--- v3 ---
Improved text handling (still takes a few generations to get it... SDXL... sigh)
Particle improvements (smoke, fire, etc...)
Better animal generations
More Prompt Flexibility with better adherence to age.
--- v2.3 ---
Significant Improvements to Prompt Adherence.
Improved Prompt Flexibility
Better reflections, glare, smoke.
Better multi-character variations.
SPECIAL NOTE: V1.3 has enhanced text generation capabilities. (not perfect, but not bad). v2.3 still has some text abilities but is not as refined as v1.3. I will be releasing a realistic XL base model soon (QuadPipeXL) that will have enhanced text capabilities.
*As always, let me know if you find any bugs and post those photos so I can see how creative you've all been! Thanks everyone!
--- v1.3 ---
I've been asked by a few people why I hadn't released an SDXL model yet. Well it was because I was working on it. This is definitely a version 1. Please notice the sample pictures though, text + sports. Text & Sports are HARD! So this has taken a long time to get to a point where I was comfortable with it.
Flux is still way better at text, but this isn't half bad without LoRAs if you find the sweet spot on prompt and CFG scale.
Anyway, I didn't want to release just another SDXL model... I wanted to move the ball forward a bit. So a few things to keep in mind. The LAION data-tags are relatively preserved. That means that it kind of knows who a lot of people are already, and you won't need as many LoRAs to render notable people and characters.
It does get confused on a few things here and there.
For example, Magic Johnson was number 32 for the Lakers and number 33 for his college team... sometimes it puts the 33 on his Laker uniform. Same for most athletes. This model will try to do text when it can, Mike Tyson is going to have "Tyson" written on his belt etc...
If you don't want text, you're probably going to have to put a negative in. I'm still working on that bit.
As always, thank you for supporting me, and please let me know what you think and if there's something I missed here. I'll definitely be updating this model moving forward.
IMPORTANT NOTE - This model was NOT trained on celebrities, any knowledge it has of celebrities or people alive or dead comes directly from the core SDXL model and the LAION 5B Dataset (which does have a lot of knowledge preserved). All I did here was focus on overall enhancements to fine-tune rendering/lighting/textures/camera positions/objects/emotions, and I did a lot of work on text generation tuning.
Description
FAQ
Comments (7)
I was looking for a cheap and easy text model and all I needed the whole time was this big chungus!
It's not perfect on text, but I did a lot of work to make it render as best as I could.
@QuadPipe I tried some models, all sucked! This one actually obeys! And this model isn't FLUX or SD3!
With the pretty good text training it does require experimenting with a couple of LoRAs for a cooler output.
Still, thanks!
@ultraRarePepe_nr420 I'm going to continue to tinker on it. I'll be happy if I can get to reliable text 1/15 or 1/20 of the time. That's about the same reliability I get on a specific message with Flux. But yeah, prompt tweaking helps and I've found that some samplers do better on specific phrases than others. So I'd recommend trying that if you're "close" but it's off by a letter or two.
@QuadPipe v1.3 obeys a lot better for text. But v2.3 produces more jank - I laughed hard, soon will let others laugh too - and sometimes it like doesn't know where to put the text, so it doesn't bother at all. Unlike v1.3 which finds a place and even more. v2.3 also loves producing words on simple background.
But I understand text focus wasn't your intention in v2.3, so It's okay, I'm happy with v1.3 being quite capable.
@ultraRarePepe_nr420 It’s good you mention it and you’re right. I may have taken the 1.3 accuracy for granted when I was building my improvements… thinking that the text quality would persist. I’ll make sure to keep text firmly in the 3.x dev cycle.
V2 is pretty neat, well done!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.