The beta7 version feels like a technological setback,
The physiological structure of women is incorrect during exercise, and there are also problems with sound control, resulting in uncontrolled noise. However, the color of the image has returned to normal, basically maintaining the original color. The color of the beta6 image will change, losing saturation and turning into a dark gray image. However, the female physiological structure of beta6 is correct during exercise, which is more accurate than beta5. Every upgraded version has problems, making people feel unsatisfied,
Don't use developer H3 style prompting, it causes the most interference. "overall_soundcape" especially is actually the easiest way to introduce extremely bad audio and will cause the model to go out of it's way and start putting cartoon sounds on the output and ignore prompt and should be deleted. Prompt very normal audio queues in prompt or at the end of the main prompt.
I find there to be a lot of noise in the background of every gen that kind of animates its way through the entire video.
beta5 still the GOAT. More consistent ref image transfer.
(I have only tried bf16 beta7)
i just tested beta6 again and it has no noise like beta7 did. so either it is something missing in my workflow, because I am using wan2gp. but beta5 and beta6 at least work well.
Don't use developer H3 style prompting, it causes the most interference. "overall_soundcape" especially is actually the easiest way to introduce extremely bad audio and will cause the model to go out of it's way and start putting cartoon sounds on the output and ignore prompt and should be deleted. Prompt very normal audio queues in prompt or at the end of the main prompt.
The color scheme of the reference image in the beta7 version is basically correct, but the physiological structure of women during exercise is incorrect, and the physical effects that appear after the body is subjected to force are more realistic than the previous version. However, the appearance of the physiological structure is incorrect, which affects the viewing effect. The turbo color of beta6 is very light, the saturation is very low, and the image is very hard. I quantified the BF16 model of beta6 as INT8 and it is still the same, indicating that it is not a quantification error. The fastH3 saturation of beta6 is still good, but the image is hard and greasy, and it can be used together with the lora of DMAD and PDMD. Moreover, you did not provide the corresponding ideal lora for use,
I find there to be a lot of noise in the background of every gen that kind of animates its way through the entire video.
beta5 still the GOAT. More consistent ref image transfer.(I have only tried bf16 beta7)
The sound generation of beta7 is indeed not good, and there may be inexplicable errors. Both beta6 and 5 are correct,
for audio issues, strictly use only video step at 12 and audio at 3 or 8 , i dont have any audio issue with it with super simple workflow euler/simple at 4step only run global comfy kitchen nothing else
Beat6完全是灾难，而baet7解决了很多问题，是绝对正向的提升，尤其是人体解剖上，只是相对的，画面感觉不够干净，且加入loar后更油腻了，甚至一些以前可以用的loar直接无法用，视频会变成频闪。希望后续beat8可以解决loar兼容问题
beta5 is still the best so far, in my opinion.
my main problem with beta7 is still the way it handles references. character facial likeness is always somewhat off. lighting is also off.
e.g. beta5 puts characters in the provided environment with correctly preserved lighting from the reference image. it is almost as good as stock H3 in that regard.
beta7 changes the lighting to some over-brightened amateur flash-always-on style.
as for characters, using this simple description:
subject_definitions:
<Subject 1> is character, whose physical appearance and attire come from <Picture 1>.
beta5 would extract that character one for one most of the time with well-preserved facial likeness, also preserving character's outfit, hairstyle, etc. again, similar to base model.
beta7 would rarely extract the character at all. i would have to give it extra anchors, for example, by describing character's outfit in the prompt. yet still the character's facial likeness would be off (EVEN with the character LoRA!)
however, beta7 is definitely way better compared to beta6. beta6 was completely unusable for me.
I thoroughly tested beta6's fastH3 and found that almost all of the Lora generation was incorrect. The video within 10 seconds was barely okay, but after more than 10 seconds, serious confusion and errors would occur. This model is almost unusable,
Beat6完全是灾难，而baet7解决了很多问题，是绝对正向的提升，尤其是人体解剖上，只是相对的，画面感觉不够干净，且加入loar后更油腻了，甚至一些以前可以用的loar直接无法用，视频会变成频闪。希望后续beat8可以解决loar兼容问题
beta7的女性在做运动的过程中, 生理解剖结构会出现问题, 计算开始就会发生了变化, 完全失去了正常的外观结构 , 感觉在视觉上就没有太多价值, beta6的女性生理结构在所有版本测试中是最准确的 , 但是turbo版本画面非常生硬, 并且饱和度被降低成死灰色 , 皮肤上还有很多杂质 , 每个版本都有错误问题, 令人很无奈 , beta6和7 , 作者只释放了turbo版本, 其他版本并没有共享
Don't use developer H3 style prompting, it causes the most interference. "overall_soundcape" especially is actually the easiest way to introduce extremely bad audio and will cause the model to go out of it's way and start putting cartoon sounds on the output and ignore prompt and should be deleted. Prompt very normal audio queues in prompt or at the end of the main prompt.
Any recommendations as to how the prompting should look like then? Just natural, prose language? Description of action after description of action?