dflow_2b is a 2B-parameter flow-matching text-to-image model. Its text encoder is a subset of T5Gemma 2 270M, and the autoencoder dinac3_96 was trained by the author. It accepts both plain descriptions and JSON layout descriptions; text within double quotes is read letter-by-letter, which helps with spelling. The maximum prompt length is 6,144 tokens. The author states that approximately 7 GB of free VRAM is required. Compiling the first image takes an additional 30–45 seconds, after which generating a single 832×1216 image takes about 10 seconds on an RTX 4090. The repository also serves as a ComfyUI custom node folder, including example workflows and a technical report.
Why It Matters It does not rely on any existing base models. Both the text encoder and the autoencoder are small, allowing it to run on consumer-grade GPUs. The layout format is compatible with MOSS-Image, enabling the creation of posters via JSON.
Hugging Face Likes 14 · Hugging Face Trending Score 14 · Hugging Face Downloads 26
License: Non-Commercial · ComfyUI: Nodes included in the repository · VRAM: ≈6 GB (estimated)
Reasons
- The model card includes an example gallery (which can be dragged into ComfyUI to restore workflows), a technical report, and update notes.
- The author explicitly states that this is a 0.215 preview version and is still under training. Image quality is unstable, requiring multiple seed attempts.
Usage
- Python: Run
hf download data-archetype/dflow\_2b --local-dir dflow\_2b, thenpip install -e dflow\_2b, and useDFlowPipeline.from\_pretrained("dflow\_2b")to generate images. Requires Python 3.12, PyTorch 2.13, and an NVIDIA GPU that supports bf16.
Risks
- The MG-BY-NC 2.0 license is non-commercial. The text encoder is additionally subject to the Gemma Terms of Use.
- For ComfyUI usage, the entire repository must be downloaded into
custom\_nodesand dependencies installed. The node code has not been audited. - This is a preview version; weights and performance may change in future versions.