⚠️ The I2V Vid extend workflow is highly experimental and requires high end GPUs.
I've created this Wan Video Image2Video workflows based on Kijai's workflow and the Native ComfyUI workflow, added an easy to use interface along with upscaling and frame interpolation.
Key Features:
1st pass - generating vanilla video
2nd pass upscaling using user selected upscaler and frame interpolation to 32FPS
This workflow comes pre loaded with my Wan14B RunPod template:
https://civarchive.com/articles/11960/wan14b-in-1-click-with-workflows-included-runpod-template
Join my Discord server for updates on new LoRAs, tips & tricks, Workflows, RunPod templates and the holy right to be close to an amazing charismatic person like myself:
https://discord.gg/fyha5Pzm
Description
Updated the workflow with:
Optional TeaCache and SageAttention
Removed the first 6 frames from the video, they are junk frames that hurt the flow of the video
Added LoRA support (1 LoRA for now, testing more)
Added some notes to explain existing and new nodes
FAQ
Comments (40)
this workflow not working well for me. makes my entire PC lock up with a 4090 and 64gb ram. it's way too intensive and slow.
Which models are you using?
@Hearmeman wan2.1_i2v_480p_14B_bf16.safetensors
@zengrath And the same model works with other workflows?
I've also struggled running this model on an RTX 4090
@Hearmeman Honestly i struggle with wan. I found best workflow for me so far is one using GGUF model instead. it has less chance of locking up my PC even at longer frames or high 512x768 resolution for example. but the non-gguf model like one used in your workflow i find is far too intensive compared to hunyuan
@Hearmeman On another workflow wan2.1-i2v-14b-480p-Q4_K_M gguf model is working much better for me.
@Hearmeman this is one i am using atm, with gguf on, and model i mentioned, it doesn't have sage attentio sadly so it's taking me around 4-5 minutes on 512x768 but it's giving me best results sso far with a lora and I2V i seen so far. https://civitai.com/models/1306165/wan-video-yaw-workflow-v2v-t2v-i2v-upscale-extend-audio-interpolate-random-lora-preview-pause-upscale-multi-res-interpolateprompt-saveload?modelVersionId=1505707
@zengrath I don’t like the GGUF models so my workflow is not made for them.
I turned on quantization so I could get this to work on a 20gb vram card, but I keep hitting an OOM error with triton:
triton.runtime.errors.OutOfResources: out of resource: shared memory, Required: 106496, Hardware limit: 65536. Reducing block sizes or num_stages may help.
where do I adjust these? can't find anything on it, my google skills are failing me at the moment
I cant seem to find the model "LOAD WANVIDEO CLIP ENCODER" - that refer to your "open-clip-xml-roberta .. "
It does not seem to be one of the clip or text encoder of WAN that Ive been using . can you help?
workflow seems awesome though
also what;s the teachcahe model you re using, this seems new . thought teachcahe is a node
Hi, I’m not exactly sure what you’re struggling with.
What’s the error you’re getting
Always stops at PathchSageAttentionKJ
How do I fix that? Triton and sageattention are correctly installed. Where is this node PathchSageAttentionKJ????
That's the node called SageAttention in the loaders section.
Disable it if it doesn't work for you.
@Hearmeman Where is that node installed from? I have sageattention and triton on my pc.
@TekHousE KJ Nodes
Thanks for this amazing workflow, have had lots of good output from it. Just a quick question, why have you set TeaCache to "0.040" and start_step 6? The official values for a 480p model is 0.13, 0.19 or 0.26 and the default start_step from Kijai is 1 - just curious!
You're referring to the Kijai workflow I assume, that's the values he defined for his TeaCache implementation, in the native workflow it's 0.26 and 2
@Hearmeman Maybe I didn't explain it properly, I just wanted to ask why in your workflow, TeaCache is set to 0.040 and start_step 6. Did you find these settings to be better than the official ones?
@Eshinio I understood you.
There are 2 workflows, a workflow that works with Kijai nodes and a workflow that works with the native ComfyUI nodes.
