I’ve developed a workflow that allows the creation of 24-second videos in just ten minutes. The core principle is based on the Diffusion Force system using Kijai’s WanVideoWrapper nodes in combination with the Skyreel V2 DF model. The process starts with generating a high-quality image using Illustrious/SDXL (for better realism—it simply looks superior to Skyreel’s TXT2VID output). This image is then used as the initial frame fed into the generation pipeline.
A continuous scene of approximately 24 seconds is created using seven sub-prompts, each generating 97 frames. These sub-prompts can either be crafted manually or generated via an LLM. I’ve been using Ollama, but this part can easily be customized.
Acceleration is achieved through CausVid LoRA V2 with high strength, combined with a CFG step switcher: the first 3 steps run with CFG 4, followed by 7 steps at CFG 1. This keeps the rendering fast while maintaining fluid and natural movement. One caveat: the system doesn’t handle abrupt movement transitions well—something like a sprint followed by a sudden stop won’t work; the character will just keep running.
At the end, the seven segments are merged into a complete video, with overlapping frames removed to ensure continuity.
It’s a complex but reasonably tidy workflow, made manageable through the use of Anything Everywhere nodes and hidden connections. If you have any questions, feel free to ask.
One tip: LLMs aren’t perfect at prompt crafting. I’ve tested this extensively, and after several days of experimentation, I’ve gone back to writing them myself—the quality is simply better when I do it manually.
Description
Skyreel V2 DF!
FAQ
Comments (4)
The Idea is nice i gues and i'm sure that one can create nice stuff with this but....the Workflow is way way WAY to overloaded, i can barely move around in the workflow, totall lag/slowdown, my 3090 might not be the latest or the greatest, but if the workflow itself is already murdering my PC than thata a good sign that this is Overkill (CPU Ryzen 7 5700X3D and 128GB Ram so not really a weak PC)
I'd add to that that making blanket statements about generation time like this is not helpful since it heavily depends on what hardware you're using. I'd wager that the "average card" that most people are using is a 3060 12GB and that'll mean significantly longer gen times.
In general, it's a neat idea, but I doubt it is practical for satisfying results. I have quite a bit experience with stitching together short clips using segments and then the last frame as starting frame for the next part. In practice, you have to generate 5 to 10 segments until you get a good one and can move on to the next segment. Simply taking the first attempt for each segment will produce crap 90% of the time, I'd estimate.
you need to look at your pc closer, maybe programs in the back or comfyui problems.
my 3080 ti, amd 5 3600 and 64gb ram can handle this with ease.
What are the vram requirements to run this? Can it use gguf variants of the DF model?