all
/
g
/
s
/
fur
/
art
/
ai
/
3dcg
/
f
/
rp
/
req
/
p2p
/
dis
/
lit
/
kaki
chat
GUROchan!
This list is WIP...
1
2
3
So use the classic menu
Link 4
Link 5
Link 6
Link 7
List 3
Link 8
Link 9
[
Settings
]
[
Home
] [
Catalog
] [
Search
] [
Thread list
] [
Stats
] [
Reports
] [
Watcher
] [
PMs
] [
Admin
]
Artificial Intelligence
[
Return
]
Report a post
Preview
File:
pool_bh_5s.png
(450 KB, 560x720)
ImgOps
Hide image
ImgOps
Hide image
▶
A little beheading with Wan 2.2... and finally, some better prompt adherence
galathien
◆0AdiozdanE
[
PM
]
2026/04/20
(Mon)
15:13:21
No.
18349
[
Report
]
AS
▶
The gory isn't as gory as the Grok stuff we've been seeing, but decent. And I could do all the video at home without fighting any content restriction filters.
https://files.catbox.moe/ll8bi9.mp4
This file has an embedded workflow, and that workflow contains the prompt for the first five of ten seconds.
A second five-second clip was generated using a still frame from the end of the first five seconds.
As this is merely Wan 2.2 there is no native audio, but I used ElevenLabs Sound Effects to generate a few bits of appropriate audio, then assembled the two video clips and the audio clips using DaVince Resolve.
Finally, I used ffmpeg to re-attach the workflow metadata from the first video clip to the final edited results:
ffmpeg -i pbh.mp4 -i pool_beheading_pt1.mp4 -map_metadata 1 -c copy pool_beheading.mp4
I suspect that the main reason for better prompt adherence is that this is the first time I've used a workflow that didn't use the Lightning LoRAs. It might be simply that I'm still not all that good at editing ComfyUI workflows, or creating my own, but I've tried to disable the the Lightning LoRAs before and just ended up with an unusable workflow.
Finding this workflow already assembled without the Lightning LoRAs, and cited as having improved prompt adherence in a Reddit discussion, I finally got good (albeit slow) results that I hadn't managed before.
Of course, the prompt adherence isn't perfect, and there's still a bit of random luck involved. My very first try gave me the weird result (which I never saw again) of the girl's hair getting completely and cleanly shaved off her entire head, instead of having her head removed.
For the second 5-second clip, I either got blood in the water as I had asked, or a reaction of surprise and horror from the girls in the background, but not both at once. (Maybe running more than three trials would have gotten me both, however.)
Here's the first prompt:
> A sharp, gleaming sword arcs through the air with deadly precision. The blade, honed to a razor's edge, swipes through the neck of the unsuspecting girl. The sword slices through her neck with a swift, fluid motion, parting flesh and bone as easily as cutting through butter. Blood sprays in an arc through the air. The head separates from the body, tumbling behind the girl's back, revealing bloody muscle and bone where the neck was sliced.
>
> The girl's headless body spasms then falls backward into the water with a splash, staining the water with blood.
>
> The entire sequence is captured in vivid detail, showcasing the brutal efficiency of the beheading.
And the prompt for the second five seconds:
> This girl has just been brutally beheaded.
>
> The girl's headless body spasms then falls backward into the water behind her. Camera framing remains fixed while body drops backwards out of view.
>
> As soon as the body drops out of frame, a gush of water splashes up into view from the body falling into the swimming pool. The water in the pool at the bottom of the frame begins to be stained red from blood.
>
> Remaining out of focus, the other girls in the background react with horror.
Post number
No.
18349
Board
Artificial Intelligence
Reason
Optional. Describe what's wrong with it.
Style: