AuraFlow¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
auraflow-v0.3 |
fal/AuraFlow-v0.3 | 2cd8588 |
Apache-2.0 | Verified |
Catalog: auraflow · Models:
auraflow-v0.3 · Task: t2i
Architecture. A UMT5 encoder (24 layers, width 2048, a relative position bias in every layer)
encodes the prompt, and its last hidden state, zeroed at padded positions, conditions a 6.8B-parameter
flow-matching transformer of width 3072 (12 heads of 256). Four MMDiT blocks process the image and
text streams jointly, then 32 single-stream DiT blocks process the two together. Eight learned
register tokens are prepended to the text, queries and keys are layer-normalized per head, and the
image patches (2×2) receive a learned positional embedding that is centre-cropped from a 96×96 grid.
Images are decoded by the 4-channel, 8× SDXL-style VAE (scaling factor 0.13025). The scheduler is
flow-match Euler with a static shift of 1.73; the transformer receives timesteps in [0, 1].
Usage.
from imagehub import pipeline
with pipeline(
"t2i", backend="auraflow-v0.3", model="checkpoints/auraflow-v0.3", device="mps"
) as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. The model defaults to 50 steps, guidance 3.5 and 1024×1024, with an empty negative
prompt. The supported options are seed, steps, guidance, width, height,
negative_prompt and max_sequence_length (default 256). The size must be a multiple of 16 and at
most 1536 per side, which is the extent of the learned positional grid. Guidance 1.0 or lower skips
the unconditional pass. Unrecognized settings raise an error.
Memory. In float32 the transformer needs about 27 GB, the UMT5 encoder about 4.9 GB and the VAE about 0.3 GB. The text encoder is stored in float16 and upcast on load.
Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with
generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA
runs are smoke tests only; there is no claim of exact equality on these devices.
Out of scope. Other AuraFlow releases, LoRA, ControlNet, fused projections, several images per prompt and fp16/bf16 parity are not supported.
License. The repository declares the Apache-2.0 license. imagehub does not distribute weights.
Reference and imagehub, side by side¶
Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.

referenceimagehub