Qwen Image 2.1 — Editing & Style v3

A flexible ComfyUI workflow for reference-based image editing, style conversion, outfit transfer, body augmentation and head replacement, with optional detailing and upscaling.

V3 brings these tools together in an organized layout with a central output controller, independent post-processing switches, and detailed instructions directly inside the workflow, along with a prompt enhacer.

Choose your main operation

• Simple Editing & Styles: make targeted changes with natural-language prompts or transform the rendering style using editable presets.

• Outfit Swap: transfer clothing from a reference image while asking the model to preserve the base subject’s identity, pose and natural body proportions.

• Body Swap: replace or augment the body using a dedicated reference and the BFS body-swap LoRA.

• Face / Head Swap: use the BFS head-swap LoRA to transfer the head and identity-defining features from a reference.

• Reference Only: send the original image directly to the optional detailers or upscaler.

The main operations are independent alternatives: select one at a time. To combine multiple main edits, reload a result as the base image for the next operation.

Independent post-processing

After the selected operation, the image follows this order:

Main output → General Detailer → Eyes Detailer → Upscale → Edited Image

Each stage has its own ON/OFF switch. Disabled stages pass the previous image through unchanged, so any combination can be used.

Want to refine only the eyes of an existing image? Select Reference Only and enable Eyes Detailer. Want only the editing result? Leave all post-processing switches OFF. With Reference Only and everything OFF, the final preview shows the original image.

References and customization

The workflow provides four reference loaders. Image 1 is the base image; the swap branches use Image 2 as their task-specific reference. Additional references can support the main Editing & Styles operation.

The LoRA loaders in Simple Editing & Styles and General Detailer are intentionally empty, ready for compatible Qwen Image 2.1 LoRAs of your choice.

The supplied swap and general-detailer prompts are recommended starting points and should initially be left unchanged. The eye prompt is deliberately generic and can be refined manually to better preserve iris color, gaze, expression and other specific features.

Resolution is configurable. The default upscaler is 2xNomosUni_span_multijpg_ldl, but it can be replaced with another compatible model.

What is included

• The main workflow with organized groups and brown instruction notes.

• An alternative workflow without Pixaroma dependencies.

• All bespoke routing nodes and the standalone Prompt Studio v2.

• A detailed English guide covering installation, model downloads, reference handling, resolution, controls and troubleshooting.

• Neutral prompt examples, model links and author credits.

• Prompt Enhacer (works with both normal and abliterated text encoders).

Pixaroma nodes are optional, but recommended: the reference switch, timer and resource monitor make the workflow more convenient to use. Without Pixaroma manage reference activation manually.

Requirements and credits

• Qwen Image 2.1, text encoder and VAE: Qwen Team / Alibaba; ComfyUI-ready distribution by Comfy-Org.

• Head-swap LoRA — required for Head Swap: BFS by Alissonerdx.

• Body-swap LoRA — required for Body Swap: BFS by Alissonerdx.

• Outfit LoRA — required for Outfit Swap: Outfit Swap Consistency LoRA by ausboss.

• Eye detector — required for Eyes Detailer: Eyeful v2 – Paired by 32Bitshifter, together with the connected Segment Anything model by Meta AI Research.

• Default upscaler — customizable: 2xNomosUni_span_multijpg_ldl by Philip Hofmann / Phips.

• External nodes: rgthree, ComfyUI Essentials by cubiq, Impact Pack, Impact Subpack, and optionally Pixaroma.

Model weights and external node packages are downloaded separately. Full installation instructions are included in the README.

Usage tips

Use the central Output Controls instead of muting or bypassing processing groups. Start with one operation and optional stages OFF, then enable only the refinements needed.

Clear references and focused prompts usually produce more consistent results. Identity, body proportions and garment details should still be checked carefully: this is generative editing, and exact preservation is not guaranteed.