - Layout adjustments.
v3.3 - Changed the seed node from "CR Seed" to "Seed (rgthree)". This change was made to align with commonly used custom nodes in this workflow, following reports of implementation issues with CR Seed.
v3.2 - Modified the layout to make it easier to disable Lightx2v Lora.
v3.1 - Modified the layout to make it easier to disable the Sage Attention node.
v3.0 released.
Video length can now be changed in each of the six generation sections, providing more flexible control over video content.
The frame rate (fps) was previously fixed at 16fps, but can now be changed arbitrarily. Accordingly, the RIFE-VFI node's scaling factor can now be changed in the input area.
GGUF model loader is now included as standard.
Version 2.0 changed the number of generation sections to six.
The layout has also been updated, allowing Seed node input to be processed in one place. Furthermore, the layout has been significantly redesigned to unify the user experience with Painter I2V versions, reducing the input burden. With this change, the wildcard prompt input method has been discontinued.
Please note that the explanations in this workflow are solely my personal opinions. I do not have expertise in AI generation, so some information may be inaccurate.
The main goal of this workflow is to achieve compact operation when performing repeated generation. It minimizes screen scrolling during operations such as prompt input, input image selection, specifying time, number of steps, resolution, and, most importantly, LORA selection. To further enhance compactness, all nodes are fixed to prevent accidental operation.
Links to the Models and LORAs and nodes used in this workflow
SVI LORA :
Wan Advanced I2V (Ultimate) :
https://github.com/wallen0322/ComfyUI-Wan22FMLF
This node was updated on January 27th, but the version available for installation from ComfyUI Manager may be an older version. While the older version will still work, you won't be able to set "SVI Motion Strength," and you'll likely experience more color misalignment. Therefore, if you can Git clone, we recommend installing the latest version.
Links to the Basic models of the Wan2.2
CLIP:
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/text_encoders
VAE:
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/vae
CLIP Vision :
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/clip_vision
You can generate up to six generations, and each generation is assigned a unique seed value. Normally, click "Randomize Every Time" to display "-1". Generation will be random. In this case, the seed value for each generation will be displayed at the bottom of the screen. If you want to fix the seed value, click the seed value field or enter the seed value directly. For example, you can fix the first and second generations and regenerate the third generation and beyond randomly. However, regenerating a section before the generation section you want to fix will change the final frame, so you cannot fix subsequent sections. As a general rule, regenerate after the section you want to fix.
By combining the six generated videos, you can create six different types of movement. For example, by generating and combining six videos of different durations, you can create a long video containing six complex movements. This is one of SVI's strengths, enabling complex processing that is impossible with a single generation.
However, SVI V2.0 PRO also has its drawbacks. Because SVI uses the first image as a reference point, the AI tries to restrict movements that deviate significantly from the reference point. As a result, the movement becomes sluggish and unnatural. Furthermore, this constraint imposed by the reference point also reduces the responsiveness to prompts.
In short, the use of excellent LORA is essential in SVI. In my experience, movements without LORA are very unnatural, lack impact, and resemble something out of a horror movie. Fortunately, there are many excellent adult-oriented motion LORAs available. However, if you want to create completely original movements, expect it to be difficult with the current version of SVI.
I hope this workflow helps make video production with SVI more enjoyable.
Description
v3.5 12-Section 2nd-Fixed - Thanks to @jasonccc's suggestion, there was an incorrect connection within the LORA area, so I've replaced the file. My apologies.
FAQ
Comments (56)
Awesome. I was struggling with another SVI workflow, slow, bad results, etc. but yours seem to solve all the issues ! One small improvement to allow landscape videos, change resize image to resize and plug the resize width and height into the generation width and height.
I don't generate them very often, but I think it supports landscape orientation as well, right?
@kenpechi It works if you route the nodes like I said before. Minor upgrade but it's usefull if you have a lot of lanscape images you want to animate.
Huh...
Well, do as you please.
@kenpechi also, small thing, since it's better to scaler on multiple of 16, you can set it up directly in the resize node with "divisible_by" value
I follow this workflow and try to create only two scenes but for some reason the second clip does not follow the second positive prompt, only extending the first. I tried to fix it but it didn't help. I want to know whether anyone facing same if second scene prompt is completely different from first scene
hello i can a newbie question, do i need to enable all 6 sections? i enabled 4 sections and it went throught all of the 4 and i ended up with only a 3 seconds video.
Enable all sections up to the number of times you want to perform the generation. If you want to perform the generation four times, enable all sections from "1st Section" to "4th Section". In this case, only "4th video" will be enabled for Video.
In your situation, the video settings may not be configured correctly.
Hello, when I use your workflow, the character's appearance always changes in the middle. Which node or model is causing this? In your video, the character consistency is very good. How is that achieved?
Whether using WAN 2.2 or LTX 2.3, character faces will fundamentally change. I'm currently working with LTX 2.3 and am directly addressing this face-changing issue.
For example, it changes depending on the resolution of the source image and the base model used. Naturally, repeating the generation process 12 times increases the risk of face distortion. The position, orientation, and expression of the face at the generation transitions are also important.
