How to run HiDream-I1 on a cloud GPU

Rent a GPU on Deploy, then open a terminal in JupyterLab or over SSH. Install HiDream-I1 from the official repository, let inference.py download the variant you pass, and copy output.png off the instance before you destroy it. Destroying the instance deletes its disk.

The README table names three variants: HiDream-I1-Full, HiDream-I1-Dev, and HiDream-I1-Fast, all run with inference.py. Other image pages: How to run FLUX.1-dev in ComfyUI on a cloud GPU, How to run Qwen-Image on a cloud GPU, and How to run Stable Diffusion 3.5 in ComfyUI on a cloud GPU. How published FLUX and SDXL weight sizes are read: How much VRAM do I need for FLUX and SDXL. The ComfyUI example for this model is a separate section below.

The README says “The code in this repository and the HiDream-I1 models are licensed under MIT License.” LICENSE is the MIT License, and its copyright line is “Copyright © 2025 HiDream.ai” (read 2026-11-12). The GitHub license field the same day was MIT. The model cards for HiDream-ai/HiDream-I1-Full, HiDream-ai/HiDream-I1-Dev, and HiDream-ai/HiDream-I1-Fast each had license mit and gated false. The Full card’s License Agreement says the Transformer models in that repository are under the MIT License, names a VAE source as FLUX.1 [schnell] under Apache 2.0, and names two text-encoder repository ids. This page does not copy those repository ids. Copy that paragraph from the model card. The same paragraph says you own all content you create with the model, that you must comply with the license agreement, and that you are responsible for how you use the models. The grant line and the model card’s “freely” sentence are quoted in the FAQ, with one note that they are not a CloudGPU price and not a storage allowance.

The commands below are copied from the upstream files named in each step. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.

What the README and the model cards say about memory

The README’s opening sentence says HiDream-I1 has 17B parameters. That is a parameter count, not a VRAM figure. The README, inference.py, gradio_demo.py, and the three model cards publish no GPU GB figure. A search of those model-card texts on 2026-11-12 found no VRAM line. The evaluation tables in the README are not copied here. This page did not re-measure a run.

GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-12T23:03:50.976Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-12T23:03:51.922Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. Every prices row has gpu_count 1. Every vram_gb below is 8 or higher. This page does not mark a row as enough, because the README publishes no GB figure to compare. Which listing you can start is on Deploy.

The catalog template named ComfyUI has id flux-1-dev, category image-gen, recommended_gpu RTX 4090 48G, and min_vram_gb 24. Those are template fields, not a HiDream README line. The template list had no HiDream template.

Catalog name vram_gb
RTX 2080 8
RTX 3080 10
RTX 2080 Ti 11
RTX 3080 Ti 12
V100 16G 16
A30 24
RTX 3090 24
RTX 4090 24
RTX 4090D 24
RTX 4080 SUPER 32G 32
RTX 5090 32
RTX 5090D 32
V100 32G 32
A100 40G 40
A800 40G 40
A40 48
RTX 4090 48G 48
RTX 4090D 48G 48
RTX A6000 48
NVIDIA RTX PRO 5000 72G 72
RTX PRO 5000 72G 72
A100 80G 80
A800 80G 80
H800 80
RTX 6000D 83
H20 96G 96
RTX PRO 6000 96G 96
RTX PRO 6000 Max-Q 96
H20-3e 141G 141

Full, Dev, and Fast

The README table, read 2026-11-12:

Name Script Inference steps Hugging Face repo
HiDream-I1-Full inference.py 50 HiDream-ai/HiDream-I1-Full
HiDream-I1-Dev inference.py 28 HiDream-ai/HiDream-I1-Dev
HiDream-I1-Fast inference.py 16 HiDream-ai/HiDream-I1-Fast

inference.py stores the same step counts, and also guidance_scale 5.0 for full and 0.0 for dev and fast. shift is 3.0 for full and fast, and 6.0 for dev. full uses FlowUniPCMultistepScheduler. dev and fast use FlashFlowMatchEulerDiscreteScheduler. The resolution strings in that file are 1024 × 1024, 768 × 1360, 1360 × 768, 880 × 1168, 1168 × 880, 1248 × 832, and 832 × 1248. The script’s own prompt example is A cat holding a sign that says "Hi-Dreams.ai". and it saves output.png.

