Fast Local Image Editing! Let's Try HiDream-O1-Image_clear_v2.
- HiDream-O1-Image uses a simple UiT architecture
- HiDream-O1-Image_clear_v2 reduces block noise
- Freely usable MIT license
Introduction
Hello, this is Easygoing.
Today, I'm introducing the HiDream-O1-Image_clear_v2 model, which I've released for fast, local image editing.
HiDream-O1-Image Uses the UiT Architecture
The HiDream-O1-Image model is an image generation and editing AI model released by HiDream.ai in China in May 2026.
About the HiDream-O1-Image Model
Unlike conventional image generation AI models, HiDream-O1-Image processes directly in RGB color space without using a VAE, and adopts the UiT architecture, which integrates a text encoder and a transformer.
Architecture of Recent Image Generation AI Models
| Model | Text Encoder | Transformer | VAE |
|---|---|---|---|
| HiDream-O1-Image | Custom based on Qwen3-VL (8B) (UiT) | - | |
| Flux-3-Image-Klein | Qwen3-VL (4B) | Flux-3-Image-Klein | Flux.2-vae |
| MOSS-Image | Qwen3.5 (9B) | MOSS-Image | Flux.2-vae-small |
| Qwen-Image-2.1 | Qwen3-VL (8B) | Qwen-Image-2.1 | Qwen-Image-2.1-vae |
| Krea 2 | Qwen3-VL (4B) | Krea-2 | Qwen-Image-vae |
HiDream-O1-Image Is Simple and Lightweight
Compared with other recently released image generation AI models, HiDream-O1-Image is especially simple and lightweight.
Image Generation Model Sizes
Key Features of the HiDream-O1-Image Model
- Fast processing
- Strong image editing capabilities
- Natural color expression
In addition to the model itself being lightweight, HiDream-O1-Image requires no model swapping mid-workflow, allowing fast processing.
Also, because HiDream-O1-Image processes both text and image inputs within a single model, combining prompts with reference images feels straightforward and makes image editing easier.
Furthermore, because HiDream-O1-Image processes directly in RGB color space without a VAE, it suffers less color loss and achieves natural color expression.
HiDream-O1-Image-Dev-2604 Is the Official Fast Distilled Model
HiDream.ai has released three official HiDream-O1-Image models:
- HiDream-O1-Image (base model)
- HiDream-O1-Image-Dev (fast distilled model)
- HiDream-O1-Image-Dev-2604 (latest version of the fast distilled model)
HiDream-O1-Image-Dev-2604 can generate illustrations at twice the speed by disabling CFG scale, and requires fewer steps, enabling fast execution.
On CFG Scale and Image Generation Time
Block Noise Appears in HiDream-O1-Image-Dev-2604
As introduced at the beginning, HiDream-O1-Image does not use a VAE, but that doesn't mean it performs no image compression.
HiDream-O1-Image splits images into 32 x 32 pixel patches to speed up processing.
Early-Stage Generation Image in HiDream-O1-Image-Dev-2604
Zooming in on an early-stage image from HiDream-O1-Image-Dev-2604 reveals grid-like block noise in 32 x 32 pixel blocks.
While this kind of block noise gradually clears up in the later stages of generation, using a fast distilled model like HiDream-O1-Image-Dev-2604 can leave block noise remaining in the finished illustration if it isn't removed enough.
Merging the Base Model and the Distilled Model!
So this time, I merged the base model HiDream-O1-Image with the fast distilled model HiDream-O1-Image-Dev-2604 to create the HiDream-O1-Image_clear_v2 model, improving image quality and reducing block noise.
Note that this newly created v2 model is completely separate from the HiDream-O1-Image_clear_v1 model I released previously, and since v2 outperforms v1 in both image quality and generation speed, I recommend using the v2 model going forward.
About the Previous HiDream-O1-Image_clear_v1 Model
Sample Illustrations!
Now let's take a look at actual illustrations.
On the left is the newly created HiDream-O1-Image_clear_v2 model, and on the right is the output from the original HiDream-O1-Image-Dev-2604 model.
Perfume Bottle
Mexican Restaurant
Compared to HiDream-O1-Image-Dev-2604 on the right, the output of HiDream-O1-Image_clear_v2 is cleaner with less overall noise.
Zooming in on the eyes also shows that HiDream-O1-Image_clear_v2 on the left has less block noise and fewer detail breakdowns.
On the other hand, HiDream-O1-Image_clear_v2 loses some of the line strength characteristic of the distilled model, and because contrast rises with lower final noise, the overall illustration tends to look slightly darker, which may come down to personal preference.
Let's Try Image Editing!
As introduced at the start, HiDream-O1-Image's biggest strength is easy image editing through unified processing of image inputs and text prompts, so let's actually test image editing now.
Image Editing Workflow for HiDream-O1-Image_clear_v2
Models (Placement Folder)
- Checkpoints (ComfyUI/models/checkpoints/)
- HiDream-O1-Image_clear_v2_BF16
- HiDream-O1-Image_clear_v2_INT8_ConvRot_HQ (lightweight version)
Custom Node: ComfyUI-uit-hidream-o1
Input Image (Red Car at Night)
Change the Background to Paris at Sunset
change background to paris sunsetChange to an Anime Illustration
change to anime illustrationChange to Black-and-White Manga
change to black and white manga artworkIn every one of these edits, it captures the features of the input image well and edits accurately according to the instructions.
Runs Smoothly Even on Mid-Range GPUs!
Here is what running the workflow above looks like using the lightweight Int8_ConvRot_HQ model on my mid-range GPU (RTX 4060 Ti 16GB).
In addition to the model itself being lightweight, HiDream-O1-Image has no text encoder or VAE processing steps or model swaps, allowing image editing at practical speeds even on mid-range GPUs.
Local Execution, MIT License!
Today, cloud tools like ChatGPT Images 2.6 and Google Nano Banana 2.1 make it easy for anyone to edit images.
However, uploading images to cloud tools involves copying the images, so you need to judge carefully regarding copyright and right of publicity / likeness rights for the images you upload.
By contrast, because HiDream-O1-Image can run locally, everything stays completely on your own PC, letting you experiment with edits freely as many times as you want until you're satisfied.
In addition, the HiDream-O1-Image model is released under the open MIT license, and Alibaba's Qwen3-VL and Google's SigLIP2—on which HiDream-O1-Image is based—are also under the open Apache-2.0 license.
As a model that can run locally with clear licensing, HiDream-O1-Image should be usable with peace of mind even for commercial illustration generation.
Summary: Let's Try HiDream-O1-Image_clear_v2
- HiDream-O1-Image uses a simple UiT architecture
- HiDream-O1-Image_clear_v2 reduces block noise
- Freely usable MIT license
Today I introduced HiDream-O1-Image_clear_v2.
About half a year has passed since the HiDream-O1-Image model was released in May 2026, and no other image generation models using the same UiT architecture as HiDream-O1-Image have appeared yet.
While HiDream-O1-Image falls behind newer models like Krea 2 in image quality and has fewer anime illustration variations, I still respect that a small AI startup completed a model with such a distinctive architecture on its own.
Why not try image editing with the HiDream-O1-Image_clear_v2 model yourself?
Thank you for reading all the way to the end!
About the Video
The video in this article was edited using FilmCraft from the free, open-source app Crafting Apps, which I introduced previously.
About Crafting Apps
Also, the background music in the video was provided from Shiki Peaceful Sounds, the YouTube channel of Shiki , who regularly provides audio tracks.
Shiki Peaceful Sounds
Shiki Peaceful Sounds has many tracks great for background music while working, so please visit and check it out!