WARNING: Hey, I've lately moved towards LTX2.3 but I will publish a fix to workflow issues mentioned in comments once I have enough free time.
I've officially switched to LTX due to slowmo issues with lightx2v loras. I might share a workflow but that model is already capable of 30s on 720p on a 4070ti so I'm not so motivated to do so. This one would still work if you found the missing packages. Thanks everyone for their support.
It finally happened!
Now there's a way for smoother continuous videos thanks to SVI team and Kijai.
We are at v1.0!
I've updated the workflow to add a few more features;
Video extend option by loading an initial video then converting it to latent that goes into first I2V(WIP)Option to switch between 3 and 2 ksampler phases by setting the initial step
Option to set cfg > 1 if you wanted to disable lightx2v
Images are saved partially in loseless format (use something like VLC to view them) and only loaded again on final merge, if something goes wrong you can merge those files to get a flowing video.
Implemented a bus system to reduce connections. Report if you have any issues but things should work as long as you have the right models and loras selected.
You can set and fix the seed for each part
There are options to upscale and interpolate before final save
Final save happens on main graph so you can preview your output
Slow motion issue probably persists. Couldnt find a consistent solution since when speed up using a third party tool every part becomes faster since they take previous latents as input until everything breaks.
Weak points of most SVI workflows right now is that it references first image in all parts so you might have background warping/chaning shape/textures on switches if the background has changed a lot.
I'll only be updating the workflow if kijai updates the node (there are two merge requests about end frame and better consistency(?)) and/or something breaks. So we can call this a semi final :)
Comfyui compatible SVI lora's;
LightX2V lora's I'm using;
FP32 vae:
Ultra Flux VAE for sharper !"Z Image"! outputs:
https://huggingface.co/Owen777/UltraFlux-v1/blob/main/vae/diffusion_pytorch_model.safetensors
GGUF still seems to be performing better than fp8 scaled in my experience.
Just share your outputs with us folks as well :)
v0.9
Left sampling on (1 + 3 + 3) steps with 4 parts (19s~). Takes around 10mins on my 4070ti with sage + torch compile. Feel free to extend it further if you need.
Everything is GGUF. Patch sage attention and torch compile are disabled by default but you are welcome to enable them back since they speed things up a lot if you have the environment set up.
You can set part specific or common lora's thanks to rgthree power lora node.
Happy generations! \('-')
Description
added video load option to extend an existing video(WIP)
added 2 ksampler flow for when the no-lightx2v-high step count is 0
added high-low-cfg to override high/low ksampler cfg if user bypassed lightx2v loras
fixed a latent issue where videos were 4 frames longer than expected due to a formula error.
added bus system to simplify connections
added per part seed control
added partial save system
added upscale and interpolate options
added final merge save system
FAQ
Comments (216)
this works great but my only question or i guess problem is the blurriness. has anyone found a fix for that? also i have face consistancy issues if the person moves there head. if they look back at the camera at the same initial angle its consistant. idk how to explain it other than when they look away the face is kindof similar but you can tell it changed alot more than it should have
Well it uses the same start latent + last latents for each part so I'm guessing it just keeps going back to the details it had. This is a consistency hack they implemented and its working better than I expected. We need further development or new models for a better consistency.
The video's I've shared have some extra realism lora on the image itself so they were a little blurred.
Also I cant see my workflow or images anywhere. Is civit having hiccups? Its like people are downloading it but I'm not sure where the traffic comes from. Can you view v1 examples?
@iLegoLoon When viewing 1.0 there are no examples only in the previous ones are there examples.
@iLegoLoon yeah idk what you did but i couldnt see it earlier. looks like its fixed now though
Awesome workflow! Really appreciate the effort and all the clean subgraphs. One suggestion. Can you make an option to remove the first sampler? I'm using SmoothMix that has the baked in light lora so I don't need the first pass sampler. Setting the high_no_light to 0 breaks the workflow since the second sampler doesn't have denoise enabled. I was able to work around it but would love the option to adjust right from the steps without digging into the each samplers subgraph to manually adjust.
Thanks! This beast ate through half my holiday :)
There are pull requests to kijai's node I'm using. If he ever merge's them into the node I'll update the workflow for end frame support and might add a boolean flag that automagically enables noise for second sampler if no lora steps is zero.
You could delete the ones except first two, edit them and duplicate the second one tho.
But until then I'ma try to rest and pass the cold that I feel like I'm catching.
Have fun.
I get the following:
WanImageToVideoSVIPro
not enough values to unpack (expected 5, got 4)
Seems the prev_samples input is empty?
Edit: Turned out to be a node conflict or something not updating correctly. Removed all custom node and installed just what is in the work flow and all issues resolved! Works excellent!
Yeah whats your video length, is it too low? Or is motion latent too high?
Its not the latent btw, its lacking one of the inputs which I'm assuming is your image, is it loaded properly?
Length was set at 81, all settings are as they are originally in workflow. The image is loaded correctly
It fails in the I2V-First WanImageToVideoSVIPro node
The node has 5 inputs
prev_samples does not have an input
the blocks following I2v-First have all 5 inputs
It may be link to a comfyui bug (not sure but I had the same meesage error). you can find a fix comfyui github discussion.
Guys,can you help me?
WanImageToVideoSVIPro module is missing and CUIManager not found this node in missing nodes, instead, it shows the KJNodes (nightly) like needs to be updated(
Installing Cuda13 did not solve problem(
Check the previous comments. People mention that deleting the kjnodes "folder" inside custom_nodes then reinstalling it fixes the issue.
@iLegoLoon it helps! Much thanks!
Even though i set WAN 2.2 fp8 versions inside the subgraphs, outside the subgraphs, it still refers to the gguf models and clips, i have also disabled those nodes; it still refers those. If I delete those, then it breaks apart. I think I need to learn how subgraphs work!
edit - Okay, edited the subgraph widgets and now everything works properly!
Can you share your experience with fp8?
Also I had a mistake with the formulas in encode subgraph. You can download fixed version or set the formula inside encode node to
"(a - 1) * 4 + 1" instead of "a * 4 - 1"
@iLegoLoon With the new formula, it works properly! With the old formula, it was not flowing properly in the subsequent parts!
