Qwen-Image - Txt2Img & Img2Img (Patreon)

Video - https://youtu.be/opE5oNlof2I

Overview

Qwen-Image is here and it seems a bit like an updated version of Flux :) Larger image sizes, better text generation, and better prompt following. What's not to like?

Model links

You can find all the models on https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main or https://modelscope.cn/models/Comfy-Org/Qwen-Image_ComfyUI/files

diffusion model:

- https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_fp8_e4m3fn.safetensors

text encoder:

- https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors

vae:

- https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors

Model Storage Location

```

📂 ComfyUI/

├── 📂 models/

│ ├── 📂 diffusion_models/

│ │ └── qwen_image_fp8_e4m3fn.safetensors

│ ├── 📂 vae/

│ │ └── qwen_image_vae.safetensors

│ └── 📂 text_encoders/

│ └── qwen_2.5_vl_7b_fp8_scaled.safetensors

```

Qwen Image - [https://github.com/QwenLM/Qwen-Image](https://github.com/QwenLM/Qwen-Image)

Custom Nodes

1 Custom requirement - install using ComfyUI Manager:

KJ Nodes - [https://github.com/kijai/ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes)

Also helpful in almost any workflow:

RGthree - [https://github.com/rgthree/rgthree-comfy](https://github.com/rgthree/rgthree-comfy)

Optional - GGUF - city96 (ComfyUI-GGUF nodes)

[https://github.com/city96/ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF)

And the optional Model files -

[https://huggingface.co/city96/Qwen-Image-gguf/tree/main](https://huggingface.co/city96/Qwen-Image-gguf/tree/main)

Prompting Guide

✓ Seems to work:

Quote-Wrapped Text: 'Dream Big' - 95% accuracy

Font Specification: bold serif, script font - Better rendering

Integration Verbs: etched, displayed, handwritten - Natural placement

Concrete Surfaces: billboard, T-shirt, glass board - Clear context

Simple Layouts: Single titles work better than complex arrangements

✗ Not as good:

Vague Text Instructions: "some motivational words"

Complex Multi-Element: Too many text blocks simultaneously

Abstract Placement: "floating text" without surface context

Overly Long Prompts: >200 words

Multiple Languages Mixed: Keep bilingual simple

Prompt example types:

1. Text-Heavy Poster:

"Core text '[TITLE]' in [FONT], tagline '[SUBTITLE]' in [STYLE]; text appears on [SURFACE] with [VISUAL ELEMENTS]. [SCENE DESCRIPTION]. Style: [AESTHETIC] with [LIGHTING]. Ultra HD, cinematic composition"

2. Product/Logo Integration:

"[SUBJECT] wearing/displaying '[TEXT]' in [FONT STYLE]; [ENVIRONMENTAL CONTEXT] with [ADDITIONAL ELEMENTS]. Style: photorealistic with natural lighting"

3. Infographic/Slide:

"[LAYOUT TYPE] featuring '[MAIN TITLE]' as center text, with sections '[TEXT1]', '[TEXT2]', '[TEXT3]' in organised layout. [VISUAL ELEMENTS: icons, colours, arrangement]. Style: clean, professional, modern design"

Advanced prompting

Basically, think like a graphic designer specifying exact placement, fonts, and integration methods. Don't hope for automatic, magic text generation.

English-First: Start prompts in English even for bilingual output

Sequential Text: Specify text placement order: "top to bottom"

Surface Specification: Always define WHERE text appears

Text-Visual Integration: Link text to visual elements explicitly

Font-Context Matching: Match font styles to scene aesthetics

Colour Coordination: Specify text colours relative to background

Iteration Strategy: Start simple, add complexity gradually

Template Reuse: Successful structures work across domains

Pattern Recognition: Note what works for your use cases

Tell A Story Using Denoise

Using Image-2-Image you can also "tell a story" via denoise :)

Example prompt:

"quantum superposition of a cat and a very british rodent wearing a tweed jacket, initial probability waves crystalising, cat-like stable geometry overlays, standing wave style transition, chaotic dyadic-colour background, cinematic photo style, trending on RodentSpace, 4K, UHD, maximum quality."

(Workflow example 4b)

Denoise 0.2 (almost like the original stock image):

Increasing the denoise...

And more... the rodent start to appear! Nice jacket

Ooopsie! Face-replace time

Almost at 0.5 denoise... the crystal passes through

And then finally at 0.5 denoise - cat & rodent:

Workflows

The attached workflows have a number at the start to indicate the order shown in the video. A,b,c simply indicate prompt variation with the same underlying workflow.

1 & 2 - Basic Text-2-Image (CFG1, vs CFG2 default settings)

3, 4 & 5- Image-2-Image with different setting for denoise, style and comparing results

Mix, match, extend and make them your own. Let your imagination go wild!