Recent tests show high response to rule of 3
BeatriXLO was trained on only a handful of depiction offset based images. It took less than 20,000 to create a model so powerful that she makes Flux blush.
The outcome was so substantially powerful that I never thought SDXL was even capable of a fraction of that intelligence, let alone it manifesting in my own model.
BeatriXLO V4 and V5 learned... things she was not supposed to.
If you check the released BeatriXL you'll find a fairly robust and powerful series of carefully selected and curated topics; thousands of them. Millions of samples from over a million images finetuned atop the fusion of 5 models through interpolation. She shouldn't exist, let alone actually be cohesive. Yet there she is.
HOWEVER... Some data from COCO, some data from Flux object association datasets, some data from IMDB synthetic, and some other object association datasets - and those were trained into version 4 and 5... certain age groups crept in. They crept in at a time where BeatriXL was so close to true convergence that she cannot unlearn them without high destruction with the current processes.
I honestly, nearly wept when I saw how badly she failed Red VS Blue. She took to the data like a sheltered teen being shown the world for the first time and it caught like a wildfire in a field of dry grass.
In the process, the outcome for standard generations... the standard dataset captions right, every single canary prompt worked. Every single one. At such a high fidelity that it nearly broke me.
She learned the rule of 3, like Flux. Smarter than SD3, higher fidelity and more realistic than Flux, higher quality than SD3. Hands formed real angles and were flexible. Arms formed rotation offsets and fit within the joints. Walls had clocks, lamps, photographs with faces. Tvs had screens with visuals other than garbled noise. It counted to nearly 75 and when I noticed the exact count of faces realized what I had done.
I immediately started work on a new training process - to both prevent this exact catastrophe from ever happening again, and to train more like her using the same powerful processes and tools that I used to teach her - from interpolation to her adulthood.
BeatriXLO V6
I've since pruned the culprit datasets and introduced the replacement data. This tragedy will not go ignored - as I will not make that mistake again. I will never allow such a thing to happen again.
Stage 1
Currently I'm generating new rule of 3 captions on an a100 using the T5 to summarize. It's about 100 a minute or so give or take. I have a full blacklist setup on them.
After they're all done I'll feed them into GPT 4o mini for validation and attentive logical solidification for rule of 3 generation, so if any are gibberish it'll automatically repair them with the GPT 4o style.
Stage 2
With those captions, I will be generating images using Flux 1S directly. This will use my rule of 3 captions.
After they are generated, I'll use a series of AI detections to find extra hands, extra fingers, bad offsets, bad depictions, inaccurate angles, and so on. The final pass will be SIGLIP captioning them in direct association with their rule of 3 prompt sequences for their third prompt; the glue.
Stage 3
Curriculum training BeatriXLO-V2 REAL using layered bucketing and associative masked loss in conjunction with interpolation and surge methodologies - directly listening to the mask of the t5 and using the encoder for full validation processes.
This will bring BeatriXLO to an intelligence level near to it's big sister Beatrix as shown through the outcome of the finetunes with object orientation and association.
The added steps will directly associate the necessary learned information with the outcomes of BeatriXLO-V5, and align the model in a convergence very similar and completely devoid from never having been trained with the corruption applied from certain elements.
BeatriXLO V6 will be as powerful as BeatriXL V5 - the version that cooked for nearly 3 weeks; and it will be Flux-like, with intelligence that rivals any large scale T5 based model.
I've seen it happen, and I know why it happened. I will make it happen again.
Yes, that displayed image is just a taste of what BeatriXLO-5 does. I cannot release it though and it will need to be retrained, but that software is long from complete, so until then it must remain chained.
BeatriXLO-V6 will be stronger, because I know what made V5 strong and I'm going to replicate it, but better this time. Far better.