Hi, in this new workflow, the video collapses and corrupts after part 7-9 or so but the last one i went to 4 minutes no problem (part 41)
unfortunately both workflows crash my comfyui when stitching together a video more than 1.5 minutes, but at least with this new one i can stitch together each part manually with the lossless parts.
I would love to go long af but cant with this new workflow because the movement screws up so bad after 30s.
I just found out that one of my formulas in workflow was wrong. Can you set lenght to 77 and try again? I think it is generating a 85 frame video instead of 81 due to putting an extra 4 frame and that could be breaking the videos. I will reupload the workflow when I have time.
Edit: reuploaded fixed version
Everything's fine, I managed to get it running on a 5060ti16vram with 32RAM. The only thing I couldn't figure out was how to make seamless videos; the stitches are still visible. That is, it doesn't continue the video at each segment, but creates a new one from the original image.
This is usually the result of not having svi lora properly set up. Also make sure all lora's are selected and use the same loras mentioned in v0.9.
Just saw the post, do you have gguf models and svi lora set up? Also I've updated the workflow for a length issue but this should not be related. Do you have any notes from your command prompt after the "got prompt" text?
Ayo lets fix it man, you are putting out a bad output example because you couldnt get it to work??
@iLegoLoon I think I used everything listed in the links, except for the models and text encoder. I have a Q6.
Thats selected:
Wan2.2-I2V-A14B-HighNoise-Q6_K.gguf
Wan2.2-I2V-A14B-LowNoise-Q6_K.gguf
SVI_Wan2.2-I2V-A14B_lora_HIGH_v2.0_rank_128_fp16.safetensors
SVI_Wan2.2-I2V-A14B_lora_LOW_v2.0_rank_128_fp16.safetensors
Wan_2_2_I2V_A14B_HIGH_lightx2v_MoE_distill_lora_rank_64_bf16.safetensors
Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16.safetensors
umt5-xxl-encoder-Q6_K.gguf
wan_2.1_vae_fp32.safetensors
@iLegoLoon sorry, it some hard to me, and my english. A dont understand about this "notes from your command prompt after the "got prompt" text?". If this about promt in 4 block first-extend, i try made a diffrent promt and 1 promt in all block.
@iLegoLoon i posted 1 video, can you please see what's wrong with him?
@iLegoLoon I just noticed I'm not using SVI Pro. As soon as HuggingFace is unlocked, I'll download Pro Lora; unfortunately, they're not available in Modelscope. I hope that's the problem. I didn't mean to offend you, my friend; you've done a great job. I'm just not very good at understanding all the processes; I'm new to Comfy, and English isn't my native language.
Well you figured it out. Its alright, the version is new so first posts are important. Cheers.
I updated the svi link to be more direct.
Can you delete the broken videos and reupload the working one when it is fixed?
@iLegoLoon i delete. try to download, when huggingface will work again
@iLegoLoon it`s amazing, all work. Problem realy was in use not prosvi lora. And that's what's needed Kijai lora, if i try use prosvi from vita_video_gen repo, it was just noise in result.
I've tried several SVI workflows at this point. I don't know why, but every time the face in the resulting video is not the face on the original image I input.
same here! I havent tried this exact workflow yet, but other 3 or 4 I tried before gives even worse results than without using SVI Pro at all. no idea why :)
Just try it but make sure the lora's and models match.
Try to shutdown any other loras maybe. This d*uckin loras changes the faces often
I was able to install the workflow and the results are stunning! I could run it on my PC with 32 GB of RAM and 32 GB of VRAM, but not enouph memory on my laptop with 16 GB RAM and 16 GB of VRAM.
Try resize pagefile to 80-100Gb
@northvnovgoroddanilov204 I recommend you set at least 200 GB virtual memory (the pagefile.sys file, that is usually hidden), preferably on a fast SSD partition, mine until recently was 150 GB and in some cases it wasn't enough (comfyui freezes), so it's now 200 GB and I would make it even 300 GB, but I don't have the space on my SSD.
You could also try quantized models and clip as it is used in the workflow. You can make shorter videos but it may shorten the action windows.
Thanks for v1.0! I keep testing and it's fun. I can see now you save the first temp video after the second part (10s), so you can make it save a video after the first part (the first one doesn't need to be lossless, just for preview), like in the previous versions, because in some cases I need to cancel after the first part (in case of slow motion) and try again with another seed. I've done dozens of tests and unfortunately the first high noise step without the lightx2v lora rarely prevents "slow motion", so it's just pure luck. I'd say 4 out of 5 are slow motion and occasionally it's the real speed.
Cheers buddy. Its an extra file in temp folder so I'm not sure :( I think unless it is requested by many you can add it yourself manually.
People mentioned that some checkpoints with lighx2v baked in did not have the slowmo effect. So I added in a 2 ksampler support that runs when the no lighx2v lora is set to 0.
Also added support for custom cfg for those 2 runs so one could run a no lighx2v lora queue but that would take too long :)
@iLegoLoon Yeah, I've added it myself. I'm relatively new to comfiui workflows (1-2 months) and i'm still not sure if i do the things the right way, but I keep learning new stuff daily from each workflow. For example until today I wasn't aware I can save lossless (ffv1) and now i can simply concatenate the parts via "ffmpeg -f concat -i mylist.txt -c copy output.mp4" and then compress. Also using buses seems super neat for complex workflows, othewise you get lost among the spagetti. 10X!
@iLegoLoon This Wan2.2 model has baked in lighx2v lora and is very good for fast motion https://civitai.com/models/2053259?modelVersionId=2477539 but I haven't tested it with your workflow (more than 5s) and I suspect the quality will degrade rapidly (most likely some wild motion after the first 5s), unlike the vanilla Wan 2.2, but who knows, I may be wrong. The vanilla model starts slow and slowly accumulate motion (it's getting faster and unpredictable) with every 5s part, which is bad, but seems unavoidable. Hopefully it will be fixed in some future Wan version that would be able to generate 20-30s out of the box.
