SDXL / SD3 Alternatives - Open Models

Open Models :

Kolors :

• https://civitai.com/models/566526/kolors?modelVersionId=631404

• https://huggingface.co/spaces/gokaygokay/Kolors

• LORA trained on Kolors : https://civitai.com/models/571029/kolors-cotton-doll-lora-trained

• merge it with pony : https://openart.ai/workflows/sanrostal/kolors-with-pony-instantid-refiner/oLRnrUdguITRZ6MWfhVE

It seems that Kolors use the same architecture as SDXL (ipadapter, lora)

Aura Flow :

• https://civitai.com/models/573326/auraflow?modelVersionId=639125

• https://huggingface.co/spaces/multimodalart/AuraFlow

PG2.5 :

• https://civitai.com/models/325263/playground-ais-playground-v25-1024px?modelVersionId=364967

• DataVoid's PG2.5 finetune : https://civitai.com/models/596963/prometheus

Pixart :

• https://github.com/PixArt-alpha/PixArt-sigma

• https://huggingface.co/spaces/PixArt-alpha/PixArt-Sigma

• 900m Sigma https://civitai.com/models/573014/900m-pixart-sigma?modelVersionId=638760

Hunyuan :

• https://huggingface.co/Tencent-Hunyuan/HunyuanDiT

• https://huggingface.co/spaces/Tencent-Hunyuan/HunyuanDiT

Lumina :

• https://huggingface.co/Alpha-VLLM/Lumina-T2I

• https://huggingface.co/spaces/Alpha-VLLM/Lumina-Next-T2I

Kandinski :

• https://civitai.com/models/529576/kandinsky-31?modelVersionId=592915

First I have no idea about the technical part, which model is the easy to finetune or not but i like to test models,

To me Kolors seems really good, it can do many styles and you get beautiful output

if kolors can be finetuned it could be a really good base !

Aura know to write but the styles are likited and the outputs are .... meh .. and i could only gen in square format ...

2.5 have beautifull results but lack prompt adherance

It seems people left out PG2.5 and Kandinsky for whatever reason ...

Still need to test more the rest, will update the article

if you know any other model please comment !