NerdyRodent

SkyReels-V1 - Text & Image To Video

Added 2025-02-22 17:18:57 +0000 UTC

Video - https://youtu.be/SIVA8fPfPU4

Overview

Even more text-to-video and image-to-video fun this week, thanks to SkyReels-V1 models from SkyworkAI!

Thanks to Kijai, there are smaller models available so you won't need 24+GB VRAM to run these, though it will certainly help! Even the Q4 is pretty decent, which weighs in at just under 8GB.

Personally, I prefer EulerCFG++ (alt) as the sampler, though your mileage may vary.

Prompting should always start "FPS-24, ", with the rest up to you.

Does it do what you tell it to? Sometimes, mostly. If you ask for a zoom, that is quite likely, but panning left or right seemed pretty random to me. Short or long prompts both seemed fine, and I too am still unsure about the real impact of negative prompts.

Longer sequences need more VRAM, but I generated up to 5s at a time with fairly reasonable resolutions, given it's a 544px960p model.

Image to video works fairly well, though things like dancing and hands may end up a bit weird :)

Links

Links for information included in the video :

Main SkyworkAI GitHub Repo: https://github.com/SkyworkAI/SkyReels-V1

Smaller Files: https://huggingface.co/Kijai/SkyReels-V1-Hunyuan_comfy

Extra Files: https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged

Extra things to try!

Not included is the video is this "Flow Edit" workflow from Zack Abrams -

https://x.com/ZackDAbrams/status/1892079837939949629

https://pastebin.com/ssc6TsyW

As well as the LTX ones manager finds, you'll need to install these custom nodes via the typical "git clone", as manager didn't find the nodes for me:

https://github.com/logtd/ComfyUI-HunyuanLoom

I've also attached a more readable version with colours and extra notes.

MS Windows + Sage Attention

For triton + sage attention on MS Windows, see this excellent guide from Purz:

https://purz.notion.site/Get-Windows-Triton-working-for-Mochi-6a0c055e21c84cfba7f1dd628e624e97

Workflows

Three workflows are attached, though only 2 are used in the video. The third one is a slightly modified version of the "Flow Edit" workflow from Zack Abrams. As I like colours and organisation, hopefully this version will make more sense to you as well :)