ComfyUI Workflow: Krea2 and Qwen3-VL for Reference Style Inpainting
TBG______ · reddit · 2026-08-12
A developer shared a powerful workflow discovered during ETUR testing that leverages Krea2 and the Qwen3-VL model for reference-style inpainting in ComfyUI.
- Core Features: This node enables reference-style inpainting while preserving the global semantics of the original image, also improving Differential Diffusion results. It recommends lowering the influence of Qwen3-VL encoded images and keeping denoise strength below 0.8.
- Inputs & Control: Users provide a Base Image (with mask) and a Reference Image (with a ref-mask). Inputs are encoded differently based on the VL Source Mode, with strength values to flexibly control the weight of each input during generation.
- Use Cases: Ideal for prompt-based inpainting fused with reference styles, though not intended for exact image placement.
The node will be included in the upcoming TBG ETUR 1.12.12 custom node update.
More from Multimodal
- Comfy MCP Supports Local ComfyUI: Agents Take Over Video Workflows — Comfy-Org · 2026-08-12
- Testing MiniMax Motion 03: A ComfyUI Workflow for Seamless Video Stitching — Cold_Zone332 · 2026-08-12
- Creator Showcases AI-Generated Short Film Clip Made with Grok and Imagine — tetsuoai · 2026-08-12
- Lightricks Announces Open Weights for LTX Video 2.5 Today — SanDiegoDude · 2026-08-12
- Maintaining Character & Shoe Consistency: Advanced AI Video Prompt — HeyZoyaKhan · 2026-08-12
- MirrorWorld: Taming Video Diffusion Models for Realistic Mirror Reflections — CityUniversityofHongKong · 2026-08-12