@bnzarev821 I tried dasiwa v8 but its not any better. It also takes away the ability to run no lora step which looks relatively more prompt adherent compared to lightx2v lora only runs. I think these settings are alright overall.
Edit: 'Works' if you use gguf models + the new SVI loras and lightx2v loras mentioned. - Would take a more detailed look at the KJ node, but oh well, gonna see and look what LTX2 has to offer for now. Otherwise this is a pretty clean workflow. Original: The implementation is wrong. anchor_samples in the extend parts should be the last frames from the previous video, not the original reference image. You ARE passing previous_images through the bus, but you're only using it for video concatenation (ImageBatchExtendWithOverlap), not for conditioning. The anchor_samples still pulls from the static DataBus value (original reference). It needs to be: take last N frames from previous_images → encode through VAE → use as anchor_samples for next clip. That's how original SVI does it.
It is not my implementation but this is what keeps it consistent.
Images are only used for overlap. However the previous latent already goes into WanImageToVideoSVIPro node.
That node creates an empty latent. Puts reference latent at start, then places number of previous latents you picked aka motion_latent_count and fills the rest with empty latent (to be denoised).
Latent is already encoded images or if you decode latent you get images.
Because I tried feeding previous latent into anchor and it just degrades too fast and loses consistency pretty quickly. Imagine if my character is spinning and last latent it is just blurry. All the node have would be a spinning blurry view two times, nothing else.
@iLegoLoon But that's not what the Original SVI Pro (2.0) does. It takes the last frames as 'images' re-encodes them and uses that. Not raw latents passed through. This implementation as it is right now also 'loops' on each start always back to the reference image, instead of providing a seamless transition between each video. Did you try decode → re-encode → anchor samples? Or just passing the raw latent directly into it? Because those are different things.
@Amatiramisu Did you try the workflow? How do you think these videos are being made? I'm not feeding them in anchor samples. I'm feeding them in prev_samples. Anchor does not move, anchor stays where it is so ship does not drift away. If I were to feed them into anchor it would be the same thing fed in the same node.
@iLegoLoon Original SVI has TWO separate image inputs: 1. anchor -> stays fixed, stylistic consistency. 2. input_image -> updates each video, visual starting point that gets encoded.
The comfyUI node only has anchor_samples, there is no separate input_image. In the node only the anchor is marked as "known" in the mask. The motion_latents from prev_samples are 'not' marked as known, they are hints, not the visual 'starting point' - That's why it would always loop back to the initial reference image on each new clip. Currently it says to the generation: first frame should look like "original reference" and not : first frame should look like "where we left off".
@Amatiramisu I've worked with wan latents before. x number of latents get decoded into ((x - 1) * 4 + 1) images. A latent is just an infromation data type and frames are stored in its 2nd dimension. Its taking those frames from the latent, not in image format directly.
In this case since it takes an anchor as well you could say it takes 4 images with 1 motion_latent_count.
This is what happens in that node;
``motion_latent = prev_samples["samples"][:, :, -motion_latent_count:].clone()
...
image_cond_latent = torch.cat([anchor_latent, motion_latent], dim=2)
...
image_cond_latent = torch.cat([image_cond_latent, padding], dim=2)``
this means;
- its cutting last x number of latents from previous node
- making a new latent with anchor + last x number of latents from previous node
- adding padding to that new latent
What this lora does in my opinion is it just helps melting that anchor image and latent from previous part and spre*d into whole latent so it feels like it is continuing. You would get the start image if you disabled svi lora but it becomes continious when it is enabled.
@iLegoLoon The issue is the mask, not the concatenation: mask[:, :, :1] = 0.0 this marks only the anchor as 'known' meaning the model is being told "frame 1 must match this exactly". The lora helps blend but it can't override the mask constraint telling the model "first frame = anchor." - I've looked at the SVI implementation again and missed something. They use CLIP vision on each clip. this node has no clip vision inputs, that path is missing entirely, which would / could explain the blurriness issues.
@iLegoLoon Holy shit, what a mess. I thought CLIP vision was missing from KJ's implementation because SVI 2.0 Pro code has it - assumed that was causing the transition issues. Also thought maybe VAE 2.2 played a role in clip consistency since 14B can't use it (only 5B TI2V does, 14B still needs VAE 2.1).
Did a deep dive and turns out Wan 2.2 I2V 14B doesn't use CLIP vision at all - the official Alibaba code has no CLIP for I2V, only for the Animate model. The img_emb layer doesn't exist in the weights. SVI 2.0 Pro code has CLIP vision but checks require_clip_embedding and img_emb first - both are False/None for Wan 2.2 I2V, so the CLIP path gets skipped entirely even though the code is there. ComfyUI detects it as model_type='t2v' not i2v. If you pass clip_vision_h, it's completely ignored.
So the CLIP vision thing isn't the issue. The transition issues must be something else. Tried your updated workflow but still not getting the results I'm supposed to get, guess I have to try some gguf models for this.
hey again, im getting a "TypeError: must be real number, not NoneType" when second pass starts on the first Extend subgraph. not sure why, i cant read these jargon errors.
Hey, does it still persists? I'm trying a few quick fixes but let me run once again to see if last upload fixed it.
Seems to be working now @negneg, looking forward to feedback about the consistency over time as well.
@iLegoLoon Working now! Just gotta check the consistency now :)
wasnt working... then i pulled the kj svi...works well now..
do you mean you bypassed it?
Guys, and what if my first start frame in video must begins with my character stands from back and turning in front? How can i set his face before turning?
That sounds a bit complex, is there any way that you can get an image where the character is facing forward? Maybe using the qwen edit etc.
If you have that what you could do is;
- feed image where character's face is visible into the encode node (the default start image)
- modify your I2V-First to expose its anchor_samples variable
- put the image where character has their back turned through vae_encode and connect output to first sample's anchor variable that you exposed
- the rest of the parts will have the inital image where his face is visible except for the I2V-First
@iLegoLoon or somehow i need to combine first last frame node with SVI
@jafdeth2030105 You could do that as well but would still need two images to not lose consistency. Just feeding a step custom image feels easier tbh. You dont even need to expose the anchor_samples. Just go in First subgraph, load and encode your image and feed it in.
