Comparison of Skyreels_hunyuan_i2v_native with 4 others. (2): A strange phenomenon occurs....
Continuing.
I will be comparing this with the 4-way comparison from this article (pyramidflow-miniflux (384p), LTXV i2v Lightricks, CogStudio + Dain-APP, KlingAI).
Although the video generation by Skyreels_hunyuan_i2v_native got off to a smooth start,
it was struck by a strange phenomenon after this.
It started with this video.
It is the "Cyber female officer image with a fantasy painting 'The most amazing scene ever created' as the background" image.
is.
outdoor, medium shot, wearing gray and black hard metal-suits, (very short bob cut hair), wide hip, large breasts, perfect slender young girl body, perfectly glossy skin ,25 yo mature girl, picture of a vivid, masterpiece, perfect lighting, asian beauty, The most amazing scene ever created , Fantasy art , Sci-Fi themes , Technological utopia , Another Planet , Timeless dimensions , Energy radiating everywhere , a Masterpiece,
I am reposting the prompt, but it does not include 'smiles'.
This is an 'unintended expression' that cannot be compared to the slight distortion in the previous article.
After this, a series of videos followed where the content changed abruptly from the start of the video.
The original image is visible only for the first instant.
I have arranged the videos in the same order as the URL article mentioned above. After seeing a glimpse of the original image,
videos that are completely different from the original image appear.
This can no longer be called i2v.
It is as if the HunyuanVideo engine of Skyreels_hunyuan_i2v_native is overriding the process by reading the prompt and performing t2v processing instead.
So, I tried leaving the prompt blank during video generation.
The image quality is rough and grainy, but the nuance of the original image has returned.
What if I remove the fantasy painting 'The most amazing scene ever created' background part from the prompt and only use the first half?
outdoor, medium shot, wearing gray and black hard metal-suits, (very short bob cut hair), wide hip, large breasts, perfect slender young girl body, perfectly glossy skin ,25 yo mature girl, picture of a vivid, masterpiece, perfect lighting, asian beauty,
It doesn't seem to work. In the most recent article, the background was different, but this part was a common prompt, and although there was some slight deformation, there was no discrepancy with the original image to this extent.
was.
I restarted the PC and also restarted ComfyUI, but this phenomenon did not disappear.
was.
From then on, I generated videos without a prompt.
I will backtrack a little from where things went wrong as mentioned above.
The background doesn't move even though they are walking....(^^;).
Even with robot images where I don't have to describe the details of facial expressions, the flaws are somewhat noticeable.
For the "Cyber female officer image with a fantasy painting 'The most amazing scene ever created' as the background"
image,
KlingAI ≧ CogStudio + Dain-APP (32 fps) ≧ Skyreels_hunyuan_i2v_native > LTXV i2v Lightricks > pyramidflow-miniflux (384p)
is probably about right.
Seeing that the inexplicable behavior could be avoided by generating videos without a prompt, I still maintain the feeling that Skyreels_hunyuan_i2v_native could meet the goal of having another reliable video generation tool for local video generation, besides CogStudio + Dain-APP (32 fps).
Will a similar mystery occur with the "Magic Array in Spaceship" image?
I will check in the next article.
It seems necessary to first test whether to provide a prompt when generating the original image or to use no prompt for each genre when generating videos with Skyreels_hunyuan_i2v_native.
Thank you for watching.
(Written on 2025/03/01)