HiDream-O1-Image

1. Quick start

Follow the SGLang Diffusion installation guide, then choose a recipe below. TP 1 and TP 2 on a single H200 are the verified shapes.

2. Model capabilities

HiDream-O1-Image is an 8B pixel-level Unified Transformer (UiT). Unlike the usual latent-diffusion layout, it has neither a VAE nor a separate text encoder: one Qwen3-VL backbone embeds the prompt and predicts clean 32x32 RGB patches directly, and the flow-matching head reads its timestep from a dedicated token in the prompt. It loads through SGLang’s native HiDreamO1ImagePipeline. Choose it for text-to-image at its trained 2K-class resolutions, including legible in-image text. The checkpoint ships a stock qwen3_vl config with no model_index.json, so the pipeline loads the transformer itself; the tradeoff is a narrower feature surface: bf16 only, one prompt per request, no sequence parallelism, and monolithic serving only.

Model	Hugging Face model ID	Notes
HiDream-O1-Image	`HiDream-ai/HiDream-O1-Image`	Text-to-image, bf16 only

3. Generate an image

One-shot generation from the CLI:

Command

sglang generate \
  --model-path HiDream-ai/HiDream-O1-Image \
  --prompt "A cat is sitting next to a sign that says 'HiDream-O1'" \
  --num-inference-steps 50 \
  --guidance-scale 5.0 \
  --save-output

Or through the OpenAI-compatible image generation API once the server is up:

Python

import base64
from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://127.0.0.1:30010/v1")

response = client.images.generate(
    model="HiDream-ai/HiDream-O1-Image",
    prompt="A cinematic photo of a quiet lakeside cabin at sunrise",
    n=1,
    response_format="b64_json",
)

image_bytes = base64.b64decode(response.data[0].b64_json)
with open("hidream_o1.png", "wb") as f:
    f.write(image_bytes)

4. Configuration tips

The model is released in bfloat16 only; --dit-precision / --component-precisions overrides are rejected rather than silently downcast.
The trained resolutions are 2048x2048, 2304x1728, 2560x1440, 2496x1664, 3104x1312, 2304x1792 and their transposes. Nothing snaps: an off-grid size is warned about in the server log and then rendered exactly as requested, and because pixel patches carry no scale invariance the quality degrades badly.
Sequence parallelism is unsupported: the prompt span is causal while the image span is bidirectional, which needs a dense [B, 1, S, S] mask that cannot be sharded along the sequence. Use --tp-size for multi-GPU instead. --cfg-parallel-size 2 is accepted at the default --guidance-scale 5.0, since it splits the two CFG branches across GPUs, but it has not been verified end to end.
Encoder/denoiser/decoder disaggregation does not apply — one component does all three, so the pipeline only runs in the monolithic role.
The defaults are the reference full recipe: 50 steps, --guidance-scale 5.0, shift 3.0 (fixed). Classifier-free guidance costs a second forward pass per step, so --guidance-scale 1.0 halves the step cost at some quality loss.
Noise is drawn on the host, so a given --seed reproduces the same image on any accelerator.

Cookbook

1. Quick start

2. Model capabilities

3. Generate an image

4. Configuration tips

5. Run in ComfyUI

​1. Quick start

​2. Model capabilities

​3. Generate an image

​4. Configuration tips

​5. Run in ComfyUI

1. Quick start

2. Model capabilities

3. Generate an image

4. Configuration tips

5. Run in ComfyUI