moonbox-dev/FLUX.1-Fill-dev-MLX

FLUX.1 Fill Dev MLX

MLX builds of FLUX.1 Fill Dev, made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's files together. Only the parts listed below are converted: the rest of the pipeline comes from the original repository unless the section below says otherwise.

Files

The last column is how far each build is from the original PyTorch model on the same inputs (see the section below for what was measured).

Precision	File	Size	sha256	Checked against the original
8bit	8bit/mflux_model.json	0.4 kB	f4ff5fbab1ee	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/text_encoder/0.safetensors	130.9 MB	29daf483b500	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/text_encoder/model.safetensors.index.json	25.0 kB	2496ca8b591e	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/text_encoder_2/0.safetensors	2.15 GB	c93d0244058a	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/text_encoder_2/1.safetensors	2.15 GB	49b24e32092e	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/text_encoder_2/2.safetensors	767.7 MB	4a896dfbcca7	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/text_encoder_2/model.safetensors.index.json	44.1 kB	310dde2f9694	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/tokenizer/tokenizer.json	3.6 MB	6d9109cc8389	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/tokenizer/tokenizer_config.json	0.4 kB	d0eda7377bb9	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/tokenizer_2/tokenizer.json	2.4 MB	2432e8aa8341	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/tokenizer_2/tokenizer_config.json	2.5 kB	e31ed19bc545	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/0.safetensors	2.14 GB	4c1b79237ceb	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/1.safetensors	2.13 GB	56a82a0053d3	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/2.safetensors	2.13 GB	b2b5f93d2ea6	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/3.safetensors	2.13 GB	09efe1817336	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/4.safetensors	2.11 GB	632c4607b56b	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/5.safetensors	2.03 GB	68a5b71bbc50	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/transformer/model.safetensors.index.json	0.1 MB	a47d31736f65	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/vae/0.safetensors	165.7 MB	6ddcda403d28	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
8bit	8bit/vae/model.safetensors.index.json	17.4 kB	0a98a344cba2	reference build: not compared with bf16 (doesn't run in 24 GB); inpaints correctly on the 3 photos
6bit	6bit/mflux_model.json	0.4 kB	9171d42971d1	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/text_encoder/0.safetensors	100.2 MB	021972a101b1	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/text_encoder/model.safetensors.index.json	25.0 kB	bb76a6a32514	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/text_encoder_2/0.safetensors	2.15 GB	d48839ac576b	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/text_encoder_2/1.safetensors	1.72 GB	4a9a542456a7	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/text_encoder_2/model.safetensors.index.json	44.1 kB	ac2d435419f7	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/tokenizer/tokenizer.json	3.6 MB	6d9109cc8389	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/tokenizer/tokenizer_config.json	0.4 kB	d0eda7377bb9	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/tokenizer_2/tokenizer.json	2.4 MB	2432e8aa8341	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/tokenizer_2/tokenizer_config.json	2.5 kB	e31ed19bc545	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/transformer/0.safetensors	2.12 GB	ff829e699345	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/transformer/1.safetensors	2.15 GB	10eb8c736234	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/transformer/2.safetensors	2.13 GB	e551955f75c7	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/transformer/3.safetensors	2.14 GB	f2c528bcbb50	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/transformer/4.safetensors	1.14 GB	c7ee01b73467	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/transformer/model.safetensors.index.json	0.1 MB	302702931b26	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/vae/0.safetensors	165.2 MB	5fb147b7776d	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
6bit	6bit/vae/model.safetensors.index.json	17.4 kB	20fb1f73bf5e	below threshold (experimental): vs 8bit PSNR 37.8 dB mean, LPIPS 0.056 mean / 0.163 worst
4bit	4bit/mflux_model.json	0.4 kB	9a500615e1c2	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/text_encoder/0.safetensors	69.4 MB	80401ef4d0cb	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/text_encoder/model.safetensors.index.json	25.0 kB	87918a51d110	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/text_encoder_2/0.safetensors	2.15 GB	daf9b170b819	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/text_encoder_2/1.safetensors	533.3 MB	3db2a6e9309a	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/text_encoder_2/model.safetensors.index.json	44.1 kB	03f5a8751ab8	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/tokenizer/tokenizer.json	3.6 MB	6d9109cc8389	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/tokenizer/tokenizer_config.json	0.4 kB	d0eda7377bb9	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/tokenizer_2/tokenizer.json	2.4 MB	2432e8aa8341	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/tokenizer_2/tokenizer_config.json	2.5 kB	e31ed19bc545	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/transformer/0.safetensors	2.13 GB	dbf7be651fcf	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/transformer/1.safetensors	2.14 GB	ae7fa499db8e	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/transformer/2.safetensors	2.13 GB	0bdccd27d8fb	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/transformer/3.safetensors	297.6 MB	6437f5d4cba1	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/transformer/model.safetensors.index.json	0.1 MB	ef05979b2043	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/vae/0.safetensors	164.7 MB	53aa93d75594	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst
4bit	4bit/vae/model.safetensors.index.json	17.4 kB	56348f08a8c9	vs 8bit PSNR 32.0 dB mean, LPIPS 0.051 mean / 0.107 worst