Install and run the inference script

On Deploy, choose Raw GPU. The deploy page says “SSH access and root within a dedicated container”. The README does not print a git clone. This page adds the clone, and it was not run:

git clone https://github.com/HiDream-ai/HiDream-I1
cd HiDream-I1

The commits API on 2026-11-12 listed 5f92bab45f1dfb1e794ee357286a5b837eaf4400 on the default branch main, committed 2025-07-16T14:38:12Z, message “Update README.md”. The quick start says to install Flash Attention first, and “We recommend CUDA versions 12.4 for the manual installation.” It then prints:

pip install -r requirements.txt
pip install -U flash-attn --no-build-isolation

requirements.txt is torch>=2.5.1, torchvision>=0.20.1, diffusers>=0.32.1, transformers>=4.47.1, einops>=0.7.0, and accelerate>=1.2.1. It does not list gradio and it does not pin a Python version. The Gradio section below adds an unpinned install because gradio_demo.py imports gradio. How to install Flash Attention on a cloud GPU is the wheel note for that second line. The README then prints:

python ./inference.py --model_type full
python ./inference.py --model_type dev
python ./inference.py --model_type fast

inference.py sets --model_type default to dev. The print line above the load call is the fixed string Loading default model (full)..., and load_models uses the argument, not that string. Passing no flag therefore loads dev while the print still says full. gradio_demo.py sets current_model = "full" and loads that value, so the Gradio script’s default really is full. This page did not run either script.

The README note says the inference script tries to download a text encoder automatically, that you need to agree to that model’s license on your Hugging Face account, and that you then run huggingface-cli login. The note and the diffusers example do not print the same repository id. inference.py and gradio_demo.py assign that id to LLAMA_MODEL_NAME. This page does not copy any of those ids. Copy each line from the file you are following. How to download a Hugging Face model is the login reference.

The diffusers block in the README loads HiDream-ai/HiDream-I1-Full, with the comment "HiDream-ai/HiDream-I1-Dev" | "HiDream-ai/HiDream-I1-Fast", height 1024, width 1024, guidance_scale 5.0 with the comment 0.0 for Dev&Fast, num_inference_steps 50 with the comment 28 for Dev and 16 for Fast, and image.save("output.png"). The two from_pretrained calls that load the text encoder are the lines this page does not copy. The README also prints pip install git+https://github.com/huggingface/diffusers.git before that block.

Gradio demo and SSH

The README prints:

python gradio_demo.py

gradio_demo.py imports gradio after torch, builds a gr.Radio of full, dev, and fast, default full, and calls demo.launch() with no arguments. generate_image returns the image to a gr.Image component and does not call save. requirements.txt does not list gradio, and the README does not print an install for that import. This page adds an unpinned install, because upstream does not pin a version. It was not run. This page still does not name a port: demo.launch() passes none, and the package is unpinned. Use the URL the process prints.

pip install gradio
python gradio_demo.py

The tunnel is the form in how to use SSH port forwarding on a cloud GPU:

ssh -N -L <port>:127.0.0.1:<port> -p <ssh-port> <ssh-user>@<ssh-host>

Then open http://127.0.0.1:<port> on your computer. This page did not run gradio_demo.py.

To keep that process after the SSH session closes, this page wraps the same two commands in tmux. The README does not print tmux, and it does not print pip install gradio:

tmux new -s hidream
cd HiDream-I1
pip install gradio
python gradio_demo.py

ComfyUI route, from the official example

The ComfyUI examples repository has a hidream directory. Its README, commit 9997687599f477ea4ff407623186516414221102 (2025-07-21T19:19:35Z), says to put four text-encoder files in ComfyUI/models/text_encoders:

  • clip_l_hidream.safetensors
  • clip_g_hidream.safetensors
  • t5xxl_fp8_e4m3fn_scaled.safetensors
  • llama_3.1_8b_instruct_fp8_scaled.safetensors

Those filenames are the example’s own list, on Comfy-Org/HiDream-I1_ComfyUI. The fourth filename is copied here as the example prints it. This page still does not copy a vendor repository id. The same page says the VAE file ae.safetensors goes in ComfyUI/models/vae/, and that this is the Flux VAE. Diffusion files go in ComfyUI/models/diffusion_models/. The example names hidream_i1_dev_bf16.safetensors for the dev workflow and hidream_i1_full_fp16.safetensors for the full workflow, and says to drag the matching example image into ComfyUI. It also names an edit model, hidream_e1_1_bf16.safetensors, and an older edit file, hidream_e1_full_bf16.safetensors. It does not name a Fast diffusion file. Fast remains the inference.py --model_type fast path above.