I noticed there is a SVI V2 and a SVI V2 Pro, i'm having a hard time finding a changeling explaining the difference between the non pro and pro version
I've updated the links to v2 pro, this one uses svi v2 pro which is new and seem to work better.
Thank you for sharing this workflow. I have been testing it on an RTX 3080 10GB and it works perfectly. Greatly appreciate the effort!
Just added an Ollama subgraph to this workflow and it now rocks in autogenerating the initial z-image and storyline for the videos. This works exceptionally well.
I do not see any videos shared :(
So I got this working but the main problem I see is that individual videos in the sequence may give bad or undesired results. What is really needed is the ability to generate the videos individually in sequence so that you can "re-roll" them over and over but still keep consistency of the original image and continuity from the previous video, and then move on to the next video in the sequence.
I guess if I had a faster card it wouldn't matter as much but trying to generate at a decent quality on my card takes around 40 minutes and so having the whole thing fail due to one or two bad clips and having to start from scratch with no real progress to show for it makes this hard to be viable.
Before each generation make all seeds fixed by clicking "new fixed random" in all 4 subgraphs, then if for example the 3rd clip is bad, manually change the seed only in the 3rd subgraph by clicking "new fixed random" and click the Run button to generate again, this way only the 3rd and all next clips will be generated again (in this case the 3rd and the 4th), so you don't need to generate all 4 clips from scratch. But once you close Comfyui, you'll have to generate from scratch, no way to save progress and continue tomorrow for example.
@bnzarev821 Great thinking! I will try this.
Hi there, how apply lora's into specific clips ?
If you open the I2V subgraphs there is lora loading there
I'm pretty excited to see this in action. I worked through getting all dependencies sorted, but when I run the workflow, I'm getting the following error from the I2V-First subgraph WanImageToVideoSVIPro node....
# ComfyUI Error Report
## Error Details
- Node ID: 24:22
- Node Type: WanImageToVideoSVIPro
- Exception Type: ValueError
- Exception Message: not enough values to unpack (expected 5, got 4)
Somebody had the same issue below and said this;
Edit: Turned out to be a node conflict or something not updating correctly. Removed all custom node and installed just what is in the work flow and all issues resolved! Works excellent!
Maybe delete the kjnodes from custom_nodes folder and reinstall it or find if something is conflicting with the node.
@iLegoLoon thanks for the response - I failed to mention that I did delete the KJNodes that was installed via ComfyUI Extensions and then installed it from Git, the issue persists.
@SloanSteddyAI Yup I'm getting a similar error about WanImageToVideoSVIPro node
KJ nodes have a few updates. I will take a look but did not see an update on the videosvipro node.
Edit: There's no difference.. are you guys sure you have connected your images into encode subgraph if you arent using ZIT?
@iLegoLoon I am using ZIT. After importing the workflow, I haven't made any changes except for adding the missing nodes and re-selecting the appropriate models where necessary (folder structure SLIGHTLY different).
So it turns out that people may use ZIT vae I've provided in Wan2.2 Loader.
Everything about ZIT is inside its own subgraph. Loader subgraph is for WAN2.2 only. The issue is caused by the UltraFlux vae I've provided being set as Wan2.2 vae.
Fixed after switching to wan2.1 fp32 vae.
@iLegoLoon Yep! Thanks for helping out, greatly appreciated!
I am running latest ComfyUI and deleted KJNodes and reinstalled it (newest versions & nightly) but still getting:
This workflow has missing nodes
This workflow uses custom nodes you haven't installed yet.
Basic data handling: CastToIntin subgraph 'Steps'
Basic data handling: MathFormulain subgraph 'Steps'
Basic data handling: SwitchCasein subgraph 'I2V-First'
Basic data handling: StringRsplitDataListin subgraph 'Final Merge'
ComfyUI-basic_data_handling its literally in the name :(
Weird that it cant find it directly tho.
@iLegoLoon Thank you! Worked!
dang thanks man, I was doubtful about SVI but now I'm a believer
If your using all the fp16 models, wondering what hardware do you have to run this workflow? Im guessing atleast an RTX 5090?
I've no idea. I remember running them since block swap and other fallback methods work in the background but it may end up slow. For speed its a good practice to use quantized on all parts.
fantastic. works perfectly. can you share prompt of first clip? the girl's walking ?
I've lost them since the metadata was turned off :( it was something like;
"woman walks forward, humble smile, being silly, vlog"
You get the vlog walk if the image looks like a selfie. I've also used realistic snapshot and lenovo as z image loras.
nice but the version is very slow
i guess this is because you load and decode the vae for each group
this guy : https://www.patreon.com/posts/146997189?utm_source=youtube&utm_medium=video&utm_campaign=20251230 made a different approach that is faster.
I would love to have your customization and his speed ;)
It gets stuck for some reason, I cancelled the run, switched the lightx2v count to 0, it ran fast, turned it back to 1 now it runs superfast. No idea why.
3 pass has a cfg>1 run at start so first sampling takes more time. But it increases the movement good enough.
There's not much time lost in saving the files. You decode the latent no matter when you do it and there's no re encoding going on. I'm simply feeding the latent to the next node with the decoded images. It is necessary to do so to prevent images stacking in memory.
Sadly, all of my videos are coming out in slow motion :/
You could increase the number of steps or go no lighx2v but thats the side effect of lighx2v loras :(
You are missing WanImageToVideoSVIPro node.
Delete kjnodes from custom_nodes folder and reinstall it.
@iLegoLoon I got a similar error but with Cannot execute because a node is missing the class_type property.: Node ID '#10:122' and Error loading RIFE model: Failed to download RIFE model: HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /lividtm/RIFE/resolve/main/rife49.pth. I tried reinstalling kjnodes and nothing changed.
awesome work man thanks. you think svi pro is possible to be used with kijai's wanvideowrapper nodes to make extended videos like this?
I have seen implementations that use wan wrapper from kijai. I do not personally use wan wrapper myself tho.
How can I remove the increase in video contrast? It starts around the third stage, and it seems like it's only going to get worse.
