Announcing HiDream-I1-Dev_clear! Comparing Composition with Flux.1, AuraFlow, and Qwen-Image

Introduction

Hey everyone, it’s Easygoing.

This time, I’m excited to introduce HiDream-I1-Dev_clear, a fine-tuned version of HiDream-I1-Dev that I’ve just released.

What is HiDream?

HiDream is an image generation AI model that was released on April 8, 2025.

gantt
   title Stable Diffusion 3 and Derivative Models
   dateFormat YYYY-MM-DD
   tickInterval 6month
		Stable Diffusion 3 : done, 2024-06-12, 2025-11-01
        AuraFlow : done, 2024-07-12, 2025-11-01
		Flux.1   : done, 2024-08-01, 2025-11-01
		HiDream   : 2025-04-06, 2025-11-01
		Qwen-Image   : done, 2025-08-04, 2025-11-01

Among locally runnable image generation models, HiDream stands out as one of the largest, boasting exceptional prompt comprehension and expressive power.

Creating HiDream-I1-Dev_clear !

HiDream-I1-Dev is a version optimized for speed by disabling CFG scale, making it runnable on mid-range GPUs on a local PC.

About HiDream

This time, I further fine-tuned HiDream-I1-Dev to produce clearer outputs.

Specifically, I reduced the influence of layers that seemed to overreact during image generation, resulting in cleaner, less ambiguous illustrations.

Sample Illustrations!

Let’s take a look at some actual outputs.

On the left is the newly tuned HiDream-I1-Dev_clear, and on the right is the original HiDream-I1-Dev.

Japanese Garden and Kimono-Clad Woman

Anime Illustration (Image-to-Image)

Close-Up View

Cleaner Illustrations!

In all examples, the left-side images (HiDream-I1-Dev_clear) are noticeably clearer, with fewer distortions in details even when zoomed in.

HiDream-I1-Dev_clear likely simplifies backgrounds due to faster noise removal, but the level of detail can be controlled via prompts or the ModelSamplingSD3 node.

Comparison with Flux.1, Qwen-Image, and AuraFlow!

Now, let’s generate six images in a row using the same prompt and compare compositional variety across different AI models.

HiDream-I1-Dev_clear

Being developed in China, HiDream-I1-Dev_clear tends to produce illustrations with a subtle Chinese aesthetic, but it excels in artistic expression and color richness.

It faithfully follows prompts—even long natural language ones—and generates delicate, nuanced illustrations.

However, compositional variety is somewhat limited; multiple generations often yield similar layouts.

FluxesCoreDev_v1.0 (Flux.1 [dev])

FluxesCoreDev_v1.0 Model is a custom version of Flux.1 [dev] created by konapieces, excelling particularly in Japanese character rendering.

The Flux.1 [dev] model supports 1440×1440 high-resolution output and delivers convincing, bold visuals with a sharp, high-contrast style.

It produces more diverse compositions than HiDream-I1-Dev_clear, but its strong visual identity can sometimes result in familiar-looking outputs.

Qwen-Image

Qwen-Image has higher hyperparameters than HiDream but falls slightly behind in composition and artistic expression compared to the previous two models.

I suspect this is because its text encoder, Qwen-2.5-VL, was primarily developed for business applications, with limited training in artistic domains.

While it responds strongly to specific keywords like “laugh” or “smile,” it struggles to reproduce subtle, nuanced prompts in long natural language inputs.

That said, Qwen-Image’s true strength lies in image editing via the Qwen-Image-Edit model, where its simpler prompt understanding proves useful.

AuraFlow_v0.3

As the oldest model in this comparison, AuraFlow_v0.3 lags in overall quality, but it shines in compositional variety and unique artistic flair, especially in color usage.

On October 27, 2025, the Pony V7 model—built on AuraFlow—also made waves.

While Pony V7 has received criticism for issues like limb distortions, it’s important to recognize that, unlike highly polished SDXL-based predecessors, it retains AuraFlow’s distinctive character.

Side-by-Side Comparison!

Here’s a summary table of each model’s characteristics:

HiDream and Flux.1 disable CFG scale for faster processing, resulting in roughly half the generation time of Qwen-Image and AuraFlow.

However, this comes at the cost of losing Negative Prompt support.

About CFG Scale and Negative Prompts

For photorealistic outputs, Flux.1 leads in both composition and quality, with HiDream as a close second.

However, Flux.1’s widespread use can make its outputs feel familiar, and its license restrictions make it less ideal for commercial use.

About Image Generation AI Model Licenses

The HiDream-I1-Dev_clear model I’ve released combines HiDream’s prompt fidelity and delicate expression under a permissive license, while aiming to approach Flux.1’s quality.

I encourage everyone to give it a try!

Workflow

Here’s the ComfyUI workflow used for HiDream-I1-Dev_clear:

• HiDream-I1-Dev_clear_20251101.json

All models are available for download here:

This workflow offloads text encoder processing to the CPU, requiring the ComfyUI-easygoing-nodes custom node.

Summary: Try HiDream-I1-Dev_clear !

• Released HiDream-I1-Dev_clear
• Excellent natural prompt understanding and delicate expression
• Next-gen models offer diverse compositional variety

I fine-tuned the HiDream-I1-Dev model this time.
HiDream is an openly modifiable and commercially usable image generation AI.

While it may not match the polished refinement of heavily tuned Flux.1 models, its raw potential feels full of room for growth.

Image generation is all about how the AI interprets your prompt—and HiDream’s prompt adherence makes creating illustrations genuinely fun. I hope you’ll give it a try!

Thanks for reading to the end!