You have a face. An outfit. A scene. You want all three combined into one image, fast enough to iterate in real time.
Up to three reference images in. One composed result out. In seconds.
Run it now on Floyo!
Why This Workflow
Most image editing workflows take one image at a time. You edit the face. Then the outfit. Then the background. Three separate runs, three separate outputs to reconcile.
This workflow takes all three at once. The Lightning LoRA drops generation to 4 steps. You get a composed result at 1920x1080 in seconds, fast enough to run a dozen variations before a standard workflow finishes one.
up to three reference images combined in one pass
4-step Lightning LoRA, results in seconds, not minutes
1920x1080 default output
character identity stays consistent across the composite
single-image edits work too, multi-image is optional
How It Works
Enable the image slots you need. Load your references. Write a prompt describing the final image. The model reads all three inputs simultaneously, fuses them into a coherent composition, and outputs a single result.
One image gives you a face. Another gives you wardrobe. The third gives you a scene. The model handles the composition, lighting match, and subject consistency across all three.
Key Inputs
Image Slots
Three slots wired in, all bypassed by default. Enable as many as you need:
1 image: single-image edits, face fix, outfit change, background swap
2 images: subject onto a new background, or face with a new outfit
3 images: full character + outfit + scene composite
Match the Image Mode setting (1_image, 2_image, or 3_image) to the number of slots you've enabled. Mismatching this causes the model to misread the input setup.
Prompt
Describe what you want the final image to look like. Be specific about subject, environment, lighting, and framing.
Examples:
"woman with curly dark hair standing in a harbor with pastel-colored houses behind, cinematic, natural light""subject from image 1 wearing the outfit from image 2, clean studio background, professional photography""product from image 2 placed in the lifestyle scene from image 3, warm natural light, editorial style""character in image 1, wearing jacket from image 2, urban night street from image 3, neon reflections, wide shot"
The more concrete the description lighting, framing, mood, the more accurate the composite.
Resolution
1920x1080: default, works for most edits
1280x720: faster output for iteration
1024x1024: square portrait format
Steps: Default 4. The Lightning LoRA is tuned for 4 steps. Pushing past 6–8 with the LoRA active hurts quality. If you want more steps, swap out the Lightning LoRA and run base Qwen Image Edit 2511 at 20+ steps.
CFG: Default 1. Lightning LoRAs are trained for low guidance. Above 2, you get burned colors and posterized outputs. Leave it at 1.
LoRA Strength: Default 1.0 for full Lightning speed. Drop to 0.7–0.8 and raise steps to 6–8 if you want more output variation at a slight speed cost.
Shift: Default 3. Lower (1–2) brings out more fine detail. Higher (4–5) gives softer, smoother results. Adjust if outputs feel too sharp or too washed out.
What This Is Great For
Character compositing: Combine a specific face, a specific outfit, and a specific scene into one image without three separate editing passes.
Fashion and lookbook: Place a subject from one reference into an outfit from another against a scene from a third. Generate multiple outfit combinations from the same face reference in one session.
Product lifestyle shots: Blend a product shot into a lifestyle scene while keeping the product accurately represented.
Rapid iteration: 4 steps means you can test a dozen prompt variations, scene references, or outfit combinations in the time a standard workflow runs once. Find the right composition before committing to a full-quality render.
Single-image edits: Face fixes, background swaps, and outfit changes on one image. The multi-image setup is optional.
What to Watch Out For
Keep CFG at 1. This is the most common mistake with Lightning LoRAs. Raising it above 2 burns the output. The model is not undertrained, it's calibrated for low guidance.
Match image_mode to the number of enabled slots. If you enable two images but leave image_mode on 3_image, the model expects a third reference and the output will be off.
4 steps is the sweet spot, not a limitation. Going higher with the Lightning LoRA active degrades quality. For maximum detail on a final hero image, disable the Lightning LoRA and run base Qwen Image Edit 2511 at 20+ steps instead.
Very complex three-image composites with hard lighting mismatches between references need more iteration. The model reconciles lighting differences automatically but extreme contrasts (bright outdoor reference + dark studio reference) take a few runs to balance.