HiDream-I1-Full-FP8

This is an FP8 / INT8 quantized version of HiDream-I1-Full. <br>

• Optimized for efficient inference with reduced memory footprint. Loading a checkpoint is bit-identical to quantizing the dense bf16 transformer on the fly, so quality matches the runtime int8/fp8 Dtype path exactly. Same-seed LPIPS vs the bf16 model (lower is better): 0.291 INT8, 0.278 FP8 (50-step suite means; all 28 per-case pairs pass per scheme).

<div> <div style="display: flex; gap: 5px; align-items: center; "> <a href="https://github.com/unslothai/unsloth/"> <img src="https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png" width="133"> </a> <a href="https://discord.gg/unsloth"> <img src="https://github.com/unslothai/unsloth/raw/main/images/Discord%20button.png" width="173"> </a> <a href="https://docs.unsloth.ai/"> <img src="https://raw.githubusercontent.com/unslothai/unsloth/refs/heads/main/images/documentation%20green%20button.png" width="143"> </a> </div> </div>

---

Samples

Prompt: "cute sloth typing on a computer"

<table> <tr> <td align="center"><b>INT8</b></td> <td align="center"><b>INT8</b></td> </tr> <tr> <td><img src="samples/int8_1.png" width="300"></td> <td><img src="samples/int8_2.png" width="300"></td> </tr> <tr> <td align="center"><b>FP8</b></td> <td align="center"><b>FP8</b></td> </tr> <tr> <td><img src="samples/fp8_1.png" width="300"></td> <td><img src="samples/fp8_2.png" width="300"></td> </tr> </table>

---

HiDream-I1 is a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds.

<span style="color: #FF5733; font-weight: bold">For more features and to experience the full capabilities of our product, please visit https://vivago.ai/.</span>

Project Updates

• 🌟 July 16, 2025: We've open-sourced the updated image editing model HiDream-E1.1.
• 📝 May 28, 2025: We've released our technical report HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer.
• 🚀 April 28, 2025: We've open-sourced the instruction-based-image-editing model HiDream-E1-Full. Experience at https://huggingface.co/spaces/HiDream-ai/HiDream-E1-Full!.

Key Features

• ✨ Superior Image Quality - Produces exceptional results across multiple styles including photorealistic, cartoon, artistic, and more. Achieves state-of-the-art HPS v2.1 score, which aligns with human preferences.
• 🎯 Best-in-Class Prompt Following - Achieves industry-leading scores on GenEval and DPG benchmarks, outperforming all other open-source models.
• 🔓 Open Source - Released under the MIT license to foster scientific advancement and enable creative innovation.
• 💼 Commercial-Friendly - Generated images can be freely used for personal projects, scientific research, and commercial applications.

Quick Start

Please make sure you have installed Flash Attention. We recommend CUDA version 12.4 for the manual installation.

pip install -r requirements.txt

Clone the GitHub repo:

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

Then you can run the inference scripts to generate images:

# For full model inference
python ./inference.py --model_type full

# For distilled dev model inference
python ./inference.py --model_type dev

# For distilled fast model inference
python ./inference.py --model_type fast

Note: The inference script will automatically download meta-llama/Meta-Llama-3.1-8B-Instruct model files. If you encounter network issues, you can download these files ahead of time and place them in the appropriate cache directory to avoid download failures during inference.

Gradio Demo

We also provide a Gradio demo for interactive image generation. You can run the demo with:

python gradio_demo.py 

Evaluation Metrics

DPG-Bench

GenEval

HPSv2.1 benchmark

License Agreement

The Transformer models in this repository are licensed under the MIT License. The VAE is from FLUX.1 [schnell], and the text encoders from google/t5-v1_1-xxl and meta-llama/Meta-Llama-3.1-8B-Instruct. Please follow the license terms specified for these components. You own all content you create with this model. You can use your generated content freely, but you must comply with this license agreement. You are responsible for how you use the models. Do not create illegal content, harmful material, personal information that could harm others, false information, or content targeting vulnerable groups.

Acknowledgements

• The VAE component is from FLUX.1 [schnell], licensed under Apache 2.0.
• The text encoders are from google/t5-v1_1-xxl (licensed under Apache 2.0) and meta-llama/Meta-Llama-3.1-8B-Instruct (licensed under the Llama 3.1 Community License Agreement).

Citation

@article{hidreami1technicalreport,
  title={HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer},
  author={Cai, Qi and Chen, Jingwen and Chen, Yang and Li, Yehao and Long, Fuchen and Pan, Yingwei and Qiu, Zhaofan and Zhang, Yiheng and Gao, Fengbin and Xu, Peihan and others},
  journal={arXiv preprint arXiv:2505.22705},
  year={2025}
}

---

Quantized transformer checkpoints (this repo)

This repo adds pre-quantized diffusion transformer checkpoints for HiDream-ai/HiDream-I1-Full, built with torchao dynamic activation quantization from the dense bf16 transformer. The official model card above is unchanged from the source repo.

Files:

• HiDream-I1-Full-INT8.pt (19.7 GB)
• HiDream-I1-Full-FP8.pt (17.1 GB)

Details:

• int8: Int8DynamicActivationInt8WeightConfig (per-token activation, per-channel weight, torch._int_mm). The routed MoE expert Linears quantize cleanly: they only ever see the concatenated image+text token stream, so the torch._int_mm M>16 minimum never binds.
• fp8: Float8DynamicActivationFloat8WeightConfig with PerRow granularity (e4m3, torch._scaled_mm). The loader must floor the dynamic activation scale (activation_value_lb=1e-12 on torchao 0.13+) so all-zero activation token rows cannot produce a zero scale.
• Loading a checkpoint is bit-identical to quantizing the dense bf16 transformer on the fly; the checkpoint skips the dense load and quantize step.
• Validated against same-seed dense bf16 renders over 28 cases per scheme (SSIM, LPIPS-vgg, CLIP delta, non-finite and black-frame checks): all 28 per-case pairs pass per scheme; INT8 LPIPS mean 0.291 / SSIM 0.877, FP8 LPIPS mean 0.278 / SSIM 0.852 (50-step trajectory-divergence band; CLIP delta means 0.007-0.008). INT8 additionally verified bit-identical to on-the-fly quantize across all 1615 state dict tensors (1073 quantized, max abs diff 0.0). torch 2.10 and torchao 0.17.

Samples

Prompt: "cute sloth typing on a computer" (1024x1024, family default steps/guidance, seeds 0-2).

int8

<img src="samples/int8_1.png" width="260"> <img src="samples/int8_2.png" width="260"> <img src="samples/int8_3.png" width="260">

fp8

<img src="samples/fp8_1.png" width="260"> <img src="samples/fp8_2.png" width="260"> <img src="samples/fp8_3.png" width="260">

Pre-cast fp8 text encoder (this repo)

HiDream-I1-Full-text_encoder_4-FP8.pt (8.6 GB) is the pipeline's text_encoder_4 (LlamaForCausalLM, assembled from the standalone unsloth/Meta-Llama-3.1-8B-Instruct mirror the pipeline expects) with the layerwise fp8 storage cast Unsloth Studio applies at load time, saved pre-cast:

• Loading it is bit-identical to downloading the dense encoder and casting on load (verified tensor for tensor: 291 tensors, 225 cast to fp8 storage).
• Cuts the text-encoder download from 16.1 GB to 8.6 GB and halves the fp8 pipeline assembly time (24.3 s vs 48.0 s measured on B200).
• Plain-tensor state dict: loads with torch.load(weights_only=True). Metadata records scheme fp8, component text_encoder_4, base unsloth/Meta-Llama-3.1-8B-Instruct.