ComfyUI / Workflow template
A two-character scene needs more than “keep the face.” This ComfyUI template connects two separate LoadImage nodes to Qwen Image 2.1 Edit, so you can describe which reference belongs on the left, which belongs on the right and what the scene should change. Build romantic story frames, visual-novel panels or a shared creator-gallery cover.
Import the JSON, select your credentials and fill the required inputs before running.
Both references enter one edit request. They are not treated as two independent generation jobs. Give each character a distinct hairstyle, outfit and position, then review both faces in the composed result. Save a usable two-person frame as the basis for the next scene or a focused refinement.
How the workflow runs
Load character A
Load character B
Compose the scene
Save the image
NSFW / Character / Style
BL/GL story frames, paired character portraits and original adult visual novels.
What to change first
Two independent LoadImage inputs feed image_urls.image_0 and image_urls.image_1. A single LoadImage batch is a different input arrangement.
Explore a character’s wardrobe without rebuilding the entire scene: a fitted evening look, a themed costume or a different fabric treatment. Reference editing lets you compare the creative change against the original.
What to change first
Describe the garment, color and fabric explicitly. Keep pose and background steady for the first comparison.
Create a recognizable male character across relaxed portraits, silk-shirt editorials and night-out looks. Keep a few signature traits so the gallery feels like one creator rather than a collection of unrelated faces.
What to change first
Keep the same reference and hair description. Change lighting or wardrobe first, then try a different framing.
Workflow tutorial
Install the SpicyAPI custom nodes, restart ComfyUI and drag in the JSON. Add your API key in Settings → SpicyAPI. The graph has two LoadImage nodes, one Qwen edit node and one SaveImage node. It uses the hosted Qwen Image 2.1 Edit endpoint.
Load the first fictional adult character in Reference A and the second in Reference B. Prefer readable faces and distinct hair or clothing. Two near-identical references are difficult to describe unambiguously. The template preserves their socket order: A is image_0 and B is image_1.
Name “the character from reference 1” and “the character from reference 2,” assign left/right positions, and say which hair and outfit belong to each. Next specify the setting, camera framing and interaction. Begin with a two-person medium shot and a simple scene; add elaborate lighting after the placement is usable.
The starter uses 3:4, 1k, opaque background, PNG and a fixed seed. Prompt expansion is disabled so the initial instruction stays easier to compare. Keep those settings while revising one character instruction. For a wider scene, try a landscape ratio; use 1.5k or 2k after you have a composition worth keeping.
Queue one execution and wait for the task. SaveImage writes the IMAGE output with the spicyapi/qwen-duo prefix in your configured output directory. Check left/right assignment, face shape, hair, hands and whether both characters remain visible. Use the selected result as a future composition reference when you want to refine that frame.
Creative example. Adapt it to your own fictional adult character and source image.
Compose a medium two-person shot of both fictional adult characters in an elegant jazz lounge. Place the copper-haired character from reference 1 on the left in her emerald evening dress. Place the dark-haired character from reference 2 on the right in her plum tailored suit. Preserve each person's distinct face and hairstyle. Relaxed affectionate eye contact, both clearly visible, warm cinematic lighting.Try this directionChoose Qwen Image 2.1 Edit to combine separate character references into one frame. If you want to animate the scene next, use the composed image as the starting frame for the H3 workflow.
The template submits one edit task carrying two reference images. Its quote reflects the actual inputs and selected settings; check the live model price rather than assuming the two-reference job costs the same as one reference. Configure Max Cost Per Run before the first queue. A fixed seed helps organize comparisons. Changing the prompt, references or settings creates a new paid edit even with the same seed.
Strengthen the distinction: reference number, left/right position, hairstyle and outfit. Reduce scene complexity first. Compare one revision at a time instead of changing both reference images and the prompt together.
Check that both IMAGE links reach separate image_urls sockets and that the prompt explicitly requests two visible people. A tight portrait crop can encourage the model to prioritize one subject.
Use the two-reference template without mask_url. For a focused mask edit, switch to a one-reference workflow using the composed frame as its single source.
LoadImage
Describe one edit
Qwen reference edit
SaveImage
ComfyUI
Outfit variants, scene changes and character frames
An edited image saved by SaveImage
Open workflowLoad starting frame
Add optional H3 LoRA
Generate a short clip
Save the video
ComfyUI
Character teasers, fashion loops and original adult visual-story openings.
A VIDEO connection saved under the spicyapi/h3-lora filename prefix.
Open workflowDownload the template or try the same model in the Playground before connecting the full workflow.