HiDreamAI Full
The undistilled 17B-parameter HiDream-I1 text-to-image checkpoint. It is the only one of the three built to run with classifier-free guidance, so the guidance scale and negative prompt both shape the image.
Overview
HiDream-I1 Full is the undistilled release of HiDream.ai's 17B-parameter text-to-image model and the teacher that Dev and Fast were distilled from. It is a sparse diffusion transformer with dynamic mixture-of-experts layers: 16 blocks handle image and text tokens in separate streams, then 32 more process them as one sequence. Four text encoders read the prompt (Long-CLIP L/14 and G/14, T5-XXL and Llama 3.1 8B Instruct), and the VAE comes from FLUX.1 schnell.
Of the three checkpoints, only Full is meant to run with classifier-free guidance. HiDream reports a GenEval overall score of 0.83, 85.89 on DPG-Bench and 33.82 on HPSv2.1 for HiDream-I1, without naming the checkpoint.
Choosing a variant: use Full when the negative prompt and guidance scale need to steer the result and 50 steps are acceptable. HiDream-I1 Dev is guidance-distilled for 28 steps, HiDream-I1 Fast goes down to 16 as the family's lighter variant, and HiDream-I1 Text to Image runs any of the three from one form.
What you can build
This page's samples include a neon cyberpunk street, a golden retriever astronaut above Earth, a castle on a floating island, a sports car on a desert road and a samurai in a bamboo forest.
- Concept art and key frames for game or film pre-production.
- Product shots where a negative prompt is meant to keep out text or people, since Full's defaults let it take effect.
Inputs
You write a prompt and can add a negative prompt for things to keep out. The advanced settings cover inference steps (1 to 500, default 45), guidance scale (0 to 20, default 5.0), flow shift (1 to 10, default 3.0), images per run (1 to 8), a seed (0 picks a random one), and width and height (up to 2048, default 1024).
Outputs
One to eight images per request, 1024 x 1024 by default.
Recommended settings
- Guidance: stay near 5.0. At 1 or below, HiDream's pipeline switches classifier-free guidance off and ignores the negative prompt.
- Steps: HiDream uses 50; the default here is 45.
- Flow shift: 3.0, the value HiDream pairs with Full.
- Size: training went up to 1024 x 1024, and the reference presets stay near one megapixel, such as 1360 x 768.
Limitations
- The weakest category in HiDream's DPG-Bench table is Global, at 76.44, against 89.48 to 93.74 for the other four.
- On GenEval, position scores 0.60, its lowest category, so placement instructions such as "to the left of" may not come out as written.
- By default the pipeline truncates prompts at 128 tokens, so put the important details early.
Safety & compliance
The transformer weights are MIT-licensed. The VAE and the T5-XXL encoder are covered by Apache 2.0, and Llama 3.1 8B Instruct by the Llama 3.1 Community License. HiDream says you own what you generate and may use it in commercial work. The card forbids illegal content, harmful material, personal information that could harm others, false information, and content targeting vulnerable groups.
Example prompts
Great starting points for HiDream-I1-Full.
API quick start
Run HiDream-I1-Full with a single API call.
{
"prompt": "A golden retriever astronaut floating in …",
"negativePrompt": "...",
"steps": 45,
"scale": 5.0
}curl -X POST "https://api.wiro.ai/v1/Run/hidreamai/hidream-i1-full" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A golden retriever astronaut floating in …",
"negativePrompt": "...",
"steps": 45,
"scale": 5.0
}
JSONOther models you might like
Explore similar image generation models.