It should not happen if you are using the same models/loras and not using any custom lora's.
Check your console to see if it does tiled vae decode instead of normal one. Then you can replace vae decode nodes with tiled vae decode but replace settings to have the least artifacts.
Use fp32 vae if you arent already.
If it still happens with these and linked models/loras you can try to change shift value and experiment. It used to become brigher as you went down from 8 in old workflows but not sure about this one.
this issue is covered here in some detail: Summary of Community Workflow Variants (Under Test), Tutorals, and Tips · Issue #51 · vita-epfl/Stable-Video-Infinity
In my testing, I'm using the default workflow but with Q5_K_M models and Q6K Clip
FP32 vae and the lightx2v '1217' models from here >> (yes I know these are T2V models but they work with I2V and the I2V ones on this page do NOT work for me..) >> lightx2v/Wan2.2-Distill-Loras at main
I personally have had bad output from Q4 models, hence I only ever use Q5_K_M and above.
how many step count should I set if I do not use Lightx2v?
If no lora at all with 2 sampler you should set no lora steps to 0, set cfg to 4 and disable lightx2v lora load nodes. default step count for wan was 20 steps total -> 10 steps per high and low.
This is the best made workflow I have ever encountered, extremely clearly structured and everything worked instantly with great results on my 3090!
When ksampler runs repeatedly and fixes a seed, it doesn't skip that seed when starting the next section, but instead starts generating the video from the beginning again. What could be the problem?
the random seed node does not save the seed it randomizes if it is -1, you have to set random fixed value for each then run it that way. it should not skip that way otherwise I have no idea.
This is a fantastic workflow! Well done! One thing that I've added was the Moe calculation (WanMoEScheduler) for the ModelSamplingSD3. I also prefer to use lora stacker from Lora Manager instead of the power lora loader.
May add a request/suggestions?
On the 'Final Merge' Could you make it stitch the first and second part separately, and then stitch the third, and so on, in a way that it doesn't need to recombine the whole full video in the end if I just want to regenerate the last part? What I mean is, in a second generation that I just changed the last part, it wouldn't stitch all parts again, just the last one... IDK if I made myself clear.
Cheers, feel free to modify it to your preferences.
You can bypass the node where it tries to load them and manually stitch the videos from your temp folder using a software like shotcut.
I'm having an issue with this workflow. As early as blocks 3-4, the video starts to drift, colors become distorted, and details are lost. Could it be because I'm using a lighting model with the LIGHTX2V LoRA disabled? The SVI LoRA I have installed is this one: SVI_v2_PRO_Wan2.2-I2V-A14B_HIGH_lora_rank_128_fp16, plus a low-noise version. I suspect WanImageToVideoSVIPro might be ignoring the anchor_samples from the initial image as a reference for color correction. The no_lightx2v steps are set to 0, and the other two are set to 2.
Does the model have extra lora's? Even lightx2v lora versions make a difference in the quality. Try to play around with the shift settings if the model doesnt have extra lora's baked in. Or maybe give the Q4 model with linked lora's a shot :)
I downloaded all the dependencies listed in the description. I get an error: "CLIPLoaderGGUF: only 0-dimensional arrays can be converted to Python scalars".
I chose the ClipLoaderGGUF - umt5-xxl-encoder-Q4_K_S.gguf. Why does this error appear?
It has to be something about the gguf nodes.
People had similar issues and there seems to be not a lot done about it;
https://github.com/Comfy-Org/ComfyUI/issues/11542
https://github.com/city96/ComfyUI-GGUF/issues/384
Check if you can run the fp8 version or if other quants give the same issue.
I made significant changes to this workflow to incorporate Ollama running Qwen3 VL and instead of having it generate an enhanced prompt per video segment I have all the user entered prompts be gathered up and sent to Ollama so it only needs to run once and then the model can be unloaded. it then goes to each video segment and parses the response from ollama to provide the correct enhanced prompt to each video.. Many upsides to this approach... It only needs to run once and it has context from all other video segments. As well as context from the start image. I used the Bus system as much as possible but since it needed to collect the user prompts first it needed a separate prompt input node for each. I added up to 12 segments(1 minute of video) with quick on/off toggles that bypass uneeded ones. I moved rgthree seed nodes outside of the subgraphs so that you can reuse a video generation easily. The only downside is if you change the prompt on any of them it regenerates the enhanced prompts for all of them and thus restarting the video generation from the start even if you have fixed seeds. I'll be uploading it soon... maybe
Great work!
Will you be posting your updated workflow? :)
@Silver_bullet666 Yeah probably in the morning. I need to write notes on it and setup a clean version
Still testing. but I have gotten around the Ollama regenerating all enhanced prompts when one prompt is added or changed with a prompt lock switch(it saved the enhanced prompt to a text file and once you lock the prompt it ignores ollama's new generation for the enhanced prompt) Since ollama only generates once for all prompts it would have had to run anyway if it was still generating once for each prompt so it acts in the same way but it always only generates once in a format as follows first enhanced prompt::SPLIT::second enhanced prompt::SPLIT::third and so on... it parses through each prompt using the ::SPLIT:: identifier but when it comes to a locked prompt it just uses the last saved prompt from the text file which only gets access to the new enhanced prompts when the prompt is unlocked.
@operation12studio300 Just checking in on this, sounds awesome!
Hi, your SVI workflow is great! Clean, organized and easy to use. Do you have any intention of creating a workflow using SVI to extend a video? V2V
I could. I'll put it in the zip if I have time on weekend.
Its just as simple as;
- load a video
- encode it to latent format
- encode first frame into latent format
- feed them to I2V-Extend subgraph after exposing variables (bypass I2V-First)
Just what I was looking for. Great work!
You can use rgthree's fast groups bypasser to bypass nodes in subgraphs from the outside which makes turning stuff on and off nicer.
Also a question,
How do I use the lora loaders inside the first/extended subgraphs? Can I put high/low loras in them, do I put the same loras in each one?
Surprisingly, this wf works well and i get pretty good image quality through a whole 30+ second generation.
But I have 1 big issue. The promt adherence is quiet poor! Never had this issue before. Is it because I am using gguf q8 model? What text encoder shouldmi use with gguf q8?
