bghira / flux-dreambooth-lora-r16-dev

flux-dreambooth-lora

This is a LoRA derived from black-forest-labs/FLUX.1-dev .

Two subjects were trained in, a character named Julia (AI) and a real person named River Phoenix.

Empirically, training two subjects in simultaneously kept the model from collapsing, though they don't train evenly - River Phoenix took longer than "Julia", possibly due to the synthetic nature of the data.

The photos of "Julia" came from Flux Pro. River Phoenix images were pulled from Google Image Search, with a focus on high resolution, high quality samples.

No captions were used during training, only instance prompts julia and river phoenix .

The main validation prompt used during training was:

julie, in photograph style

Validation settings

• CFG: 3.0
• CFG Rescale: 0.0
• Steps: 28
• Sampler: None
• Seed: 420420420
• Resolution: 1024x1024

Note: The validation settings are not necessarily the same as the training settings .

You can find some example images in the following gallery:

The text encoder was not trained. You may reuse the base model text encoder for inference.

Training settings

• Training epochs: 108
• Training steps: 2500
• Learning rate: 0.0001
• Effective batch size: 2
  • Micro-batch size: 1
  • Gradient accumulation steps: 2
  • Number of GPUs: 1
• Prediction type: flow-matching
• Rescaled betas zero SNR: False
• Optimizer: AdamW, stochastic bf16
• Precision: Pure BF16
• Xformers: Not used
• LoRA Rank: 16
• LoRA Alpha: 16.0
• LoRA Dropout: 0.1
• LoRA initialisation style: default

Datasets

julia

• Repeats: 0
• Total number of images: 34
• Total number of aspect buckets: 1
• Resolution: 512 px
• Cropped: True
• Crop style: random
• Crop aspect: square

riverphoenix

• Repeats: 0
• Total number of images: 12
• Total number of aspect buckets: 1
• Resolution: 512 px
• Cropped: True
• Crop style: random
• Crop aspect: square

Inference

import torch
from diffusers import DiffusionPipeline

model_id = 'black-forest-labs/FLUX.1-dev'
adapter_id = 'ptx0/flux-dreambooth-lora-r16-dev'
pipeline = DiffusionPipeline.from_pretrained(model_id)
pipeline.load_lora_weights(adapter_id)

prompt = "julie, in photograph style"

pipeline.to('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu')
image = pipeline(
    prompt=prompt,
    num_inference_steps=28,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(1641421826),
    width=1024,
    height=1024,
    guidance_scale=3.0,
).images[0]
image.save("output.png", format="PNG")

SimpleTuner Config

The configuration used to train this model:

export MODEL_TYPE='lora'
export TRAINING_SEED=420420420

export CHECKPOINTING_STEPS=500
export CHECKPOINTING_LIMIT=10

export LEARNING_RATE=1e-4

export FLUX=true
export MODEL_NAME="black-forest-labs/FLUX.1-dev"

export VALIDATION_SEED=420420420
export VALIDATION_PROMPT="julie, in photograph style"
export VALIDATION_NEGATIVE_PROMPT="blurry, cropped, ugly"

# How frequently we will save and run a pipeline for validations.
export VALIDATION_STEPS=500

# Validation image settings.
export VALIDATION_GUIDANCE=3.0
export VALIDATION_GUIDANCE_REAL=3.0
export VALIDATION_NUM_INFERENCE_STEPS=28
export VALIDATION_GUIDANCE_RESCALE=0
export VALIDATION_RESOLUTION=1024x1024

export ALLOW_TF32=true
export PURE_BF16=true

export CAPTION_DROPOUT_PROBABILITY=0

export MAX_NUM_STEPS=0
export NUM_EPOCHS=1000

export OPTIMIZER="adamw_bf16"
export LR_SCHEDULE="constant"
export LR_WARMUP_STEPS=500

export TRAIN_BATCH_SIZE=1

export RESOLUTION=512
export RESOLUTION_TYPE=pixel

export GRADIENT_ACCUMULATION_STEPS=2
export MIXED_PRECISION="bf16"
export TRAINING_DYNAMO_BACKEND='inductor'

export USE_XFORMERS=false
export USE_GRADIENT_CHECKPOINTING=true
export VAE_BATCH_SIZE=8
export TRAINER_EXTRA_ARGS="--aspect_bucket_worker_count=48 --lora_rank=16 --lora_alpha=16 --max_grad_norm=1.0 --gradient_precision=fp32 --base_model_default_dtype=bf16 --lora_init_type=default --flux_lora_target=all+ffs --user_prompt_library=user_prompt_library.json  --webhook_config=webhooks.json --compress_disk_cache"

export USE_EMA=false