Qwen 2511 · Single Image to Character Dataset

ABOUT THE WORKFLOW

Generate a LoRA Training Dataset from One Photo
Upload a single image of any character. Write a list of scene and pose descriptions, one per line. Qwen Image Edit 2511 generates the character in each described scene, preserving their face, hair, outfit, and proportions across every output using the FluxKontext multi-reference latent method. Each image saves automatically with a matching caption file. One reference photo in, a complete training dataset out.

Model

• Qwen Image Edit 2511 (bf16) by Alibaba. A multi-reference image editing model paired with the Lightning LoRA for 4-step generation. Uses FluxKontext multi-reference latent conditioning for strong character identity lock across diverse poses and settings.

HOW IT WORKS

Step 1. Upload your character reference image
One clear photo, illustration, or render of the character you want to train a LoRA on. Qwen Image Edit 2511 reads the identity from this image and carries it into every generated view.
Works great with: portrait photos · character illustrations · anime designs · AI-generated characters · product shots

Step 2. Set the character name (output folder)
Enter a name for the output folder. All images and caption files save here.

Step 3. Edit the scene list (optional)
The default list covers 10+ views including city front-facing, park bench side view, forest trail rear view, beach run, wall lean, classroom, rainy street, and more. Each line in the list generates one training image. Edit, add, or remove lines to match your target use case.

Step 4. Hit run
Qwen Image Edit 2511 generates each scene in sequence at 1.6MP in 4 steps. Each output saves as a PNG with a matching TXT caption file.

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

First time? Upload your character image, set a folder name, and hit run. The default scene list generates a usable starting dataset immediately.

RECOMMENDED SETTINGS

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

• Standard character dataset — 1.6MP output, 4 steps, CFG 1, Lightning LoRA at 1.0, seed randomized. Upload your reference, set the folder name, run with the default scene list.

• More diverse dataset — Add more lines to the scene list. Cover the full range: front, back, left, right, three-quarter, low angle, high angle, close-up, full body, action pose, seated, walking, running. Diversity in the training set produces a more generalizable LoRA.

• Add a style prefix to all captions — Put style keywords in the prepend_text field. "Anime style, detailed illustration," at the start of every prompt keeps all generated images in the same aesthetic.

• Add a trigger word to all captions — Put your trigger word in the prepend_text field. "TOK character," at the start associates the trigger with the character across the full dataset.

• Reproduce a specific view — The seed is randomized by default. Set it to a fixed value to reproduce the same output when refining a particular scene description.

• Non-anime characters — The workflow works with photorealistic and stylized characters equally. Write scene descriptions that match the character's visual style rather than specifying "anime" if the reference is photorealistic.

Prompt: Write each line as a plain instruction starting with "make this character." Include pose, setting, expression, camera angle, and framing. "Make this character walk confidently down a forest trail, arms swinging naturally, focused expression, rear view" is specific. "Character in a forest" 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

👤 Character LoRA from a Single Photo
Generate a full multi-view training dataset from one portrait or character image, covering the diverse poses, settings, and angles a LoRA trainer needs for a strong concept.

🎮 Game and VTuber Character Datasets
Build a training set for a game character, VTuber avatar, or original character from a single reference, covering front, back, side, and action views.

📸 Personal Portrait LoRA
Upload a photo of a person and generate a diverse set of that person in different environments and poses for a personal AI avatar or portrait LoRA.

🎨 Illustrated Character Training Sets
Turn a single piece of character art into a multi-angle dataset. The FluxKontext conditioning preserves illustrated style and identity across all generated views.

WHAT WORKS BEST / WHAT TO AVOID

✅ Works great

• Clear, well-lit reference images with visible face, outfit, and body

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

• Each description specifying pose, setting, expression, and camera angle

• Both photorealistic and illustrated character references

⚠️ May produce softer results

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

• Scene descriptions with no pose or camera direction

• Very short scene lists (fewer than 10 views limit training diversity)

• Duplicate or near-identical scene descriptions that produce redundant training images

FAQ

How is this different from the Flux Kontext character dataset workflow?
Both workflows generate multi-view character datasets from a single reference image. This workflow uses Qwen Image Edit 2511 (Alibaba) with the FluxKontext multi-reference latent method and a Qwen-specific text encoder. The Flux Kontext workflow uses FLUX.1 Dev Kontext (Black Forest Labs) with dual CLIP-L and T5-XXL encoders. The core dataset output and scene list format are identical between both. Choose based on which base model produces better identity preservation for your specific character reference.

What is the FluxKontext multi-reference latent method?
A conditioning technique that encodes the reference image into the latent space with index_timestep_zero, allowing the model to read the character's identity at the structural level rather than as a pixel overlay. This produces stronger identity lock across diverse poses and settings compared to standard image conditioning.

How many training images do I need for a character LoRA?
Most LoRA trainers produce strong results with 15 to 30 images for a character concept. Fewer than 10 may underfit. More than 50 with redundant views can overfit. Prioritize diversity over quantity.

Does the workflow include a negative prompt?
Yes. The negative prompt guards against: face change, identity drift, broken anatomy, extra fingers, wrong scale, cutout look, pasted subject, flat lighting, harsh shadows, low realism, and blur. It is already configured and does not need editing for standard use.

How do I add a trigger word to the captions?
Put your trigger word in the prepend_text field on the CR Prompt List node. "Chl0e, " at the start adds the trigger to every generated caption automatically.

Is Qwen Image Edit 2511 licensed for commercial use?
Yes. Qwen Image Edit 2511 is open-source by Alibaba. The Lightning LoRA has its own license terms. Check each component's license on its model page for commercial use in your specific project.

How to run character LoRA dataset generation online?
You can run character LoRA dataset generation online through Floyo. No installation, no setup, no local GPU needed. Open the workflow in your browser, upload your reference image, 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 image, set the output folder name, and hit run. The default scene list generates your first dataset.

→ Launch Workflow, Free

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