dflow_2b

Preview, version 0.21. The model is still training, so checkpoints and behaviour will change between releases. Generations are still uneven, expect some seed hunting.

dflow_2b is an indie text-to-image model, trained from scratch on a single GPU, mostly on photographs and classical art.

It takes ordinary captions, and also JSON layout captions (the same format as MOSS-Image).

The package contains the weights, the dflow inference code and ComfyUI nodes.

Example gallery · Manual · Technical report · Release notes

Boardwalk to a rugged coast Tortoiseshell cat in tall grass Rooster among tropical plants Studio portrait of a man
Narrow street of wooden houses Landscape oil painting Jazz festival poster from a JSON layout Woman holding a red perfume bottle

What's included

  • model.safetensors: the 2B denoiser.
  • text_encoder/: the part of T5Gemma 2 270M the model uses, under the Gemma Terms of Use.
  • dflow/: the Python inference package.
  • comfyui/: ComfyUI nodes and example workflows.

The autoencoder, dinac3_96, is our own and downloads automatically on first use. The technical report covers the architecture and training.

Requirements

Python 3.12 or later, PyTorch 2.13 with CUDA, and an NVIDIA GPU with bfloat16 support and about 7 GB of free VRAM.

Generation is compiled with torch.compile. The first image takes an extra 30 to 45 seconds while it compiles; after that, an 832×1216 image takes about 10 seconds on an RTX 4090.

ComfyUI

The repository is a ComfyUI custom-node folder. Download it into custom_nodes and install its requirements:

cd ComfyUI/custom_nodes
hf download data-archetype/dflow_2b --local-dir dflow_2b
pip install -r dflow_2b/requirements.txt

Load dflow_2b/comfyui/workflows/dflow_default.json, or drag any image from the gallery onto the canvas to get the workflow that made it. Start ComfyUI with --preview-method auto for live previews.

Python

hf download data-archetype/dflow_2b --local-dir dflow_2b
pip install -e dflow_2b
from dflow.inference import GenerationSettings
from dflow.pipeline import DFlowPipeline

pipe = DFlowPipeline.from_pretrained("dflow_2b")
image = pipe("A lighthouse on a cliff at dusk.", GenerationSettings(seed=7)).images[0]
image.save("lighthouse.png", pnginfo=pipe.png_info(image))

from_pretrained also takes the Hub id data-archetype/dflow_2b, in which case the files go to the Hugging Face cache. If you have diffusers installed, a thin diffusers-style wrapper is available as dflow.diffusers_pipeline.DFlowDiffusersPipeline.

Prompts

Write captions the way you would describe a photo or a painting: the subject, the setting, the light and the medium. Long prompts are fine, up to 6,144 tokens.

Put text you want in the image in double quotes, as in A shop sign that reads "OPEN LATE". The model reads quoted text letter by letter, which helps with spelling.

A layout describes the picture as a whole and then lists its elements, each with a label, a caption and a box [y0, x0, y1, x1] on a 0–1000 scale. Text elements carry the exact string to render:

{"summary": {"canvas": "single",
             "caption": "A travel poster of a mountain above a lake."},
 "frame": {"caption": "The peak sits in the upper middle.",
           "palette": ["#F3E6C8", "#2F5D62"]},
 "items": [
   {"label": "object", "caption": "A snow-capped mountain.",
    "box_2d": [250, 180, 620, 820]},
   {"label": "text", "caption": "The title in tall serif capitals.",
    "text": "SILVER LAKE", "box_2d": [120, 110, 220, 890]}]}

Use it as the prompt in Python, or paste it into the DFlow Layout Encode node. The manual covers every field and works through a full poster.

Limitations

  • It is a preview, and quality varies from prompt to prompt and seed to seed.
  • Hands and anatomy often go wrong.
  • Small text and long passages of text are hit and miss.
  • Above about one megapixel results get less reliable. For large images, move the self-attention gain towards 0.
  • Styles far from photography and classical art are weaker.

You are responsible for how you use the model and its outputs.

Licence

Third-party material is listed in ATTRIBUTION.md.

Citation

@misc{dflow_2b,
  title  = {dflow\_2b},
  author = {{Data Archetype}},
  year   = {2026},
  url    = {https://huggingface.co/data-archetype/dflow_2b},
  note   = {Preview release, version 0.21}
}
Downloads last month
26
Safetensors
Model size
2B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for data-archetype/dflow_2b

Finetuned
(19)
this model

Space using data-archetype/dflow_2b 1