How to run

Each folder is the whole FLUX.1 Fill [dev] pipeline (T5 and CLIP text encoders, transformer, VAE, tokenizers) saved by mflux 0.24.0 from the original weights at the revision below. Run it with mflux; the mask is white where the picture should be repainted:

pip install mflux==0.24.0
hf download moonbox-ai/FLUX.1-Fill-dev-MLX --include "8bit/*" --local-dir fill
mflux-generate-fill --model fill/8bit --base-model dev-fill --image-path photo.png --masked-image-path mask.png \
  --prompt "an empty street" --steps 20 --seed 7 --output out.png

Add --low-ram on a 24 GB Mac. 8bit/, 6bit/ and 4bit/ quantize the transformer's and text encoders' linear layers (MLX affine, group size 64); the VAE stays bf16.

No bf16/ folder. The lab's Mac (24 GB) can't run the 12B transformer unquantized (it thrashes swap), so a bf16 build couldn't be checked, and building it needs ~85 GB of free disk. The original repo's weights are already bf16 and mflux runs them directly (--model black-forest-labs/FLUX.1-Fill-dev).

How they were checked. Three photos with their subject masked out, 512x512, guidance 30, 20 steps, the same seed: 6bit/ and 4bit/ against 8bit/ (nothing bigger fits in 24 GB). Pass: LPIPS <= 0.15 on every picture. 8bit/ itself wasn't compared with bf16; it fills the masked areas correctly.

Precision	Size	PSNR vs 8bit (mean)	LPIPS (mean / worst)	Status
8bit/	17.0 GB	reference	-	not compared with bf16
6bit/	13.8 GB	37.8 dB	0.056 / 0.163	just below threshold, experimental
4bit/	9.6 GB	32.0 dB	0.051 / 0.107	passes

About 5 minutes per 512x512 picture (20 steps) for 8bit/ on an Apple M5 with 24 GB in low-RAM mode.

How these were made

From FLUX.1 Fill Dev at revision 358293da03: the original weights loaded in MLX and stored as bf16 or fp16, or quantized with MLX's affine quantization (8bit, 6bit, 4bit, group size 64; layers MLX can't quantize, such as convolutions and norms, stay bf16). Every build was checked against the original before upload; the sha256 column above is what the files must hash to.

License

FLUX.1-dev Non-Commercial (gated). These files are a derivative of the original model and carry its license: it is copied in this repository, and it governs your use of these files. The original model card follows unchanged.

---

Original model card

FLUX.1 Fill [dev] is a 12 billion parameter rectified flow transformer capable of filling areas in existing images based on a text description. For more information, please read our blog post.

Key Features

1. Cutting-edge output quality, second only to our state-of-the-art model FLUX.1 Fill [pro].
2. Blends impressive prompt following with completing the structure of your source image.
3. Trained using guidance distillation, making FLUX.1 Fill [dev] more efficient.
4. Open weights to drive new scientific research, and empower artists to develop innovative workflows.
5. Generated outputs can be used for personal, scientific, and commercial purposes as described in the FLUX.1 [dev] Non-Commercial License.

Usage

We provide a reference implementation of FLUX.1 Fill [dev], as well as sampling code, in a dedicated github repository. Developers and creatives looking to build on top of FLUX.1 Fill [dev] are encouraged to use this as a starting point.

API Endpoints

The FLUX.1 models are also available in our API bfl.ml

Diffusers

To use FLUX.1 Fill [dev] with the 🧨 diffusers python library, first install or upgrade diffusers

pip install -U diffusers

Then you can use FluxFillPipeline to run the model

import torch
from diffusers import FluxFillPipeline
from diffusers.utils import load_image

image = load_image("https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/cup.png")
mask = load_image("https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/cup_mask.png")

pipe = FluxFillPipeline.from_pretrained("black-forest-labs/FLUX.1-Fill-dev", torch_dtype=torch.bfloat16).to("cuda")
image = pipe(
    prompt="a white paper cup",
    image=image,
    mask_image=mask,
    height=1632,
    width=1232,
    guidance_scale=30,
    num_inference_steps=50,
    max_sequence_length=512,
    generator=torch.Generator("cpu").manual_seed(0)
).images[0]
image.save(f"flux-fill-dev.png")

To learn more check out the diffusers documentation

---

Limitations

• This model is not intended or able to provide factual information.
• As a statistical model this checkpoint might amplify existing societal biases.
• The model may fail to generate output that matches the prompts.
• Prompt following is heavily influenced by the prompting-style.
• There may be slight-color shifts in areas that are not filled in
• Filling in complex textures may produce lines at the edges of the filled-area.

Out-of-Scope Use

The model and its derivatives may not be used

• In any way that violates any applicable national, federal, state, local or international law or regulation.
• For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; including but not limited to the solicitation, creation, acquisition, or dissemination of child exploitative content.
• To generate or disseminate verifiably false information and/or content with the purpose of harming others.
• To generate or disseminate personal identifiable information that can be used to harm an individual.
• To harass, abuse, threat