The example README does not name an output folder. ComfyUI folder_paths.py on commit c1eece62cbf1805d4f64286d7f5592e6a215e6ba (2026-11-12T20:46:52Z) sets output_directory to base_path joined with output. Installing ComfyUI is How to install ComfyUI and ComfyUI-Manager on a cloud GPU. The latest commit on the examples repository was f9431bb000ce792094ff345446e22cac1ea6cef3, committed 2025-11-26T20:33:58Z. The top-level list that day included hidream.

When it fails

CUDA out of memory. The README and inference.py do not print an out-of-memory flag or a GB target. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.

“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. The README recommends CUDA 12.4 for the manual Flash Attention install. requirements.txt allows torch>=2.5.1 and does not name compute capability 12.0. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and needs a PyTorch build for CUDA 12.8 or newer. This page did not install a torch wheel and did not re-test HiDream-I1 on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. That is the catalog number only.

Text-encoder download refused. The README note says the automatic download needs an accepted license and huggingface-cli login. The id to accept is the line this page does not copy. Copy it from the README note or from LLAMA_MODEL_NAME in the script.

Save the images, then destroy

inference.py and the README diffusers block write output.png in the working directory. gradio_demo.py returns the image to the page and does not call save. A ComfyUI save node writes under the output directory above. Copy those files off before you destroy the instance. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

The pricing page FAQ states the billing rule in these words:

Per-minute options have a 1-minute minimum; unused prepaid minutes are refunded when you destroy the instance. Some configurations use whole-hour billing with a 1-hour minimum; unused time within a charged hour is not refunded on voluntary stop. The billing period, minimum charge and initial prepayment are shown before confirmation. No subscription.

Prices are not copied here.

Frequently asked questions

How much VRAM do the HiDream-I1 README and model cards publish?

The README, inference.py, gradio_demo.py, and the Full, Dev, and Fast model cards publish no GPU GB figure. The README says HiDream-I1 has 17B parameters. That is a parameter count, not a VRAM figure. Every vram_gb in https://cloudgpu.app/gpu-prices.json (updated_at 2026-11-12T23:03:50.976Z) is 8 or higher. This page does not mark a row as enough. The catalog template named ComfyUI has id flux-1-dev, recommended_gpu RTX 4090 48G, and min_vram_gb 24. Those are template fields. Which listing you can start is on https://cloudgpu.app/deploy.

What are the Full, Dev, and Fast variants?

The README table names HiDream-I1-Full at 50 steps, HiDream-I1-Dev at 28 steps, and HiDream-I1-Fast at 16 steps, all through inference.py. The script also sets guidance_scale 5.0 for full and 0.0 for dev and fast. The argparse default is dev, while the print string still says full. The Gradio script defaults to full. The ComfyUI example names a dev file and a full file and does not name a Fast diffusion file.

What is the license, and where is the text-encoder id?

LICENSE is the MIT License, copyright line Copyright © 2025 HiDream.ai. The grant says permission is granted free of charge, and the model card says you can use your generated content freely. Those two sentences are license text, not a CloudGPU price and not a storage allowance. The README note, the diffusers example, and LLAMA_MODEL_NAME in inference.py and gradio_demo.py each print a text-encoder repository id, and the note does not match the diffusers example. This page does not copy those ids. Copy the line from the file you are following. The model card also names text-encoder repository ids. Copy that paragraph from the model card.

Do the images stay on the instance after I destroy it?

No. inference.py writes output.png in the working directory. gradio_demo.py returns the image to the page and does not call save. A ComfyUI save node writes under the output directory, which folder_paths.py joins to base_path. Copy those files off before you destroy the instance. The terms say data on local instance storage is permanently deleted when the instance is destroyed.

Sources (read 2026-11-12, Asia/Shanghai)

Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-12 (Asia/Shanghai).