Could be because of svi lora's with 3 phase sampling setup. You get 1 latent less than usual because it is comes from the previous video and have only 1 high no lora step, rest is lightx2v loras. You could try with more trigger words than plain prompting. But its trial and error with some seed luck involved.
Is there a way to avoid using SVI for the first clip?
You can try to replace the svi node with default node in the first node but it will have the lora regardless. You will need to disable lora from model loader and enable it in each part. But if you just dont use the previous latents it generates the same latent.
@iLegoLoon I tried this as well as by passing the first node entirely and generating a new latent and passing that instead to to extend node similar to your recommendation for V2V extend. In either case the output appears broken ie weird yellow color and artifacts. Would love to see an example.
@OwnPear8996980 Is there a reason why you want to avoid svi? It should not make a difference since there are no previous latents in the first one. Doesn't it work as expected?
@iLegoLoon From my testing SVI does seem to affect prompt adherence on the first video, maybe it's just me.
Low vram rtx3080 10go ram passed your way, this workflow is too heavy.
I do 15 secondes with my rtx for info with svi lora.
In theory it doesnt use incremental vram until the very end so its using same amount of memory and processing.
At the end it tries to load all part videos that are temporarily saved and stitch them together. That could be where you had an OOM. You can bypass "Load Video (Path) 🎥🅥🅗🅢" Node inside Final Merge subgraph to prevent it from trying to load all videos.
If you go into the temp folder you can find the part files (deleted after comfyui restarts)
You can view those part files using a software like VLC and merge them using something like shotcut.
it works beautifully!! best workflow ever! thanks you so much. usually video quality is degrading at each extend, this workflow doesn't! marvellous!
I was hoping this would prevent faces and penises from slowly changing every time they were obstructed, but I'm finding it hard to generate any penetration with this workflow, even with appropriate loras.
If the properties are missing in the first image they will not be preserved. If they are in the first image it doesnt matter if at some point your subjects decides to go off camera. It will always be fed into the latent.
Loras are not trained to make videos longer than 5 seconds. Any lora will affect the output in some way. It only depends on how long you want to make the video.
@iLegoLoon I guess the implication of that is if I want to have a subject preserving video in which penetration occurs later on, I need to manually splice in at least one section using a different, non-preserving method?
Either that or generate an image variant where the features you want to preserve are in the image. Feed it into encode subgraph. Then go inside the I2V-First and encode the actual first image you want to use/feed it into anchor_samples. This way you will have a different starting image but the anchor sample will have the features you want to preserve.
Runs great, but scene restarts from the image instead of continuing from where it left in every stage, other times there is a very easy identifiable scene change.
I am, running Q5 quants.
Make sure you have svi v2 pro loras. If thats clear then try to see if you have the exact same lightx2v loras. If it is restarting/ghosting something must be wrong.
I've just tried a new idea to generate 1-2 min consistent vids: instead of cancelling the generation after each bad part and then regenarating, I let it generate a 8-10 min vid (~100 parts) with all body horror etc., then simply choose the longest consistent sequence. It works fine for simple repetitvie plots that don't require deferent prompts for each part, for example "a catwalk". "the river flows", "a busy street" etc.
I see the enthusiasm. It is good to have generic prompts if you are aiming for longer videos indeed. For short range videos lora's help a lot steering the video. Its funny to still see videos with noticable transitions get uploaded while this is possible :)
Seems you have removed the video tho.
getting this error "got prompt
Requested to load AutoencodingEngine
loaded completely; 25014.71 MB usable, 159.87 MB loaded, full load: True
!!! Exception during processing !!! not enough values to unpack (expected 5, got 4)
Traceback (most recent call last):
File "/workspace/pinokio/api/comfy.git/app/execution.py", line 518, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
File "/workspace/pinokio/api/comfy.git/app/execution.py", line 329, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
File "/workspace/pinokio/api/comfy.git/app/execution.py", line 303, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "/workspace/pinokio/api/comfy.git/app/execution.py", line 291, in process_inputs
result = f(**inputs)
File "/workspace/pinokio/api/comfy.git/app/comfy_api/internal/__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "/workspace/pinokio/api/comfy.git/app/comfy_api/latest/_io.py", line 1582, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
File "/workspace/pinokio/api/comfy.git/app/custom_nodes/comfyui-kjnodes/nodes/nodes.py", line 3042, in execute
B, C, T, H, W = anchor_latent.shape
ValueError: not enough values to unpack (expected 5, got 4)
Prompt executed in 2.46 seconds"
you are not using the right vae, dont pick vae meant for z image inside the model loader. everything about z image is inside its own subgraph.
I think it can be optimised by encoding all the promts first and only then starting generation for every part. Because it takes quite a long time to load/unload text encoder every time (and unload VAE/Wan from vram by this as well).
Yeah that could improve a little but the biggest issue is the high/low model switches for now. It'd be great if we could do all high first then all low later on.
@iLegoLoon It should be doable with saving high-noise latents to ram/disk? The same for promts latents, as no need to keep them all in vram.
Thanks for making this. I can see there was a lot of work put into it. My results from using it are ok. I'm not sure if I will end up using this extensively but I'm glad it worked very easily out of the box and it is something I can play around with.
this is the best wan workflow on civit.
IT. JUST. WORKS.
And it's very simple. +1
first things first , its the BEST SVI workflow I came across, Cheers! I am trying to load these checkpoints https://civitai.com/models/2053259/wan-22-enhanced-nsfw-or-camera-prompt-adherence-lightning-edition-i2v-and-t2v-fp8-gguf?modelVersionId=2584698 but not getting results I expect , when using default with lora faces & movements changes a lot & without bypassing loras in loaders & changing to 4 still not giving desired outputs , any further step I am missing ?
Thank you.
It could be the embedded loras degrading the quality. If you have a clean first image and those details are being lost that is most likely the issue. Because even the lightx2v lora you use matters.
Thank you for your great work !
Have you guys noticed that videos generated from SVI Pro have one pretty obvious flaw: the further you go toward the later clips, the video quality drops noticeably? The details of objects get way blurrier compared to the first clip, and in some more complex videos, the character's facial features even start getting messed up.
