Qwen Image Edit 2509 · Build a LoRA Dataset

ABOUT THE WORKFLOW

Generate a LoRA Training Dataset
Upload one reference image of a character. Write a list of scene descriptions, one per line, each describing a different view, pose, or setting while preserving identity. Qwen Image Edit 2509 generates the character in each described scene, keeping their face, hair, eye color, and design consistent. Each output saves automatically as an image and caption pair in the format required for LoRA training. One reference image in, a full dataset out.

Model

• Qwen Image Edit 2509 (fp8) by Alibaba. A multi-reference image editing model paired with a Lightning LoRA for 4-step generation and an anime illustration LoRA for consistent anime-style character rendering across views.

HOW IT WORKS

Step 1. Upload your character reference image
One clear, well-lit photo or illustration of the character you want to train a LoRA on. The model preserves this character's identity across every generated view.
Works great with: portrait photos · character illustrations · anime designs · AI-generated characters

Step 2. Set the character trigger word
Enter the trigger word that will activate the trained LoRA during generation (e.g., "Chl0e"). This name gets embedded in each caption file alongside the description.

Step 3. Write your scene list
Enter one scene description per line in the prompt list. Each line generates one training image. "Anime-style left profile portrait of the same character standing on a rooftop at sunset, wind blowing through her long brown hair, preserving exact facial identity, blue eyes, hair clip, and character design." Repeat key identity anchors in every line.

Step 4. Hit run
Qwen Image Edit generates each scene in sequence at 512x512. Each output saves as a PNG with a matching TXT caption file, named with the trigger word and description.

Step 5. Use the dataset for LoRA training
Take the saved image and caption pairs and feed them into any LoRA trainer (Z-Image LoRA, Kohya, SimpleTuner, etc.).
Ready for: Z-Image LoRA Trainer · Kohya SS · SimpleTuner · Comfy LoRA trainers

First time? Upload your character image, set the trigger word, write 10 to 20 scene descriptions, and hit run. The workflow handles the rest.

RECOMMENDED SETTINGS

Quick-start guide. Find the goal that matches yours and copy the settings.

• Standard dataset generation — 512x512, 4 steps, CFG 1, Lightning LoRA at 1.0, anime LoRA at 1.0. Write 10 to 20 scene descriptions and run.

• More diverse dataset — Write descriptions that vary the angle (front, left, right, three-quarter, back), the setting (interior, exterior, urban, nature), the lighting (golden hour, night, overcast), and the expression (neutral, smiling, focused, emotional). Diversity in the training set produces a more generalizable LoRA.

• Include the trigger word in every caption — The workflow embeds the trigger word automatically in each caption file. Keep the same trigger word across all runs so the LoRA associates it consistently.

• Non-anime characters — Remove the anime illustration LoRA from the pipeline (set its strength to 0). The base Qwen Image Edit model handles photorealistic characters without the anime adapter.

• Higher resolution output — Change the ImageScaleToTotalPixels target above 1MP. 512x512 is standard for most LoRA trainers, but some trainers accept higher resolution. Match the output to your trainer's requirements.

• Repeat identity anchors in every caption — Write "preserving exact facial identity, blue eyes, brown hair, hair clip, and character design" at the end of every scene description. Consistent captions improve LoRA training signal.

Prompt: Each line in the prompt list is one training image. Write it as a complete scene description that includes: art style, scene context, character identity anchors, and any specific visual detail. "Anime-style three-quarter view of the same character sitting at a desk studying, warm desk lamp light, preserving exact facial identity, blue eyes, hair clip, slight frown of concentration" is complete. "Character at a desk" is not.

LEARN

📹 Videos

• Intro to Floyo

• ComfyUI 101 Free Course ft. Sebastian Kamph

• Floyo 101 for Team Collaboration

✨ Quick links

• Help docs

• LoRA Training guide

• Discover more workflows

USE CASES

👤 Anime Character LoRA Datasets
Generate 20 to 50 views of an original anime character from a single reference illustration, covering all the angles, expressions, and settings a LoRA trainer needs for a strong concept.

🎮 Game Character Training Data
Build a training set for a game character from a single concept art or screenshot, with varied poses, environments, and camera angles for a generalist character LoRA.

📸 Portrait-to-Anime LoRA Preparation
Upload a real person's photo and generate an anime-style dataset of that person across multiple scenes, ready to train a personal anime LoRA.

🖼️ Consistent Character Illustration Sets
Generate a multi-scene illustration set from one character design for webtoons, novels, or presentation decks, with matching caption files for future fine-tuning.

WHAT WORKS BEST / WHAT TO AVOID

✅ Works great

• Clear, well-lit reference images with visible facial features

• 15 to 30 diverse scene descriptions covering multiple angles and settings

• Repeating identity anchors in every scene caption

• Trigger words that are unique strings unlikely to appear in pretraining (e.g., "Chl0e," "Xyrn," "Tobek")

⚠️ May produce softer results

• Reference images with heavy occlusion, extreme angles, or low resolution

• Duplicate or near-identical scene descriptions (the dataset will lack diversity)

• Common words as trigger names (e.g., "girl," "character") that overlap with model pretraining

• Very long descriptions with too many competing visual targets

FAQ

What is a LoRA dataset and why does it need image-caption pairs?
A LoRA dataset is the set of images and text captions used to fine-tune a model on a specific concept, character, or style. Each image needs a matching caption that describes what is in it, including the trigger word, so the trainer associates the trigger with the visual concept. This workflow generates both the images and the captions automatically.

How many training images do I need for a character LoRA?
Most trainers produce good results with 15 to 50 images for a character LoRA. Fewer than 15 may underfit. More than 100 with redundant views can overfit. Prioritize diversity: different angles, expressions, lighting, and settings over quantity.

Why repeat identity anchors in every caption?
The LoRA learns associations between the caption text and the image content. If every caption mentions "blue eyes, brown hair, hair clip," the model learns to associate those features with the trigger word. Inconsistent captions produce weaker identity lock in the trained LoRA.

What trigger word format works best?
Use a unique string that does not appear in the model's pretraining vocabulary. "Chl0e" (with a zero), "xyr3n," or "tok_person" work better than "girl" or "character," which are too common. The trigger activates the LoRA during generation when included in a prompt.

Can I use this workflow for non-anime characters?
Yes. Set the anime illustration LoRA strength to 0 to remove the anime style filter. The base Qwen Image Edit 2509 model handles photorealistic and stylized characters without the anime adapter.

Is Qwen Image Edit 2509 licensed for commercial use?
Yes. Qwen Image Edit 2509 is open-source by Alibaba. The Lightning LoRA and anime illustration LoRA have their own license terms. Check each component's license on its model page for commercial use in your specific project.

How to run LoRA dataset generation online?
You can run LoRA dataset generation online through Floyo. No installation, no setup, no local GPU needed. Open the workflow in your browser, upload your reference image, write your scene list, and hit run. Free to try.

WHY FLOYO?

Floyo is the only platform with team collaboration for ComfyUI in the browser. You run workflows with no install. You share run history, assets, and models across your team. You pay only when you generate. Floyo supports open-source and closed-source models.

A designer runs an edit and likes the result. A teammate opens that exact run from shared history and keeps going. No file handoffs. No version confusion.

For studios and enterprise teams, Floyo adds private workspaces, pooled resources, and a team usage dashboard. Other ComfyUI cloud tools run for one person at a time. Floyo runs for the whole team, with transparent per-generation costs.

Ready to try it?
Upload your character reference, set the trigger word, write your scene list, and hit run.

→ Launch Workflow, Free

Questions? Watch the free course or check the FAQ above.