First, I recommend the fp8 model and LightX2V Lora that I use. My workflow can be used with other models, but it was optimized for the fp8 model. Additionally, keeping the video length per generation short, trying to keep the character looking at the camera as much as possible, and performing the generation process 6 times or less should help minimize face changes to some extent.
@kenpechi Thank you for your help. I have one more question. When using the official FP8 model, undressing scenes are very hard to generate. The girl usually only makes small, subtle movements and rarely actually removes her clothes. However, after switching to the Remix_NSFW FP8 model, the undressing performance became much better and more reliable. Should I stick with the official FP8 and keep refining the prompts, or would it be better to switch to the Remix model?
@billy19850317605 I think using Remix would be a good idea. Everyone has different preferences, so it's best to use a model that delivers the performance you like. Quality issues can be addressed through means other than the model itself.
@kenpechi Will you release a workflow for LTX 2.3
@billy19850317605 We plan to release it, but we are still learning how to do it, so please wait a while.
If anyone is experiencing this error on COMFYUI DESKTOP APP: CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.For debugging consider passing CUDA_LAUNCH_BLOCKING=1Compile with TORCH_USE_CUDA_DSA to enable device-side assertions
This fix worked for me: https://github.com/Comfy-Org/ComfyUI/issues/13661
@kenpechi thank you for your worklfows! I am testing both of them (SVI & Painter) and looking forward to use them but I've noticed a problem and want to ask did you have the same issues in past. In both workflows I am using Smooth Animations model, your Litghtx and smooth animations loras. And for tests generating 2 videos. In each videos I am using the same loras and the same seeds. And in result 1 video is fine but 2nd video speeds up and results are weird. Even when using random seeds. This is happening in your both workflows. Any idea?
I recall someone asking a similar question in this comment section before. (It seems older comments are not visible.) To be honest, I still don't understand why it happens.
@kenpechi hiya. It was Smooth Animations model fault. When I switched to Dasiwa checkpoint everything works nice and easy :)
@Inperfector I see, thank you for the information. Have fun!
@kenpechi oh by the way! What wan model do you use for your anims?
@Inperfector I don't create anime myself, but if I were, I'd probably use "Dasiwa". His models are very particular about facial consistency, so I think they're perfect for anime.
@kenpechi I meant for your videos what model do you use? :)
@Inperfector I use the official fp8 model. Ultimately, since I train all my techniques, such as painterI2V and SVI, on the official model, it best utilizes these techniques. While merged models might have advantages in some areas, they add impurities, preventing them from reaching their full potential.
For example, with Dasiwa, while the facial consistency is excellent, I don't think the movement is as good as the official model, so I don't use it.
@kenpechi I think the same. Dasiwa model is just excelent and so clear with quality. But the speed... I actually experiment myself with official model on your Painter workflow and I get interesting results :)
thank you for this workflow dude!
thank you for this workflow, I'd like to know if this workflow supports both a starting frame and an ending frame.
No, it's not supported. That's because I don't use the end frame.
@kenpechi Thank you for the clarification.
I have another question. I'm experiencing a gradual drift in colors and character features as the video progresses. For example, hair color, eye color, clothing colors, or facial details slowly change over time.
Could you explain what typically causes this issue? Is it related to the model itself, CFG settings, denoising strength, frame interpolation, or some other factor?
I'd appreciate any suggestions for reducing feature and color drift during longer generations.
@Nmst0409 This is a typical problem with video generation AI itself, and the contributing factors are varied.
First, SVI uses the last five frames of one generation as the starting image for the next generation, so the consistency is inherently shifted each time images are joined, leading to a degradation in image quality. When you join 12 generated videos together, as in the 12-section version, it's actually natural for the figures and colors to change, as you pointed out. According to community evaluations, my workflow is relatively better than others, but many people face this problem.
It also varies depending on the underlying diffusion model, resolution, and video length.
There's no fundamental solution, but to mitigate the consistency shifts, you can shorten the length of the video created in each generation (around 3 seconds to minimize quality degradation), reduce the number of videos joined together (I think the 6-section version is sufficient), minimize changes in the actions instructed to the AI, especially ensuring the character's face is as large and facing the camera as possible. Of course, the starting image should also be one where the face is as large and facing the camera as possible.
Incidentally, in my videos, the characters are probably looking at the camera to an unnatural degree. I instruct them to "look at the camera" in almost every prompt. This is to minimize changes in the characters' faces and maintain consistency.
As you can see, problems can arise from various factors, so you just have to keep trying, but I think it's best to be careful from the very beginning when choosing your starting reference images.
@kenpechi Got it, thanks for the explanation.
I'll experiment with a few different settings and see if I can improve the consistency.
Thanks again for your help!
很棒的工作流，能稳定出视频
Thanks for create such a GREAT workflow! It makes me much easier to create long video with WAN2.2.
