ERNIE-Image¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
ernie-image |
baidu/ERNIE-Image | 5346b31 |
Apache-2.0 | Verified |
Catalog: ernie-image · Models:
ernie-image · Task: t2i
Architecture. The language model of Ministral 3 (26 layers, width 3072, grouped-query attention with YaRN rotary embeddings and a position-dependent query scale) supplies the second-to-last hidden state of the prompt. An 8B single-stream diffusion transformer with 36 blocks, width 4096 and 32 heads of width 128 processes the image tokens and the text tokens together, with shared AdaLN modulation and 3-axis RoPE. The FLUX.2 VAE decodes 32-channel latents that are patchified to 128 channels. The flow-match Euler scheduler uses a static shift of 4.0. The native image size is 1024×1024.
Usage. The repository is not gated. The weights are published in bfloat16 and are upcast on loading; the snapshot is about 32 GB, and float32 needs about 46 GB of memory.
from imagehub import pipeline
with pipeline("t2i", backend="ernie-image", model="checkpoints/ernie-image", device="mps") as pipe:
result = pipe("a poster that says 'HELLO', flat design", seed=42)
result.images[0].save("poster.png")
Settings. The model defaults to 50 steps, guidance 4.0 and 1024×1024; width and height
must be multiples of 16. Guidance above 1.0 also runs the negative prompt (default empty).
The other supported options are num_images, seed and latents. Unrecognized settings raise
an error.
Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with
generator=torch.Generator("cpu").manual_seed(seed) and use_pe=False (validation).
MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.
Out of scope. The optional prompt enhancer (the pe model that rewrites a short prompt, which
diffusers enables by default with use_pe=True) is not implemented: it samples text with the global
random generator, and its weights are not part of the snapshot. Images match the reference run with
use_pe=False. Image input, LoRA and fp16/bf16 parity are not supported either.
License. Apache-2.0. imagehub does not distribute weights.