FLUX.2 Image Generation Workflow | Generated by AI

Home 2026.07

Question: Explain download_flux.py, run_flux.py, and the /workspace files.

Answer:

This is a typical Hugging Face Diffusers FLUX.2 image generation workflow:

/workspace
├── download_flux.py     # download model weights
├── run_flux.py          # load model + generate image
├── FLUX.2-dev/          # downloaded model checkpoint
├── venv/                # Python virtual environment
└── lost+found/          # filesystem recovery directory

The flow is:

HuggingFace
    |
    | snapshot_download()
    v
/workspace/FLUX.2-dev
    |
    | Flux2Pipeline.from_pretrained()
    v
GPU memory (partially)
    |
    | pipe(prompt)
    v
flux_output.png

1. `download_flux.py`

Purpose:

Download FLUX.2-dev model files from Hugging Face into local disk.

import os, time, subprocess
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"

This changes Hugging Face endpoint:

Normally:

https://huggingface.co

becomes:

https://hf-mirror.com

Useful if huggingface.co is slow/unreachable.

Import downloader

from huggingface_hub import snapshot_download

snapshot_download() downloads an entire Hugging Face repository.

Similar to:

git clone https://huggingface.co/black-forest-labs/FLUX.2-dev

but optimized for model files.

Download

path = snapshot_download(
    "black-forest-labs/FLUX.2-dev",
    local_dir="/workspace/FLUX.2-dev",
    resume_download=True,
    local_dir_use_symlinks=False
)

Meaning:

Source:

black-forest-labs/FLUX.2-dev

Destination:

/workspace/FLUX.2-dev

Example structure:

FLUX.2-dev/
├── model_index.json
├── scheduler/
├── text_encoder/
├── transformer/
├── vae/
└── tokenizer/

The model is not one file.

A diffusion model consists of multiple components:

             prompt
               |
               v
        Text Encoder
               |
               v
       conditioning vectors
               |
               v
Random noise ---> Transformer ---> VAE Decoder
               |
               v
            image

Resume download

resume_download=True

If interrupted:

50GB model
downloaded 30GB
network fails

restart continues:

30GB -> 50GB

instead of restarting.

Check size

subprocess.run(["du", "-sh", path])

Equivalent:

du -sh /workspace/FLUX.2-dev

Example:

Size: 80G

2. `run_flux.py`

Purpose:

Load FLUX.2-dev and generate an image.

Environment

os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"

Same mirror setting.

Import PyTorch + Diffusers

import torch
from diffusers import Flux2Pipeline

Flux2Pipeline is the high-level inference wrapper.

It combines:

Flux Transformer
+
Text Encoder
+
VAE
+
Scheduler
+
Tokenizer

into one API.

Load model

pipe = Flux2Pipeline.from_pretrained(
    "/FLUX.2-dev",
    torch_dtype=torch.bfloat16
)

Important:

Your model path:

/FLUX.2-dev

NOT:

/workspace/FLUX.2-dev

This means your image probably has:

/FLUX.2-dev

mounted already.

Check:

ls -lh /FLUX.2-dev

If not:

change to:

pipe = Flux2Pipeline.from_pretrained(
    "/workspace/FLUX.2-dev",
    torch_dtype=torch.bfloat16
)

Why bfloat16?

torch_dtype=torch.bfloat16

FLUX is huge.

FP32:

1 parameter = 32 bits

BF16:

1 parameter = 16 bits

Memory:

FP32:
20B params × 4 bytes
≈80GB

BF16:
20B params × 2 bytes
≈40GB

Almost half VRAM.

3. CPU offload

pipe.enable_sequential_cpu_offload()

This is the key line.

Without it:

GPU VRAM:

Text encoder
+
Transformer
+
VAE

all loaded

=> OOM

With offload:

CPU RAM
 |
 |
 v

Text encoder
      |
      v
GPU
      |
      v
remove

Transformer
      |
      v
GPU
      |
      v
remove

VAE
      |
      v
GPU

Only one component stays on GPU.

Tradeoff:


VRAM	much lower
speed	slower
RAM	higher

Useful for Radeon 7900 / MI300 / small VRAM cards.

4. Generate image

prompt = "A cute cat wearing a spacesuit on Mars, digital art"

Text prompt.

image = pipe(
    prompt=prompt,
    num_inference_steps=28,
    guidance_scale=3.5,
).images[0]

The diffusion loop:

Start:

random noise

x_T

28 steps:

x_T
 |
 v
denoise
 |
 v
x_27
 |
 v
...
 |
 v
x_0

Finally:

latent
 |
 VAE decoder
 |
 RGB image

`num_inference_steps`

num_inference_steps=28

More steps:

10 steps
fast
lower quality

50 steps
slow
higher quality

28 is a reasonable default.

`guidance_scale`

guidance_scale=3.5

Controls prompt following.

Low:

2
|
more creative
less obedient

High:

8+
|
follows prompt strongly
sometimes artifacts

FLUX usually likes lower CFG compared with Stable Diffusion.

Save

image.save("/workspace/flux_output.png")

Output:

/workspace/flux_output.png

5. `venv`

venv/

Python virtual environment.

Usually contains:

venv/
├── bin/python
├── bin/pip
└── lib/python3.12/site-packages/

Activate:

source venv/bin/activate

Then:

python run_flux.py

uses the isolated packages.

6. `lost+found`

lost+found/

Normal Linux filesystem directory.

Created by ext4.

Used by:

fsck

for recovering orphaned files after disk corruption.

Ignore it.

For your AMD Radeon Cloud setup

Given your previous server:

AMD Radeon Cloud
ROCm 7.2.1
AMD GPU

the important checks are:

rocminfo | head

and:

python -c "import torch; print(torch.cuda.is_available()); print(torch.version.hip)"

Expected:

True
7.2.x

Then:

python run_flux.py

should use ROCm.

The design here is specifically to make a huge FLUX model run on limited VRAM by trading speed for memory using enable_sequential_cpu_offload().

Back Donate