FLUX.1-dev was distilled into an n-step model using AnyFlow (a MeanFlow-style forward stage followed by on-policy DMD) on the e621 furry split. Two deliberate choices shaped the run:

Why start from a distilled base? FLUX.1-dev is itself a distilled model. This run was a check for mistakes in the AnyFlow implementation — specifically, whether already-distilled models simply can't be AnyFlowed. They can.
Why e621? It was readily available via webshart in SimpleTuner.

The adapter was intentionally not trained long enough to resolve the remaining 2-step and 1-step generation difficulty: the objective above was validated, and training stopped there.

bghira/flux-anyflow-e621-dmd

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

The main validation prompt used during training was:

masterpiece, best quality, score_7, safe, anime portrait of a young woman with blue hair wearing a white jacket, clean line art, detailed eyes, soft daylight

Validation settings

CFG: 3.5
CFG Rescale: 0.0
Steps: 4
Sampler: AnyFlowValidationScheduler (FlowMatchEulerDiscreteScheduler)
Seed: 42
Resolution: 1024x1024
Skip-layer guidance:

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: 0
Training steps: 1500
Learning rate: 2e-06
- Learning rate schedule: constant_with_warmup
- Warmup steps: 250
Max grad value: 1.0
Effective batch size: 4
- Micro-batch size: 1
- Gradient accumulation steps: 1
- Number of GPUs: 4
Gradient checkpointing: True
Prediction type: flow_matching (extra parameters=['shift=3.0', 'flux_guidance_mode=constant', 'flux_guidance_value=3.5', 'flux_lora_target=all'])
Optimizer: adamw_bf16
Trainable parameter precision: Pure BF16
Base model precision: no_change
Caption dropout probability: 0.0%
LoRA Rank: 128
LoRA Alpha: 128.0
LoRA Dropout: 0.0
LoRA initialisation style: default
LoRA mode: Standard

Datasets

flux-e621-1024

Repeats: 0
Total number of images: 65536
Total number of aspect buckets: 1
Resolution: 1.048576 megapixels
Cropped: True
Crop style: random
Crop aspect: square
Used for regularisation data: No

Inference

python

import torch
from diffusers import DiffusionPipeline

model_id = 'black-forest-labs/flux.1-dev'
adapter_id = 'bghira/flux-anyflow-e621-dmd'
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # loading directly in bf16
pipeline.load_lora_weights(adapter_id)

prompt = "masterpiece, best quality, score_7, safe, anime portrait of a young woman with blue hair wearing a white jacket, clean line art, detailed eyes, soft daylight"


## Optional: quantise the model to save on vram.
## Note: The model was not quantised during training, so it is not necessary to quantise it during inference time.
#from optimum.quanto import quantize, freeze, qint8
#quantize(pipeline.transformer, weights=qint8)
#freeze(pipeline.transformer)
    
pipeline.to('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu') # the pipeline is already in its target precision level
model_output = pipeline(
    prompt=prompt,
    num_inference_steps=4,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(42),
    width=1024,
    height=1024,
    guidance_scale=3.5,
).images[0]

model_output.save("output.png", format="PNG")

Exponential Moving Average (EMA)

SimpleTuner generates a safetensors variant of the EMA weights and a pt file.

The safetensors file is intended to be used for inference, and the pt file is for continuing finetuning.

The EMA model may provide a more well-rounded result, but typically will feel undertrained compared to the full model as it is a running decayed average of the model weights.