Endless Wan 2.2 I2V (SVI 2 Pro)
A simple workflow to create Wan 2.2 videos of unlimited duration, using SVI 2.0 Pro.
The workflow has a 5 sec "Initial" block and 8 more optional "Extend" blocks of 5 sec each that can create almost 45 sec of video (some frames are lost in the connection).
The video generation can starts either from an initial image, or from an already existing video.
The Initial Image block, has the "Start Frame and End Frame" image-to-video feature, that allows you to also use an "End" image, to guide the generation from beginning to end. After this Initial block, the other blocks just Extend the video.
If more seconds than the ~45 provided are needed, you can copy an "Extend" block, connect it with the others and continue..
Every block has its own Prompt selector and Length control in seconds (don't use more than 5.0).
Every block has a fixed noise seed number, that lets you experiment with that block without re-generate all the previous, already generated blocks. You generate the video until that block, and if you're satisfied and need more time, you enable the next one. After that, only the next one will be generated (if you don't change something in the previous blocks or the LoRAs).
Every block has its own independent LoRA section in addition to the Main LoRA section.
Select between
GGUF loadersfor low VRAM systems orSafetensors loaders(didn't test the safetensors, but they should work).Accelerated Generation: Supports deeply optimized, distilled LoRAs (like Wan-Lightning) that generate high-quality video in as few as 4 steps using lightx2v 4-step LoRA.
Warning: The LoRAs already loaded in the Main LoRA section are mandatory (for 4-steps & Linked blocks), except for the
Wan2.1_I2V_14B_FusionX_LoRAthat is there to speed up the movements. If you don't need extra speed you can turn its value lower or turn it off entirely.Warning: If the workflow in your system does not look like the screenshot I provide, that means that you are using a more current, but unfortunately broken version of comfyui-frontend.. (You can search google for the subgraph issues with the 1.4x.xx releases of their frontend). The last frontend version, that the subgraphs were working OK for me, was 1.39.2. To install this version, you must do
pip install comfyui-frontend-package==1.39.2in your..\venv\Scripts\folder. After that you will see a warning once, but other than that, everything will work fine..
Version 3.5
Added "End Frame" Image option (for the Initial Section only).
In the "Start from Video" section, you can now select a video from the Input directory, or Upload one (copy to Input directory), with the "Choose video to upload" button.
Version 3.0
Added independent LoRA per (5sec) video section.
Removed Extra LoRA 1/2 sections.
Version 2.5.1
Added the option to extend already existing videos.
Removed some leftover Crystools nodes so, no more compatibility problems with the RTX 50xx cards.
Tried to fix the "missing prompts" problem.
Version 2.1
Added another extra LoRA section to select from, in every 5 sec block.
Speed additions to counteract the slow-motion effect a little:
Changed the
HIGH_lightx2v_4step_lora_260412with theHIGH_lightx2v_4step_lora_v1030because it has more coarse movements. You can change the strength from 1.0 to 1.5.Added the
Wan2.1_I2V_14B_FusionX_LoRA(to the high noise path only), that gives additional speed in the movements. Use a strength of 2.0 to 3.0. This LoRA was created for the Wan2.1 model but works fine with Wan2.2 too. It produces a lot of warnings in the console for missing keys. This is because Wan2.2 misses some Wan2.1 keys, but it is just a warning nothing more. The generation works fine. For those of you that want to fix this in the code of ComfyUI, you can rename thelogging.warning("lora key not loaded: {}".format(x))line in theComfyUI\comfy\lora.pyfile, tologging.debug("lora key not loaded: {}".format(x))(always backup your files before editing them, for safety).
Models used:
SVI_v2_PRO_Wan2.2-I2V-A14B_HIGH_lora_rank_128_fp16.safetensors
SVI_v2_PRO_Wan2.2-I2V-A14B_LOW_lora_rank_128_fp16.safetensors
Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_v1030_rank_64_bf16.safetensors
Wan_2_2_I2V_A14B_LOW_lightx2v_4step_lora_260412_rank_64_fp16.safetensors
Custom Nodes used:
Description
Initial release
FAQ
Comments (14)
Bernini at 24fps with svi 2 pro would be awesome.
Call me an idiot, but i cant find where to add the prompt.
Hi, you will put the first description in the "initial" block. That describes the first five seconds. For each additional "Extend" block, you will put the description at the bottom of those that you activate. So you control each blocks action.
"The workflow has a 5 sec "Initial" block and 8 more optional "Extend" blocks of 5 sec each that can create almost 45 sec of video (some frames are lost in the connection).
Every block has its own Prompt and Length control in seconds (don't use more than 5.0)."
@Cyphermodx - when you open the workflow none of the text boxes work. They're not.. active.. not there. If you then go into the sub-graph and "open" or "unhide" the text encoder node, and then go back to main workflow page, then the text box is there. I tested with several and I had to do this for each new segment I enabled. I have no clue why and only discovered it by accident.
@Symbiot78 I hate when workflow designers collapse major nodes that you need to modify form time to time or all the time. Such as upscale nodes, like, people's directories are different so we'll need to change that... don't hide them behind other nodes and pin everything in place. I just delete the workflow when I see a bunch of pinned nodes.
@Symbiot78 Apologies, I haven't used this workflow, I was just going by the image of the workflow and trying to help as I was checking out what was new for the day.
The workflow should look like in the screenshot.
There, you can see that every section has a prompt.
If it's not like the screenshot, that means that you are using a more current, but unfortunately broken version of comfyui-frontend..
You can search google for the subgraph issues with the 1.4x.xx releases of their frontend.
The last frontend version, that the subgraphs were working OK for me, was 1.39.2.
To install this version, you must do pip install comfyui-frontend-package==1.39.2 in your ..\venv\Scripts\ folder.
After that you will see a warning once, but other than that, everything will work fine..
@Symbiot78 There is a "Extend Video Duration" node that you can use to enable the blocks you need.
And I thought it was the crazy
@noembryo mins is grayed out in those boxes. the WF runs fine just cant add any prompts. i'll do some searching around. thx. @Cyphermodx i'll check this and see.
@Cyphermodx this has worked thanks.
Is there any way to speed up the motion? I'm getting super slow motion even with the sampling set to 9. I've tried a few different LoRA, changing the weights, nothing works. Thanks!
Nice Workflow. I like this concept. However I do have problems with the additional loras. I have so far tested it with two loras but I got a significant quality loss and even changing the character. Without loras, I didn't run into this. It would be nice if it was possible to assign the loras sperately to each extention.
It would also be nice to manually define the video size to get out the maximum of my VRAM and adapt better to the Image.
Ready for V2?
I can't comment for the effect that a LoRA can have on the video (depends on the actual LoRA), but you can enable the Extra LoRAs independently for each section.
And you can definitely define the video size manually at the "Auto Rotate" node, in the "Loaders" section of the workflow..
What would you think a v2 should add to the mix?