Experimental NoobAI with Rectified Flow + EQ-VAE

A finetune using NoobAI v-pred 1.0 as base and converting it to Rectified Flow by default uses Anzhc's EQ-VAE-B7 this experiment struggles with dark scenes due to data still being out of distribution but shows that Rectified Flow is indeed the next step in SDXL's improvement efforts (beyond just EQ-VAE), this model somehow recovers v-pred from its innate blue-shift and blue bias, but still lacks improvements in the uncond, and needs fixing for Zero Out negatives which can be fixed in a proper finetune.

Usage Tips:

YOU'LL NEED TO USE COMFYUI/SWARMUI or any webui that can do ModelSamplingSD3 at 2.5

• Positive Prompt: masterpiece, best quality

• Negative Prompt:worst quality, low quality, worst aesthetic, old, early, blurry, lowres, signature, artist name, watermark, twitter username, sketch, logo, furry, text, speech bubble, censored, ai-generated, censorship, censor, mosaic censoror similars as needed

Technical Settings:

• Sampler: Euler, Euler CFG++ DO NOT USE ANCESTRAL SAMPLERS!

• Steps: 20-28

• Scheduler: Simple, Normal

• CFG Scale: 4.5-6.0

• CFG++ Scale: 1.0 - 2.0

• Resolution: 1024x1024 or any aspect ratio that falls within ~~1.048.576 pixels.

• ModelSamplingSD3: 2.5

Is this another of your shitty models bluvoll? In a way, yes, but this is meant to show that Rectified Flow SDXL is completely possible, NoobAI recovers to this easily if we can hit it with enough data, how much data? The full NoobAI dataset which I posses but can't use due to limited money, therefore if you believe this has future I humbly ask for donations so I can make it a reality how much does this need? 2 Epochs for full dataset, or 20 million steps, some 1600 to 1900 USD due to issues that will inevitably appear.

"Why should I trust you Bluvoll?" You shouldn't, we have a stain in the community called Unstable Diffusion and another being Resonance Cascade, therefore, if someone is willing to donate compute that would be perfect, if you want to donate either, I have a Ko-Fi or you can find me in NoobAI's server or Arc en Ciel's, please consider supporting this if you want a usable Rectified Flow, not like certain slop model that has a 7 on its version.

Training information:

• Used shift 2.5 for training.

• User Cosine Optimal Transport.

• Used Anzhc's B7 VAE

Future Improvements that can be done:

• Better VAE using more data.

• Finetuned Clip Text Encoders to recover Clip L and further improve the performance of BigG

• Improved training with contrastive flow matching and more newer practices.