SRPO by Tencent: In the world of AI art, a groundbreaking new model, Direct-Align, is changing the game by teaching diffusion models to paint with human-like flair, while sidestepping two major creative roadblocks. Instead of the usual slow and expensive process of painstaking, step-by-step corrections, Direct-Align leaps ahead with a clever shortcut, using a predefined noise prior to instantly "interpolate" stunning visuals from any point in the creative process. Even more revolutionary is its ability to learn on the fly. By introducing Semantic Relative Preference Optimization (SRPO), the model can listen to text-based feedback - like a master artist adjusting to a client's whims - and make real-time changes to its style. This eliminates the need for endless, repetitive training sessions, making it remarkably efficient. The results speak for themselves: in a dazzling display, Direct-Align fine-tuned the FLUX.1.dev model, boosting its realism and aesthetic appeal by over three times.
👇
Resources:
1) Model GGUF: https://huggingface.co/befox/SRPO-GGUF/tree/main

or the refined version: https://civarchive.com/models/1953067?modelVersionId=2210446
2) VAE GGUF: https://huggingface.co/calcuis/pig-vae/blob/main/pig_flux_vae_fp32-f16.gguf
3) Flan GGUF: https://huggingface.co/silveroxides/flan-t5-xxl-encoder-only-GGUF/tree/main
3) UltraSharp v2: https://huggingface.co/Kim2091/UltraSharpV2/tree/main
4) for NSFW (here is what you can do: https://civarchive.com/posts/22174353) use this lora: https://civarchive.com/models/1295758/nsfw-fluxorwan-22orqwen-mystic-xxx?modelVersionId=2009929 or this one: https://civarchive.com/models/754919/photorealistic-nude ... or any other one you like ;)