Search Results
7/14/2025, 4:39:01 AM
>>715405257
It's normal to run batches with a varying seed or prompt to compare what works. Each of those took maybe 5 seconds to generate even on my old card.
>>715405694
>People just work around AI fucking them up, by adding direct reference images or rotoscoped motion cap footage so the AI never has to generate hands at all, it only adds a filter on top of original manually drawn or photographed hands
You don't know what you're talking about. To fix hands you add a controlnet, which is a neural network that gives spatial conditioning control over the base model. Controlnets came from a 2023 paper and are now a pretty standard part of the suit of tools used when generating good images. A controlnet trained on pose estimation or depth estimation can be used to make sure the diffusion model produces exactly the hands you want. No direct reference images or rotoscoping involved. It's a neural network. Training is learning. It literally learned to do hands.
All the good proprietary image generators like dalle or midjourney or whatever are doing this kind of thing under the hood too. Your prompt is only part of the input.
It's normal to run batches with a varying seed or prompt to compare what works. Each of those took maybe 5 seconds to generate even on my old card.
>>715405694
>People just work around AI fucking them up, by adding direct reference images or rotoscoped motion cap footage so the AI never has to generate hands at all, it only adds a filter on top of original manually drawn or photographed hands
You don't know what you're talking about. To fix hands you add a controlnet, which is a neural network that gives spatial conditioning control over the base model. Controlnets came from a 2023 paper and are now a pretty standard part of the suit of tools used when generating good images. A controlnet trained on pose estimation or depth estimation can be used to make sure the diffusion model produces exactly the hands you want. No direct reference images or rotoscoping involved. It's a neural network. Training is learning. It literally learned to do hands.
All the good proprietary image generators like dalle or midjourney or whatever are doing this kind of thing under the hood too. Your prompt is only part of the input.
Page 1