This workflow is my take on organizing and building an I2V model running against a "smaller" graphics card (specifically 12 GB). I use Q5 GGUF models, though I have also tested against smaller Q4 and Q3 - each having a slight decrease to prompt adherence and quality, but still usable. This includes using the lightspeed lora and notes on where all models / loras can be downloaded from.
I suggest if you have limited RAM or VRAM, to run ComfyUI with the parameter: --cache-none
While this will mean multiple batches against the same video will be slower, you get a much more consistent overall generation speed of your videos (3-4 minutes for 5-6 seconds on moderate home PC configurations).
This also includes using Florence2 (LLM) for image detection and auto-prompt assistance. You only need to add your action to the manual prompt (if desired).
There are a lot of nodes I have seen in various workflows ... but in Wan 2.2 I2V, at least, they tend to have no effect and only increase overhead.
I run my videos typically in 480p @ 480 x 832, and this workflow then upscales by 2x to 960 x 1664.
Custom Nodes Used:
ComfyUI-GGUF (https://github.com/city96/ComfyUI-GGUF)
rgthree-comfy (https://github.com/rgthree/rgthree-comfy)
ComfyUI-KJNodes (https://github.com/kijai/ComfyUI-KJNodes)
ComfyUI-Florence2 (https://github.com/kijai/ComfyUI-Florence2)
ComfyUI-VideoHelperSuite (https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
WAS Node Suite (https://github.com/WASasquatch/was-node-suite-comfyui)
Description
Version 1.2 Notes:
Cleaned up node layouts:
General cleanliness
Organized prompts and settings (frame, width, height) to one section
Minor changes to custom nodes
Added new (improved) node to show final prompt (LLM + Manual, if you leave the LLM enabled)
FAQ
Comments (7)
There is no advantage to using tiny braindamaged versions of models UNLESS you have silly small amounts of system RAM. With 64GB of RAM using better quants or FP8 has no speed downside (your bottleneck is unlikely to be the time taken to stream the model across the PCIe bus per iteration unless you have a very bad GPU with some prehistoric PCIe interface on a portable- and then video generation is unlikely to be an option fullstop.
Small quants are really only for people with 32GB of system RAM or less., and no-one into this hobby should have such a silly amount of RAM, given how cheap it is.
This is an interesting comment. The workflow works - exactly as advertised - for those with less availability to disposable income. But ... I guess I don't know what point you're trying to make by commenting like this?
Not everyone has systems up to your implied demands, your comment smacks of silly elitism and is ultimately little more than noise in the context given.
16 to 32 ram is actually what most people have. 32 gb is definitely not a silly small amount of vram. Its a perfectly fine and normal amount. even above the average. smaller models work fine too. there is almost no degredation compared to fp8. so basically you are talking out of your ass with your whole comment
When I load this workflow, it reports an exception. May I ask what is going on? the info:
Some Nodes Are Missing
When loading the graph, the following node types were not found
Int To Stringin subgraph 'File Names'
when i run it , the info: Cannot execute because a node is missing the class_type property.: Node ID '#230:222'
The node you are missing is:
comfyui-node-int-to-string-convertor
You can clone or download this node (put it in your ComfyUI/custom_nodes):
https://github.com/IvanRybakov/comfyui-node-int-to-string-convertor
It is a simple int to string converter for the file name, in case you want to include the seed in your file name. Alternatively, if you never plan to include the seed, you can adjust to exclude this node. I have personally found the int-to-string to be fairly useful for other projects, so I use it more often than I thought I would.