v17 Consistency
V17's character consistency compared to earlier versions is insanely good. Well done on putting together such an excellent blend.
which workflow are you using?
110% agree. Such an amazing update, truly a christmas gift from Phr00t to us 🙏🙏🙏. It's so good. I really can't believe we're living in such times where amazing work is being done by Phr00t and everyone else who has contributed... It's leading to us being able to run such an amazing image edit model locally....
which workflow are you using?
I'm in the process of building out a LORA source image generator - jibMix RealisticXL to generate an initial headshot, and then Qwen to take the headshot and generate the rest. I'm building out my own because I want body consistency as well as face consistency, and there are not many workflows I've seen that do that. Previous versions struggled to translate the SDXL-generated headshot accurately to Qwen. v16 and v17 are doing it really well. This workflow does a lot of looping and uses AI / wildcards / text replacement for the image gen so it's a bit of a mess. If I get it working well and manage to clean it up satisfactorily I'll share it.
very good work with V17, thank you Phr00t 😀
which workflow are you using?
I'm in the process of building out a LORA source image generator - jibMix RealisticXL to generate an initial headshot, and then Qwen to take the headshot and generate the rest. I'm building out my own because I want body consistency as well as face consistency, and there are not many workflows I've seen that do that. Previous versions struggled to translate the SDXL-generated headshot accurately to Qwen. v16 and v17 are doing it really well. This workflow does a lot of looping and uses AI / wildcards / text replacement for the image gen so it's a bit of a mess. If I get it working well and manage to clean it up satisfactorily I'll share it.
I have one that allows you to have 1 input images and it generates 16 based on that with definable prompts for each image. Body consistency is pretty decent as long as your body proportions are not out of control otherwise you need to add that specific to each prompt
The results look very good! But for me the workflow increased from ~140s to 1200s? Only thing I switched is the model as well as the recommended use of 'euler_ancestral'
// nevermind, was swap use
which workflow are you using?
I'm in the process of building out a LORA source image generator - jibMix RealisticXL to generate an initial headshot, and then Qwen to take the headshot and generate the rest. I'm building out my own because I want body consistency as well as face consistency, and there are not many workflows I've seen that do that. Previous versions struggled to translate the SDXL-generated headshot accurately to Qwen. v16 and v17 are doing it really well. This workflow does a lot of looping and uses AI / wildcards / text replacement for the image gen so it's a bit of a mess. If I get it working well and manage to clean it up satisfactorily I'll share it.
I have one that allows you to have 1 input images and it generates 16 based on that with definable prompts for each image. Body consistency is pretty decent as long as your body proportions are not out of control otherwise you need to add that specific to each prompt
...
i adjusted the multi angle with one click template in comfy, wich does essentally the same, but i added another prompt box that concatenates to every individual prompt so i can make additions that applies to all on top of the main individual prompt. Could be a neat addition to yours. It works with the multi angle lora and 2509, unless you overwhelm it with too much prompt stuff.