Type | Workflows |
Stats | 1,016 |
Reviews | (81) |
Published | May 22, 2024 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 75DC2A1640 |
Easy SDXL Animation with HotshotXL
Simple workflow for txt2vid, img2vid or vid2vid. The video input can be swapped for an image input or bypassed for text only. if you have enough vram you can replace a controlnet with QR Monster, if not you can incorporate concepts from my FAQR workflow. You could combine this with my other workflows for more advanced features.
The real breakthrough I want to share with this workflow comes from the custom sampler section:
AlignYourSteps seems to add smoothing, and PertubedAttentionGuidance seems to add composition consistency (admittedly my testing is limited, if you find that this isn't exactly correct please let me know!). The slower samplers like dpm don't seem to offer enough detail over faster ones like euler to justify the 2x time to render.
I have also abandoned the lightning checkpoints at this time. The increased speed is nice for playing around but the quality loss is making them less viable, They tend to be especially bad when combined with AlignYourSteps.
Tested with 16gb vram. You can maybe get away with 12gb if you lower frames or resolution.
If you benefit from this workflow please consider sharing videos on this post. I would love to see what you can do with more vram than me.