Comparison of Skyreels_hunyuan_i2v_native with 4 others. (1)
[Extra Edition] Interrupting the series to return to chronological article publication.
On Saturday, 2025/03/01, I performed video generation using Skyreels_hunyuan_i2v_native and compared its expressive power with the 4 others I have been using.
Since I have already performed a 4-way comparison with pyramidflow-miniflux (384p), LTXV i2v Lightricks, CogStudio + Dain-APP, and KlingAI, I will make this a 5-way comparison by using the same original images to generate videos with Skyreels_hunyuan_i2v_native. Please refer to the series of articles starting from the URL below for the other 4.
Each video is generated by providing the prompt used during the original image generation.
(The prompts have already been published in past articles.)
The cfg values for the two-stage custom sampler are 12.0 for the first stage and 1.0 for the second stage.
I will evaluate the pros and cons sequentially by genre.
First is the "Space Station Image."
It does not rotate, but the image quality is not bad.
Even adding 'rotating' to the prompt did not make it rotate.
The movement is subtle, but there is no perspective collapse or detail degradation like in LTXV i2v Lightricks, and it does not explode like CogStudio + Dain-APP (32 fps). I was surprised that the ground was filled in at the bottom, making it no longer a flying object.
The movement is smooth and dynamic, but although not as much as LTXV i2v Lightricks, there is an impression that the details become coarser over time.
For the space station image, I would say:
KlingAI ≧ Skyreels_hunyuan_i2v_native ≧ CogStudio + Dain-APP (32 fps) = pyramidflow-miniflux (384p) > LTXV i2v Lightricks.
Next is the "Alien Frontier Base Image." This is quite dynamic.
It is at the highest level for video generation from this original image.
This one is the same.
For the alien frontier base image, I would say:
KlingAI = Skyreels_hunyuan_i2v_native ≧ CogStudio + Dain-APP (32 fps) = pyramidflow-miniflux (384p) > LTXV i2v Lightricks.
Next is the "Cyber Female Soldier Image with a Background of Machines at an Alien Frontier," where flaws are easiest to detect.
Are the prompt's 'dark and foggy environment, heavy fog environment' a bit too effective?
The strange flickering of the image is concerning. The facial expressions have also deviated from the atmosphere of the original image.
This one does not seem to have as much flickering.
In both cases, I was concerned that the facial expressions changed slightly from the original image. This is important, but it is definitely a flaw that cannot be detected in original images of other themes. I feel that the root cause is the same as the phenomenon where the face turns into someone else halfway through, rather than a sense of collapse.
For the cyber female soldier image with a background of machines at an alien frontier, I would say:
KlingAI = CogStudio + Dain-APP (32 fps) ≧ Skyreels_hunyuan_i2v_native = LTXV i2v Lightricks >> pyramidflow-miniflux (384p).
I feel it has more reliable potential than LTXV i2v Lightricks and pyramidflow-miniflux (384p). There were also some instances where it exceeded the expression of CogStudio + Dain-APP (32 fps).
If I encounter original images where video generation requests in KlingAI are rejected, I feel that Skyreels_hunyuan_i2v_native can meet the goal of having another reliable video generation tool for local video generation besides CogStudio + Dain-APP (32 fps).
However, after this, I was struck by a strange phenomenon.
That will be in the next article.
Thank you for watching.
(Written on 2025/03/01)