HiDream-O1-Image-Dev-2604 — FP8 Mixed (ComfyUI)

This is the FP8 mixed-precision quantization of HiDream-ai/HiDream-O1-Image-Dev-2604 — the updated distilled Dev variant of HiDream-O1-Image released May 13, 2026 — for use with ComfyUI. At only ~10 GB VRAM and 28 steps, this is the most accessible way to run the latest HiDream O1 Dev update locally. 9B parameters.

⚠️ PyTorch 2.9.x is not recommended — known compatibility issues exist. Use 2.8.x or earlier.

⚠️ Editing note: For instruction-based image editing tasks, the upstream team recommends using the full model instead of Dev.

Custom ComfyUI Node: Saganaki22/HiDream_O1-ComfyUI

What's New in Dev-2604

Accelerated IP inference — faster subject-driven personalization
Layout conditioning — place subjects at specific bounding box regions
Skeleton conditioning — OpenPose-based pose control for try-on and character workflows
Updated editing scheduler — improved Dev editing behaviour

Dev vs Full — Key Differences

	Full Model	Dev-2604 (this repo)
Parameters	9B	9B
Inference Steps	50	28
Guidance Scale (CFG)	5.0	0.0 (disabled)
Shift	3.0	1.0
Scheduler	FlowUniPCMultistepScheduler	FlashFlowMatchEulerDiscreteScheduler
Speed	Slower, more detail	~2× faster

CFG is disabled in Dev mode — negative prompts have no effect.

VRAM Requirements

Precision	Approximate VRAM
BF16	17 – 20 GB
FP16	17 – 20 GB
FP8 Mixed (this repo)	~10 GB

This is the recommended variant for GPUs with less than 16 GB VRAM. Combined with the Dev model's 28-step schedule, it is the lowest-cost way to run HiDream O1 Dev-2604 — roughly 2× faster and half the VRAM of the full BF16 model.

What is FP8 Mixed? Weights are stored in float8_e4m3fn. Sensitive layers (norms, embeddings, output heads) retain higher precision for stability. On RTX 40xx / H100 (Hopper/Ada), FP8 compute is hardware-accelerated. On older GPUs, weights dequantize on-the-fly — still saving VRAM, with a small speed penalty. Do not set config.json dtype to float8_e4m3fn; keep it as bfloat16 — the node detects FP8 from the safetensors tensors directly.

Quick Start — ComfyUI

1. Install the Custom Node

cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/HiDream_O1-ComfyUI.git
cd HiDream_O1-ComfyUI
python -m pip install -r requirements.txt

Or search for HiDream O1 in ComfyUI Manager.

Suggested transformers version: 4.57.1 – 5.3 (newer versions may break compatibility).

2. Download the Weights

Download the entire model folder (all files, not just the safetensors) and place it in ComfyUI/models/diffusion_models/:

huggingface-cli download drbaph/HiDream-O1-Image-Dev-2604-FP8 \
    --local-dir ComfyUI/models/diffusion_models/HiDream-O1-Image-Dev-2604-fp8

The folder must contain the full Hugging Face support files alongside the weights: config.json, chat_template.json, generation_config.json, preprocessor_config.json, tokenizer.json, tokenizer_config.json, vocab.json, merges.txt, model.safetensors

3. Load in ComfyUI

Use the workflow provided in the custom node repository. The loader will detect dev in the folder name and automatically apply Dev settings (28 steps, no CFG, Euler scheduler). Point the model loader to HiDream-O1-Image-Dev-2604-fp8.

For the fastest inference on supported hardware, set precision to fp8_e4m3fn_fast in the model loader node.

About HiDream-O1-Image

HiDream-O1-Image is a natively unified image generative foundation model built on a Pixel-level Unified Transformer (UiT) — no external VAEs, no disjoint text encoders. It encodes raw pixels, text, and task-specific conditions in a single shared token space, supporting:

Text-to-image generation up to 2,048 × 2,048
Instruction-based image editing (full model recommended)
Subject-driven personalization with layout and skeleton conditioning
Long-text and multilingual text rendering

It debuted at #8 in the Artificial Analysis Text to Image Arena (2026-05-05).

Key Features

🧬 Pixel-Level Unified Transformer — end-to-end on raw pixels, no VAE, no disjoint text encoder
🎨 One Model, Many Tasks — T2I, editing, personalization, layout, skeleton, storyboard
⚡ 28-Step Distilled Dev — ~2× faster than the full model
💾 FP8 Quantized — ~half the VRAM of full-precision variants
🖼️ Native High Resolution — direct synthesis up to 2,048 × 2,048
🧍 Skeleton & Layout Conditioning — OpenPose control and bounding-box subject placement

All Model Variants

Full Model

Repo	Precision	VRAM	Steps
drbaph/HiDream-O1-Image-BF16	BF16	17–20 GB	50
drbaph/HiDream-O1-Image-FP16	FP16	17–20 GB	50
drbaph/HiDream-O1-Image-FP8	FP8 Mixed	~10 GB	50

Dev Model (original)

Repo	Precision	VRAM	Steps
drbaph/HiDream-O1-Image-Dev-BF16	BF16	17–20 GB	28
drbaph/HiDream-O1-Image-Dev-FP16	FP16	17–20 GB	28
drbaph/HiDream-O1-Image-Dev-FP8	FP8 Mixed	~10 GB	28

Dev-2604 Model (updated, this series)

Repo	Precision	VRAM	Steps
drbaph/HiDream-O1-Image-Dev-2604-BF16	BF16	17–20 GB	28
drbaph/HiDream-O1-Image-Dev-2604-FP16	FP16	17–20 GB	28
drbaph/HiDream-O1-Image-Dev-2604-FP8 (this repo)	FP8 Mixed	~10 GB	28

License

The original HiDream-O1-Image model and code are released under the MIT License. This FP8 quantization inherits the same license.

drbaph
/

HiDream-O1-Image-Dev-2604-FP8