HunyuanDiT¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
hunyuan-dit-v1.1-distilled |
Tencent-Hunyuan/HunyuanDiT-v1.1-Diffusers-Distilled | 527cf2e |
Tencent Hunyuan Community License | Verified |
hunyuan-dit-v1.2 |
Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers | 5e96094 |
Tencent Hunyuan Community License | Verified |
hunyuan-dit-v1.2-distilled |
Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers-Distilled | ba991d1 |
Tencent Hunyuan Community License | Verified |
Catalog: hunyuan-dit ·
Models: hunyuan-dit-v1.1-distilled, hunyuan-dit-v1.2, hunyuan-dit-v1.2-distilled · Task: t2i
Architecture. Bilingual Chinese/English BERT-large and mT5-XL text encoders condition a 1.5B
diffusion transformer with 40 blocks, width 1408, 16 attention heads of width 88, long skip
connections and 2D RoPE. It also uses timestep, image-size, style and pooled-text conditioning and
predicts a learned sigma. A 4-channel VAE with scaling factor 0.13025 decodes the latents. The DDPM
scheduler uses v-prediction, and requested resolutions are binned to supported shapes. The v1.2
releases drop the image-size and style conditioning (use_style_cond_and_image_meta_size is false)
and use a smaller final beta (0.018) in the scheduler.
Usage.
from imagehub import pipeline
with pipeline(
"t2i",
backend="hunyuan-dit-v1.1-distilled",
model="checkpoints/hunyuan-dit-v1.1-distilled",
device="mps",
) as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. steps defaults to 25 (50 for hunyuan-dit-v1.2), guidance defaults to 5.0 and width/height default to
1024. The other supported options are seed, latents, negative_prompt, guidance_rescale and
use_resolution_binning. 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) (validation). MPS and CUDA
runs are smoke tests only; there is no claim of exact equality on these devices.
Out of scope. Other HunyuanDiT releases (v1.0 and the non-distilled v1.1), ControlNet, PAG, LoRA and fp16/bf16 parity are not supported.
License. The weights use the Tencent Hunyuan Community License, which excludes use in the European Union, the United Kingdom and South Korea; see the 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

referenceimagehub

referenceimagehub