Skip to content

HiDream-O1

Model Source Revision Weight license Parity (CPU, fp32)
hidream-o1-image HiDream-ai/HiDream-O1-Image 0b0901d MIT Verified

Catalog: hidream-o1-image · Models: hidream-o1-image · Task: t2i

Architecture. HiDream-O1 uses a Qwen3-VL-8B decoder with 36 layers, width 4096 and grouped query attention. It appends noisy RGB image patches directly to the chat-formatted text sequence. Each patch contains one 32×32 RGB region; a two-stage linear projection maps it to the decoder width, and a linear output head predicts the clean pixels. Text tokens retain causal attention, while the timestep token and generated image tokens attend to the complete sequence. There is no VAE: the final patch sequence is rearranged directly into an RGB image.

Usage.

python -m imagehub.pretrained hidream-o1-image checkpoints/hidream-o1-image
from imagehub import pipeline

with pipeline(
    "t2i", backend="hidream-o1-image", model="checkpoints/hidream-o1-image"
) as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. The checkpoint defaults to 50 Flow-UniPC steps, classifier-free guidance 5.0, schedule shift 3.0 and 2560×1440 output. width and height must be multiples of 32. The supported options are seed, caller-supplied noise, steps, timesteps, guidance, shift, width, height and num_images. Noise is drawn on CPU with seed + 1 and multiplied by 8.0, matching the published sampler. Unknown options are rejected before checkpoint loading.

Prompt and positions. A single user message is rendered with the checkpoint's Qwen chat template and followed by <|boi_token|><|tms_token|>. Generated patches use the checkpoint's fixed-point multimodal rotary positions beginning at 4096. Guidance above 1.0 adds an unconditional branch whose prompt is one space.

Image input. Not supported yet. Although the published checkpoint contains the Qwen3-VL vision encoder and the official implementation can condition on reference images, this backend currently implements text-to-image only. The catalog therefore declares no image input.

Memory and validation. The eight float32 shards total about 32.8 GiB. The published 759-tensor layout, new pixel modules, tokenizer and positions, Flow-UniPC trajectory, guided and unguided sampling, two seeds and non-square output have offline exact-parity tests. All 5 real-checkpoint tests passed with every recorded comparison exactly 0; the run took 1 minute on an M2 Max (CPU) (validation).

Out of scope. Reference-image, subject, layout and skeleton conditioning, the prompt-refiner agent, distributed inference, flash attention, Dev checkpoints and reduced-precision parity are not supported.

License. The author repository declares the model weights and official code MIT. imagehub does not distribute weights.