The native workflow is set with the values you mentioned.
The workflows that is set to work with Kijai's nodes is set with the values Kiaji recommended in his github.
@Hearmeman This isn't related to the node type, it's dependent on version and whether or not the version you are using implemented coefficients or not.
Man oh man we need a way to "extend" the clips we generate. Maybe use the last frame as input.
Hi there
I’m working on a workflow that does this, results are so so…
What's really critical for longer vids is to be able to control both ends; that way one can guarantee keyframes are upheld then cherrypick from amidst the gens. I'm a bit perplexed as to why open models are lagging behind on that, though admittedly it hasn't been very long since hunyuan's release. Back when 1.5 came out, controlnet was all the rage.
Has there been any talk of this feature?
@fedupscribe687 @firemanbrakeneck I just released another version with video extension
@Hearmeman maybe for second video can you try to use another prompt?
Heads up guys, someone just managed to modify kijai's nodes to do start + end frames! And kijai may soon merge that into his repo.
https://civitai.com/models/1374287/comfyui-wanvideostartendframes
So good workflows and all, but how do I upscale videos that are already generated?
You can remove the sampling nodes the add a load video node, connect it to the upscaling nodes.
I still struggeling with this. Got Sageatt installed, but it seems to trip over Teacache? I've got that installed it works with another WF, but can't find what node uses it and whats wrong.
This is the Error I receive:
Failed to validate prompt for output 97:
* WanVideoTeaCache 242:
- Value not in list: cache_device: '-1' not in ['main_device', 'offload_device']
Output will be ignored
Failed to validate prompt for output 80:
Output will be ignored
Failed to validate prompt for output 95:
Output will be ignored
Failed to validate prompt for output 94:
Output will be ignored
Failed to validate prompt for output 237:
Output will be ignored
Prompt executed in 0.02 seconds
So Output '97' is there a way to search for this in the WF? Hopefully someone knows what I'm doing wrong now, I'd love to get the kind of results that are shown here.
Did you update your ComfyUI and WanVideoWrapper nodes to the latest version?
@Hearmeman I did, but something is defenitely wrong there:
The update for the following custom nodes has failed:
ComfyUI-WanVideoWrapper
Any idea on how I can solve this?
Thank you so much, love your work!
Mm, it seems after updating I'm missing a node now:
'ComfyUI-WanVideoWrapper': With the current security level configuration, only custom nodes from the "default channel" can be installed
at least I know where to search now.
@tiskloteweer588 Install it using git clone, not from the manager.
@Hearmeman That did the trick, thank you for your patience.
Everything is updated, Teacache is installed (works perfectly with another WF),
Sageattn is installed and working according to controlscript from https://github.com/woct0rdho/triton-windows?tab=readme-ov-file#install-from-wheel:
---
D:\ComfyUI\python_embeded>python test-triton.py
tensor([0., 0., 0.], device='cuda:0')
If you see tensor([0., 0., 0.], device='cuda:0'), then it works
---
I've tried to WF's the Sageattn/Teacache gives:
--
Can't import SageAttention: No module named 'sageattention'
--
And the Teacache:
---
Failed to validate prompt for output 97:
* WanVideoTeaCache 229:
- Failed to convert an input value to a INT value: end_step, offload_device, invalid literal for int() with base 10: 'offload_device'
Output will be ignored
Failed to validate prompt for output 80:
Output will be ignored
Failed to validate prompt for output 95:
Output will be ignored
Failed to validate prompt for output 237:
Output will be ignored
Failed to validate prompt for output 94:
Output will be ignored
Prompt executed in 0.02 seconds
---
I might be reaching too high above my abilities.
@tiskloteweer588 Can you expand the TeaCache node below the sampler and see if one the values is NaN?
@Hearmeman Endstep says NaN
Edit: sorry, no that is the V2V version.
I2V says endstep -1
I'm past all the errors now, but unfortunately it was never meant to be. It stops during generation with an OOM (RTX-3090).
Solved the OOM with quantization.
The results are really nice and worth the effort.