By the way, I've got an error below when I use:
# ComfyUI Error Report ## Error Details - Node ID: 815:871 - Node Type: WanVideoNAG - Exception Type: RuntimeError - Exception Message: RuntimeError: self and mat2 must have the same dtype, but got Float8_e4m3fn and Half
Here is what I've tried:
- Re-install whole ComfyUI→it seems to be solved but happened again without change any settings.
- Update KJnodes
- Disable SagedAttn
- Run ComfyUI on forced fp16 mode
- Re-install clip encoder
Idk if it helps but my environment:
- RTX5080 VRAM 16GB with 64GB RAM
- ComfyUI ver.v0.24.1 on ComfyUI-Ezi-Desktop v3.7.1
Very appreciate if someone could teach me idea/information/solution, thanks! :)
EDIT:I always used WAN 2.2 Enhanced NSFW Fast Move Q8 High/Low.safetensors, but changed to gguf model then it works.
Also i tried Wan2.2-I2V-A14B fp8 model and it works too.
Tried re-installation of enhanced NSFW, but it did not work. Maybe just compatibility issue.
Thanks for this great workflow.
It is possible to add a colour match node to the start image and the other images from each sections?
You can add the node, but I don't think it will work very well. Originally, the community was trying to mitigate the severe color shift of SVI using color matching nodes. My workflow is highly regarded partly because the degree of color shift is small. Even so, it probably won't be completely eliminated.
Of course, it's worth trying, but I think it will be quite difficult.
Todavía debo seguir estudiando, porque no entiendo como con mi imagen generada previamente cargarla en este workflow... debo generar la imagen ya con la penetración antes de ponerla en este workflow?
No estoy seguro de si esto responde a tu pregunta, pero te daré mi respuesta. Este flujo de trabajo genera un video a partir de imágenes que ya han sido creadas, por lo que necesitas crear las imágenes por separado.
Thanks for this awesome workflow. But which computer requirements for run this workflow. I have ran on L40s 30GB RAM but not work
It break when caching after run 12 section
It's hard to say definitively as it depends on the final video length, but with my 16GB of VRAM and 64GB of noRAM, a total of about 30 seconds should be fine. However, I haven't actually made many long videos, so I'm not entirely sure.
@kenpechi thank you I will check it again
Hello. Your work is wonderful, and I'd like to try recreating it myself.
I know this is a bold request, but would it be possible to upload the workflow, including the starting image and prompts, somewhere?
Most of the works I've posted on this page probably have metadata embedded, so you can download the videos and drag and drop them into ComfyUI to see the workflow with prompts and LORA weights.
For the starting image, you can either prepare a similar image yourself, or I've posted some in the image section of my profile, so you can download and use those.
In other words, these have already been shared with you.
Thank you.
I did manage to load the workflow, but ran into many errors without the starting images, so the output didn't turn out well.
Your tip about the profile images is very helpful. I'll give it another shot. Appreciate your kindness!
one day i'm going to figure out how to run this :D
Newbie question, but can't seem to figure out why the output video is either a total freakshow hallucination or orange/yellow pixel fog. Any advice from more experienced users on what to look into to fix this type of issue?
There's a high probability of a model configuration error. The most common mistake is confusing HIGH and LOW. Other issues include installing a text encoder that isn't WAN-compatible, so please check that the models are configured correctly.
@kenpechi Thanks! Thankfully was not that newbie as HIGH & LOW were set correctly. But it seems the SVI LORA I got from https://github.com/vita-epfl/Stable-Video-Infinity/tree/svi_wan22?tab=readme-ov-file as per note in workflow was different although it was SVI Pro (Wan 2.2 14B), I took one from the direct link on this page and also changed the low lightx2 lora to be exact same as in original workflow and that solved the issue. Also adding to this one, cheers for an awesome workflow!
我也是出现一样的问题，请问楼主是怎么解决的
Thank you so much for these. These have become my primary Wan2.2 workflows. One issue I'm having is I'm unable to see TAESD previews. i can see them in other workflows, just not this one. There is on-screen text that says "Disconnected" under each section and i'm wondering if that is an indicator why previews aren't showing. Any insights are appreciated!
I am sorry, but I have absolutely no idea.
Love the workflow! Easy to use and works wonderfully well. I wanted to ask though, what's the best way to fight degradation? I don't expect to be fully rid of it but any recommended combination of loras or specific settings that might help greatly?
I've been using this workflow for a while now and it's been great, but lately I just update my comfyUI and I don't know why everytime I want to regenerate new section with different seeds, the workflow always generating again from the first section even though I don't change anything from the previous setting, it makes generating takes so long. Before this I have no issue, it generate fast because it doesn't regenerate the section that already generated. How to fix this problem?
I had the same problem. Solved it through comfyui discord. The solution is to add "--cache-ram 0" in Startup Arguments. Mine look like this now: "--enable-manager --cache-ram 0" and it doesn't restart the whole thing after every segment/stage.
@jerkoffalltrades how to do that? I'm not computer expert.
Radeon users, add "Clean VRAM" where necessary. It works well. (This also applies to most i2v workflows.)
I just want to create one single 5sec video thats all but when I try its blurry broken videos why I have 3090 anyone can help pls