HiDreamAI Dev
A guidance-distilled checkpoint of the 17B HiDream-I1 text-to-image model, trained from Full and set up for 28 sampling steps with guidance off and a flow shift of 6.0.
Overview
HiDream-I1 Dev is one of two distilled checkpoints of HiDream.ai's 17B-parameter image model. HiDream trained it from Full with distribution matching distillation plus an adversarial loss and describes it as guidance-distilled for 28 steps. The transformer configuration and the four text encoders are the same as Full's.
What sets Dev apart is its sampler setup. HiDream's reference configuration uses guidance 0.0, a flow shift of 6.0 (twice the value used for Full and Fast) and a flash flow-matching Euler scheduler.
Choosing a variant: Dev sits in the middle by step count. Pick it when you want fewer steps than Full and the prompt alone can describe the image. When the negative prompt has to take effect, go to HiDream-I1 Full, the teacher model and the only checkpoint built around classifier-free guidance. HiDream-I1 Fast, the lighter sibling, is distilled to 16 reference steps. To switch checkpoints from a single form, use HiDream-I1 Text to Image.
What you can build
Dev's samples lean toward photographic scenes: a lion on sunlit rocks, a log cabin in a snowy pine forest, a steam locomotive in a green mountain valley, a Tokyo market street hung with lanterns, and a palm-lined beach at sunrise.
- Destination and hospitality imagery for travel websites and brochures.
- Wildlife and landscape pictures for editorial layouts and blog headers.
- Crowded street and nightlife scenes for game environments or film mood boards.
- A prompt-to-image feature where users only type a description, with guidance left off as in Dev's reference settings.
Inputs
The only required field is the prompt, and a negative prompt is optional. Advanced settings let you change inference steps (1 to 500, 30 by default), guidance scale (0 to 20, 0 by default), flow shift (1 to 10, 6.0 by default), how many images to make (1 to 8), the seed (0 means random) and the output size (1024 by 1024 unless changed, at most 2048 per side).
Outputs
A set of 1 to 8 images, each a 1024-pixel square by default.
Recommended settings
- Guidance 0, matching how Dev was distilled.
- Flow shift 6.0. Reusing Full's 3.0 moves Dev away from its reference setup.
- Around 28 steps, HiDream's reference count. This page starts at 30.
- Sizes near one megapixel, for example 1360 x 768 or 832 x 1248.
- Keep the prompt tight: by default only the first 128 tokens reach the text encoders.
Limitations
- HiDream's pipeline applies classifier-free guidance only when guidance is above 1, so at Dev's default the negative prompt has no effect. Raising guidance takes Dev outside the settings it was distilled for.
- The benchmark tables on HiDream's card are labeled "HiDream-I1" and give no separate numbers for Dev.
- The training schedule topped out at 1024 x 1024, while this page accepts up to 2048 per side.
Safety & compliance
Dev ships under the MIT License. Its VAE (from FLUX.1 schnell) and the T5-XXL encoder carry Apache 2.0, and the Llama encoder follows the Llama 3.1 Community License. According to HiDream, the images you generate are yours to use, commercial projects included. HiDream's terms rule out illegal or harmful content, false information, personal information that could hurt others, and content aimed at vulnerable groups.
Example prompts
Great starting points for HiDream-I1-Dev.
API quick start
Run HiDream-I1-Dev with a single API call.
{
"prompt": "A vintage steam train moving through a fo…",
"negativePrompt": "...",
"steps": 30,
"scale": 0
}curl -X POST "https://api.wiro.ai/v1/Run/hidreamai/hidream-i1-dev" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A vintage steam train moving through a fo…",
"negativePrompt": "...",
"steps": 30,
"scale": 0
}
JSONOther models you might like
Explore similar image generation models.



