OnePunchMonk101010/dptlab-klein-dora-subject5
028
dptlab-klein-dora-subject5
Post-trained with `dptlab` using the lora recipe with a dora adapter, on top of black-forest-labs/FLUX.2-klein-4B.
Trainable parameters: 14.5M.
Training config
{
"model_key": "flux2-klein-4b",
"recipe": "lora",
"dataset_path": "/root/data/syncd/subject-5",
"output_dir": "/root/outputs/klein-peft/dora/subject-5",
"resolution": 512,
"learning_rate": 0.0001,
"train_batch_size": 1,
"gradient_accumulation_steps": 4,
"max_train_steps": 500,
"lora_rank": 16,
"lora_alpha": 16,
"peft_method": "dora",
"max_grad_norm": 1.0,
"use_masks": true,
"mixed_precision": "bf16",
"seed": 42,
"checkpointing_steps": 500,
"validation_prompts": [],
"validation_steps": 500,
"extra": {
"sampling_steps_for_shift": 4,
"logit_mean": 0.0,
"logit_std": 1.0
}
}Benchmark
Split: heldout (6 prompts). Prompts describe settings no training image shows.
CLIP-T and DINO pull in opposite directions: an adapter that learned nothing scores well on the first, one that memorized its training shots scores well on the second. Read them together.
See RESULTS.md in the repo for all six methods and the confounds.
Usage
import torch
from diffusers import DiffusionPipeline
from huggingface_hub import snapshot_download
# `dora` is not a format `pipe.load_lora_weights()` can read -- it would
# log "no LoRA keys found" and silently serve the base model. Inject via peft.
from dptlab.training.peft_methods import load_peft_checkpoint # pip install dptlab
pipe = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16).to("cuda")
denoiser = pipe.unet if hasattr(pipe, "unet") else pipe.transformer
load_peft_checkpoint(pipe, denoiser, snapshot_download("OnePunchMonk101010/dptlab-klein-dora-subject5"), "dora", for_training=False)
image = pipe(prompt="your prompt here", num_inference_steps=4, guidance_scale=1.0).images[0]