Does anyone know if there's any method to reduce or fix this issue? Please help me out !
I've noticed that and i don't see much people talking about it, all the details of the face disappear and the color tone of the face changes and become brighter.
@juanitaz Yeah that's right brother !
The details of the final clip was very bad :(( .
I think clip joiner using VACE is the good way to make high quality long videos for now .
If you aren't using the linked base model and lighx2v it may not work as good. Quality depends on the embedded loras and lightx2v loras.
There are 2 minute video examples shared showing almost no quality loss at all.
@iLegoLoon Thank you for your reply ! <3
@iLegoLoon Ohhh ! it worked ! Amazing work mate ! <3
@TheBodyHunter do you have a workflow for vace long video, that has consistency and no color shift problems?
@grasshopper85116 You can try VACE workflow from : WAN VACE Clip Joiner - Native workflow . This is a good WF for long video combination .
How do I use Power Lora Loader with this workflow? I can't get the 'model_bus' or 'data_bus' to connect to the 'model' on Power Lora Loader. Any help appreciated. Thanks
Each part has a power lora loader inside it if you go in.
First two are high noise lora loaders and the third is low noise lora loader.
If you want to add a general lora the loader is in model loader subgraph.
@iLegoLoon Oh wow, that's awesome. I've never played with the subgraphs before. Learned something new. Thank you. For anyone who doesn't know what this is, just click the little box at the top right of any of the I2V-First/I2V-Extend nodes.
I tried your settings - exact match, and then played with many others, but all I get is ribbons and explosions! I then made changes... if the CFG High-Low is set at like 0.5-1, but now it blurs. I have already bypassed the light2xv and have tried all the values of high_no_light2xv from 1-4, tough, unsure a I have tried other workflows and haven't had this issue, but I love the neatness of this workflow!...
Just can't get camera motion to work consistent. Any ideas? I've heard lightx2v can mess up camera instructions in prompt. Anyone found successful settings? I'm running the default settings and almost the same models as in the original workflow.
It becomes a bit rebellious with both lightx2v and svi lora itself. Best approach for consistency would be to have fixed seed for each part and try different prompts and seeds for the failed parts.
@iLegoLoon thanks. I had issues with pull back. It improved things somewhat by having the same seed but I think but the main thing was being more explicit in the prompt, e.g. the camera pulls back to reveal more of her body and more of the surroundings. It seemed like telling the model to also show more of the surroundings (which it must when it pulls back) did the trick. Or it was something else... :-)
Does v1.0-3Ksampler work for a model that has merged lightx2v Lora already? I noticed the fitst Ksampler just bypasses the HIGH LIGHTX2V node in the workflow. But it does improved slow motion, comparing to Wan MoE KSampler
It does not, you have to set high no lora steps to 0 to bypass the first sampler. Different lightx2v loras may give different results but you can try it with 0 high no lora steps.
Just a question about disabling lightx Loras in your workflow. As the Note box mentions, this can by done by bypassing those nodes in the model loader subgraph. Is just doing this enough, or is a higher "high-low-cfg" required for it to work?
You need to;
- disable lora loaders inside the model loader subgraph.
- set high no lora steps to 0
- set high-low-cfg to something like 4
Does disabling lightx2v lora significantly increase generstion time?
Noob question. im getting an error
!!! Exception during processing !!! [GetNode] ✗ Variable 'upscale' not found!
Available: MOTION FRAME, steps, scheduler, CLIP, MOTION AMPTITUDE, VAE, LENGHT, second_steps
Tip: Make sure SetNode runs BEFORE GetNode in the graph.
any directions on where to look?
I dont have get / set nodes in my subgraph as far as I know.
Thank you for the response. it is fixed, was a class name collision between ComfyUI-KJNodes and PromptModels Studio, after updating PromptModels Studio to 1.0.6 it worked fine. Awsome work btw.
Is this workflow still working? I get Cannot execute because a node is missing the class_type property.: Node ID '#29' using latest comfyui. I have installed all nodes.
Details:
Error: Prompt execution failed
at ComfyApi.queuePrompt (http://127.0.0.1:8000/assets/index-ClJ8cazm.js:4:5485)
at async PromptService.api.queuePrompt (http://127.0.0.1:8000/rgthree/common/prompt_service.js:88:28)
at async ComfyApp.queuePrompt (http://127.0.0.1:8000/assets/index-ClJ8cazm.js:277:17632)
at async app.queuePrompt (http://127.0.0.1:8000/extensions/rgthree-comfy/rgthree.js:478:24)
at async Proxy.function (http://127.0.0.1:8000/assets/GraphView-BJurfd1H.js:15:74005)
at async http://127.0.0.1:8000/assets/index-ClJ8cazm.js:75:267894
at async Proxy.execute (http://127.0.0.1:8000/assets/index-ClJ8cazm.js:75:287723)
at async keybindHandler (http://127.0.0.1:8000/assets/keybindingService-DnU9p4Gw.js:1:2956)
Node ID #29 is bus node from Comfyroll Studio. It worked just fine in a fresh comfyui setup I had.
@iLegoLoon if I remember correctly, Comfroll Studio wouldn't install at all for me due to conflicts so I'll try a different route. My knowledge is too limited to troubleshoot your workflow. Thanks anyway for the reply :)
@snobbias124 Check out v0.9 in that case. It doesnt use bus system and may lack a few other features but svi works.
I have tried so many svi 2.0 pro wfs, and this seems to have the highest quality video outputs!
But I am still suffering from very weak promp adherence. Tried so many things, to no avail. Is there any solution to this yet, does anyone else experience this?
I guess it has something to do with first few frames coming from previous latent and also svi lora itself. I hope we get multiple input frame support for native models in the future since svi is technically just a workaround.
