Big improvement. Looks good-ish, nametags can be prompted (see example, it's a bit less easy. Works open, closed.... It's not perfect but it works dramatically better than the first one.
It seems a bit too fixated on older, slightly fade uniforms, maybe the changed dataset after shortleggate, or it is qwen2.1 We'll see. Generally 0.80 weight seems to do well while remaining flexible..
UPDATE 3300 step version- extended run from the old one with more images to fight the scruffiness. Doing old/new it per trigger word does not work (later), but it generally creates less scruffy ones. If you like scruffy, worn look use the other lora.
Select which version to load on the right under Download. STILL WiP all.
OLD desc, still fits. Change the existing nametags on the jacket to right: "US ARMY" , the left nametag to "PromptMe"; works fairly reliably (even without reference images). Unlike krea2 or boogu a simple left nametag "blahbla" does not work very well.
Top tip: Qwen image 2.1 is set in comyui's default template and even according to the og github with cfg1. But the model is NOT distilled, higher cfg (3 or so) massively improve the quality, at the cost of speed. (It's still fairly fast for how many steps it uses).
Description
Big improvement. Looks good, nametags can be prompted (see example, it's a bit less easy. Works open, closed.... It's not perfect but it works dramatically better than the first one.
Change the existing nametags on the jacket to right: "US ARMY" , the left nametag to "PromptMe"; works reliably (even without reference images). Unlike krea2 or boogu a simple left nametag "blahbla" does not work very well.
Top tip: Qwen image 2.1 is set in comyui's default template and even according to the og github with cfg1. But the model is NOT distilled, higher cfg (3 or so) massively improve the quality, at the cost of speed. (It's still fairly fast for how many steps it uses).