Alissonerdx/flux.1-dev-SRPO-LoRas

Flux.1-Dev SRPO LoRAs

These LoRAs were extracted from three sources:

• the original SRPO (Flux.1-Dev): tencent/SRPO
• community checkpoint: rockerBOO/flux.1-dev-SRPO
• community checkpoint (quantized/refined): wikeeyang/SRPO-Refine-Quantized-v1.0

They are designed to provide modular, lightweight adaptations you can mix with other LoRAs, reducing storage and enabling fast experimentation across ranks (8, 16, 32, 64, 128).

Notes:

• The Loras version for Nunchaku was converted using the official Nunchaku conversion tool but it is something experimental I still need to test and analyze the results, I do not recommend using it for now it is only for testing.
• These loras allow you to use the quality of SRPO using the official flux dev as a base, without the need to use the base flux SRPO, that is, in my opinion, it is not very advantageous to use any of these loras + flux SRPO as a base, unless you want to apply the quality of, for example, SRPO RockerBOO in the base flux SRPO model.
• The version I recommend is RockerBOO but I advise you to test the others, because the original version will give you different results than the other versions.
• According to some reports it seems to work well with Flux Krea, the report was with rank 256, I haven't tested it yet to confirm.

Example comparison between Flux1-Dev baseline and LoRA extractions

use with 🧨diffusers:

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)

pipe.load_lora_weights('Alissonerdx/flux.1-dev-SRPO-LoRas', weight_name='srpo_128_base_R%26Q_model_fp16.safetensors')
pipe.to("cuda")

prompt = "aiyouxiketang, a man in armor with a beard and a beard"

image = pipe(
    prompt, 
    num_inference_steps=28, 
    guidance_scale=5.0,
    generator=torch.Generator("cpu").manual_seed(0)
).images[0]