US$2.99
Attachment 1
Attachment 2
Attachment 3
Attachment 4
Yesterday, I saw the bundle and thought it was a discount system , where buying the bag early and then buying the bundle would result in a bigger discount. I was originally planning to put three of the taller ones in the bundle for Singles' Day, but after the discount, the third one is basically free ( 65-35-30=0 ). It's more cost-effective for existing users.
The result is a discount system ... Later buyers get a bigger discount, while existing customers get a smaller discount ( 35 + 30 + 20.53 = 85.53 ). Early buyers basically don't get any discounts, which is outrageous. The plan was disrupted, and the pricing is now awkwardly stuck.
So after calculating the costs when I got back tonight, I refunded everyone at once, lowering the price to (35 + 20 (15 pre-paid) + 20 (15 pre-paid)). Existing users' costs remain unchanged. The bundle feature is temporarily not being considered; please do not download duplicates already downloaded. The content remains unchanged.
There's no exaggeration in this analysis. The long V-shaped performance wasn't great; it performed better without LoRa but still crashed. It just wouldn't pick up, so we stopped and let the device recover.
A different approach to the long V-process, with the addition of NSFWlora, resulted in even worse performance and disjointed flow; it was actually inferior to the one in Kolomi... I can't seem to get the H3 NSFW performance right! And those prompts are truly awful; even with the skill-based agent, it still can't properly derive small-scale prompts. Modifying the process and prompts has caused me to lose hair and burn out my equipment.
If these issues can't be resolved soon, I might abandon the continuity and vocal reference capabilities and switch back to LTX for a longer V-shape, and try it for a few more days...
I've recently shifted to workflow setup and haven't had the energy to fine-tune models. My models don't have much room for improvement, and I can't tweak other architectures either. Therefore, I'll no longer use generated assets for iterations, but instead use existing assets for iterations (when necessary). This means I can use the low-loss WebP format to reduce file size. This loss is invisible to the naked eye and won't affect viewing. I temporarily turned it off because I was worried it would affect iterations, but the WebP format will be added after the existing inventory is organized.
Preview package: https://www.fansky.co/shoujikong/210
Preview 1: https://www.pixiv.net/artworks/151857472
Preview 2: Water Return