I've discovered what exactly is wrong with this when using the 3 sampler method. If you use that method in this workflow, the first sampler is not configured well. It has the denoising schedule all wrong for that first sampler. The first sampler needs to be expecting a longer schedule since there are no lightx2v loras on it. I personally rewired everything to use a TripleKSampler node, but you can do it manually by setting the samplers to do something like:
Stage 1: Base high model - steps 0-3 of 24 (denoising 0.0%–12.5%)
Stage 2: Lightning high model - steps 1-4 of 8 (denoising 12.5%–50.0%)
Stage 3: Lightning low model - steps 4-8 of 8 (denoising 50.0%–100.0%)
Then i use very low strength loras in the base stage. Once I did that, prompt adherence and camera motion got much better.
@ImSoDan hi buddy, that's awsome! Would you mind sharing the workflow? Please feel free to message me with a link to it. I would really appreciate it, thank you!
Is version 1.0 still able to make extended videos? I may just be confused on how the nodes work but I don't understand how to make extended videos with you v1.0 workflow.
It isn't. I tried to have extend and first image setups both functional at the same workflow but comfyui logic and validation are not my favourite. I dont think I'll be adding it in near future.
@iLegoLoon Gotcha, thanks for responding. I appreciate the work your doing. I've been playing with v0.9 and it's been great!
Is there a way to save and continue a run without having to regenerate each part when reopening comfyui?
I dont think so, video extend would make it work where you could load the last generated part or the whole video to continue only if comfyui node validation did not suck. But for now I'm too busy to implement it in an acceptable way.
How would I fix error.
VHS_PruneOutputs
Paths don't have the same drive
I know the gist of comfy nodes but not specifics so I couldnt figure it out. Any help would be appreciated
https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/issues/340#issuecomment-3115431003
Try saving in the default output folder if you have modified it or check the path where it saves and manually set it if it is wrong.
You can disable prune outputs if it keeps giving errors, you'll just have extra images next to your video in the output folder.
@iLegoLoon Thanks for the reply, I tried that fix, but still got the error. Ill just disable the node. I appreciate the help.
修改你的启动批处理文件（.bat），在开头加入两行：
@echo off
:: 在这里设置临时文件夹到 D 盘
SET TEMP=D:\ComfyUINeo\temp
SET TMP=D:\ComfyUINeo\temp
:: 如果没有这个文件夹，记得先手动创建一个
mkdir D:\ComfyUINeo\temp
:: 后面是你原来的启动命令，比如 python main.py ...
I'm having this issue as well. I had to remove my D drive output and just let it go to default. Really rather it go to the D drive as normal. Tried disabling as well but got errors. any other ways to fix this?
How can i use this workflow 1.0 withou svi lora? When I bypass svi lora, the video gets a lot of artifacts, I know that svi is good for consistency and transitions, but most part of the time "blocks" the custom loras, Is it possible to use this workflow without svi?
You can use v0.4 for plain last frame stitch method but I dont remember the model and lora references there, it was a while ago.
Just wondering if this is normal generation time for a RTX 5090.
If i have 4 segments for 20 sec video total. 1280x720p resolution. Light x2v v2022. Euler, simple.
It takes about 20 minutes to generate. Quality is high, but i feel like this is a long time for an rtx 5090, or is that about normal?
Depends on the model size, are you using Q4? For me sage makes the biggest difference in speed. Other than that my default resolution is 480x832 and it takes about 10-15mins for 3 parts depending on the sage + torch combination (torch is broken on my current setup)
I think you have some space for improvement.
@iLegoLoon i am using q8 gguf, q8 text encoder. No sage/torch.
im curious how long mine would take on 1280x720
I'm using Q8 just testing a 480x480 "Prompt executed in 289.98 seconds" Running 5090 - 64G DDR5 on i7 14700k.
No, it's not long, it's fast))
@sizyking ok thanks. But is there any better svi 2.0 workflows? Ones with better prompt adherence?
What do I have to do so my video is not slow motion?
Unfortunately slow motion comes from lightx2v and svi loras. You can;
- use some keywords on first part to trigger faster motion like rapid, fast etc.
- increase no lora step count, it is set to 1 because it has a >1 cfg and takes longer than the lora steps.
- keep an eye on the generation to make sure the first step goes with normal speed.
There were also custom nodes for speed which you could add to the workflow and use on the first part only.
@iLegoLoon can you remember which custom nodes speed up motion?
Set the high noise Lightx2v lora strength to 1.3-1.5. More will speed it up
Ok so I have been getting much better motion speed and prompt adherence by using
Set to 3.5 on high:
lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16
Set to 1.5 on low:
1217 low lightx2v
It's not perfect but it is much better motion and prompt adherence than what i was getting before, so this is in the right direction for me.
How to avoid character change? That generate a completely different character in my case
if the first image has your character and it still changes you must've set something wrong. this is an image to video workflow so make sure to use i2v models, linked svi and lightx2v loras and you should be fine.
I get it, there a was Z-I-T section to me text to video
Thanks. This generate better video compare to other SVI workflow, i used exact same models but the video come out so different. Great job
This is the way.
Used your templates for a long time, now I've updated to v1.0 and getting "Required input is missing
pipe", didn't change anything .
Comfyui update broke the workflow, needs an update
EDIT: I left the comment below thinking this was a totally different workflow... I'm leaving it up because people are actually having issues.
I love this workflow but it isn't working anymore with the recent ComfyUI changes. Does anyone know of another workflow that accomplishes the same thing?
yea it sucks, idk it seems broken. From what I can tell it's the comfyroll studio bus, but I think maybe the whole thing needs to be replaced which is like 50% of the workflow lol
Not as good as this one since you can't set different lora for each section, but at least works on the latest update of ComfyUI and you can modify it manually and add lora loaders on high/low inside each section, I don't see why it wouldn't work.
https://civitai.com/models/2418122/wan-22-svi-long-form-video-image-to-video-diffusiongguf-workflow-w-upscaling?modelVersionId=2718673
Thank you))
For those facing the problem when trying to use this with a newer version of comfyui. I got this to work by doing a fresh ComfyUI 15.1 install. I had to manually clone ComfyUI_Comfyroll_CustomNodes into custom nodes but other than that it works without a problem.
Required input is missing
pipe
and i have no idea how to fix it
One of the updates to ComfyUI broke passing some values to subgraphs. All I had to do to make this wf work again is unpack all of the subgraphs.