🖼️ Qwen Image Outpainter – ComfyUI Workflow
Created by Me – a filmmaker & AI workflow designer. This tool extends your images beyond their original borders, intelligently generating new content that seamlessly blends with the existing scene. Perfect for creating wider shots, changing aspect ratios, or expanding a composition for cinematic framing – all with a single image input.
---
### 🧩 Overview
This workflow takes an input image, expands its canvas in any direction (left, right, top, bottom) by a specified number of pixels, and uses Qwen‑Image 2512 with the LanPaint inpainting technique to fill the new areas with context‑aware content. It then blends the original and generated parts smoothly, producing a larger, cohesive image that feels like it was always that way. The process is guided by your positive/negative prompts and supports a Lightning LoRA for faster generation.
---
### 1️⃣ Required Models & Prerequisites
Make sure these files are placed in your ComfyUI models/ folders:
| Folder | Required Files |
|--------|----------------|
| diffusion_models/ | qwen_image_2512_fp8_e4m3fn.safetensors |
| loras/ | Qwen-Image-2512-Lightning-8steps-V1.0-fp32.safetensors (recommended for speed) |
| text_encoders/ | qwen_2.5_vl_7b_fp8_scaled.safetensors |
| vae/ | qwen_image_vae.safetensors |
> ⚠️ VRAM Requirement: ~10–12 GB recommended – the workflow uses the full Qwen‑Image pipeline with LanPaint sampling.
---
### 3️⃣ How It Works (Outpaint in 4 Steps)
1. Load your image – any image you want to expand.
2. Set expansion values – define how many pixels to add to the left, right, top, and bottom. The workflow uses ImagePadForOutpaint to create a padded canvas with a feathered edge mask for smooth blending.
3. Generate new content – the padded image is VAE‑encoded, a noise mask is applied to the empty areas, and LanPaint_KSampler generates new content based on your positive and negative prompts. The Lightning LoRA speeds up the process (fewer steps).
4. Blend & save – the original and generated images are blended using LanPaint_MaskBlend to ensure a seamless transition, and the final expanded image is saved with your prefix.
---
### 6️⃣ Advanced Settings (Optional)
- Expansion values – individually control left, right, top, and bottom padding (in pixels).
- SEED – fix the seed for reproducible results.
- Positive Prompt – describe the scene you want to generate in the new areas (e.g., "A breathtaking landscape, dramatic golden hour lighting...").
- Negative Prompt – exclude unwanted elements (e.g., "low quality, bad anatomy...").
- Paint Force – adjust the LanPaint_NumSteps (default 4) to control how aggressively the model fills the new areas – higher values give more creative freedom but may diverge from the original.
- Model Shift – the ModelSamplingAuraFlow shift is set to 3.5 for optimal outpaint quality; you can fine‑tune it.
---
### 📦 Final Output
After running, you’ll get a larger, extended version of your original image saved as Qw_IO_*.png – ready for use in wider shots, different aspect ratios, or creative compositions.
---
Why I built this: In filmmaking and concept art, I often need to expand a reference image to fit a new framing or to visualise a wider scene. This outpainter gives me that flexibility instantly, without having to reshot or manually paint in the missing areas. It’s a powerful tool for pre‑visualisation, storyboarding, and even final asset creation. I hope it opens up new compositional possibilities for you too! 🎨🖼️
---
📬 Contact: If you’d like to donate, ask a question, or order a custom workflow, reach out to me on Civitai (SaeedKhalili), visit my website SaeedKhalili.com, or email me at [email protected]. I’m always happy to help fellow creators!