A heavy finetune with nearly 500k samples on strictly a smaller liminal dataset to refine BeatriXL into an artistic and tasteful liminal image generating monster.
This tool is far beyond simply generating depictions of interesting areas or sectors of an image; you have advanced control of your world in many shapes and forms. Albeit stubborn at times, this beast can be tamed; as it's compatible with most loras if you set the lora to low strength from most SDXL models.
You won't just get regurgitated art from a list with this one. It's generating unique cross-entropic contaminated sectors more often than not; and the outcomes reflect high complexity differences - beautiful differences if you want them to be.
Euler or Euler A usually works, but I advise using the RES4LYF sampler pack to really make this model's showcase powerful.
https://huggingface.co/AbstractPhil/Liminal-Full/tree/main/Beatrix-LORA
https://huggingface.co/AbstractPhil/Liminal-Full/tree/main/Beatrix-LORA-V2
You can pick the mix out yourself if you want. It's a few loras; where one was merged into the core and then the final lora epochs trained from that.
liminal, no humans, stuffThe dataset was trained with a multitude of timesteps for a long time with multiple different datasets. Careful though, BeatriXL is still fairly unstable - the refinement helps a whole lot, but if you aren't careful you'll see things you don't want to see. However, this version is a lot more SFW by default than the others - meaning it's more likely to default to a SFW topic than a NSFW topic, but there are no guarantees.
If you use humans you're probably going to see something you don't want to see, so be careful with those.
Roughly 5k different images; so not many. However, those 5k images were issued a series of different captions from plain English to full description.
Joycaption 2 + siglip, GPT4o, LLAMA2 LLAVA B-OMEGA + Siglip500, CLIP_L interrogation, CLIP_G interrogation, and pure raw tags from the WD14 systems.
The curriculum training went really well.
For the fp8_e4m3fn
if your ComfyUI gives you any major errors place these in your gpu launch bat;
--fp8_e4m3fn-unet --fp8_e4m3fn-text-enc --fp32-vaeIt's not a true fp8 utilization; as ComfyUI did not autocast to fp8, but instead cast to bf16 manually for me.
model weight dtype torch.float8_e4m3fn, manual cast: torch.float16Yet I definitely launched with float8_e4m3fn as the unet. I really don't know what to say about that, other than it's likely not supported by a 4090 so it'll only run on higher-end cards.
The text encoders run just fine on fp8.
Likely in this code here, as it's running "should_use_bf16", which probably forces it into bf16 mode. Until I have a proper workaround it'll have to be upscaled to bf16, but at least the size is tiny.
Description
If ComfyUI gives you any major errors; run it with these as your launch commands.
--bf16-unet
--bf16-text-enc
--fp32-vae
FAQ
Comments (17)
do you have one half the size?
Oh, yeah sorry I uploaded late. This is the fp32 version. I'll make the bf16 version and an fp8 version so you can all play with it.
Alright I uploaded a much smaller one at fp8 and it's been formatted to fp8_e4m3fn.
If you have any major errors run your comfyui with these args.
--fp8_e4m3fn-unet --fp8_e4m3fn-text-enc --fp32-vae
Actually what I want is the sdxl as a smaller size. I have space issues.
@frankmike The fp8 version is tiny. 3.6 gigs.
@AbstractPhila oh i mistook that for a flux model. My bad
Drop me a line after u upload the bf16 version. looks like my wildcards will love it!!
FP8 looks so much better than I thought. I've implemented a proper converter and it looks fantastic. I'll upload bf16 next.
Do you have a list of specific triggers?
Or are they not necessary?
for this particular checkpoint?
@Daemona I really don't know if they're needed yet or not. I pumped so many captions into it that I haven't really had a good area of breathing room. I posted some of the specific trained triggers in the fp32 version's trigger list, but there are many beyond that.
On top of that BeatriXL is a bit of an anomaly on the best of days. It can run epsilon loras and vpred loras from most models on low heat.
If it acts up, you can get some realism loras, grab some liminal loras, and just kind of frankenstein the exact style you want with a bit of fiddling. It knows so much plain English and the clips are so heavily fintuned that it'll make pretty much anything if given the attention shifts and the unet tweaks.
Which of the beatrix "real" loras do you reccommend from your HF?
They are object association and heavy realism revamps. Just get the newest one and grab the parent model beatrix merge that I'm training it with, it's in the parent repo folder.
Inference with unet at bf16, text encoders at fp32 and vae at fp32.
Also use these laion flavors for negatives; it's a very different model.
naked cartoon gross nasty intrusive, overly dry from incorrect context, anatomically incorrect bad anatomy, entirely completely full face image, extensive bloated indirect fallacy, translated text misdirection bias, uninterrupted liquids with flowing, shattered broken dilapidated uncomfortable, straight lines only direct, surgical scarring or excessive tissue damage, yellow tinting encompassing, jewelry wearing and encrusted, flat and perfect encompassing facial features,
@AbstractPhila Thanks! Do I need to set fp32 bs bf16 manually in comfy, or does it handle it automatically?
@civitcensersebryting847 Set it manually, it won't look as good without it manually set.
If you set the UNET to fp32 as well, you can see some really really cool effects and it yields a higher quality output too - but it takes longer than flux.
--bf16-unet --fp32-text-enc --fp32-vae
@AbstractPhila Is there a recommend workflow?
@AbstractPhila Any positive keywords whould be used for realism like with beatrix?
@null There's a definite workflow. I'll release it in an hour or two.