Puts your own character into an existing scene frame with one Qwen-Image 2.1 edit pass. The frame is the canvas: the pose, the other people, the camera, the light and the room stay; only the woman is replaced by the character on your reference picture. A depth map of the frame holds the pose and the sizes.

The package is the working recipe, not only a graph: the standing instruction texts, the prompt template, the questions that let any vision model write the prompt for a frame, and the two rules that decide whether a swap comes out clean.

WHAT IT DOES

• Input: a scene frame, a picture of your character (portrait or full body), and four short text boxes.

• Output: the same frame with your character in it, at the frame's own pixel size.

• One pass, 6 steps (Qwen-Image 2.1 int8 + the Viggle turbo LoRA), no CFG.

WHY IT IS MORE THAN A GRAPH

• Classify first. A frame where the woman's face is not visible (back view, hidden face, head out of frame) must not get a face swap - it turns her toward the camera. The workflow has a separate hair-only mode for these frames.

• One pass. A second full pass over the result degrades style and anatomy, so the graph is built for exactly one.

• Keep rules. The instruction says what must stay exactly as in the frame, so the swap does not tidy the picture up.

THREE MODES, ONE LINK TO CHANGE

• A - face visible: the swap (default wiring)

• C - oral scene: the swap plus an extra keep rule

• B - no face: hair only, no face is ever drawn

WHAT IS INCLUDED

• character_swap_qwen21.json - the ComfyUI workflow (groups: Inputs, Depth, Models, Instruction + Prompt, Sampler, Output; notes inside)

• character_swap_qwen21_api.json - the same graph in API format, for scripts

• PROMPT_KIT.md - instruction texts, prompt templates with placeholders and one filled example, the vision questions, the decision rule, a per-frame checklist

• README.md - requirements, step-by-step use, the three modes, scripting, tips, troubleshooting

REQUIREMENTS

• ComfyUI with the core nodes Text Encode Qwen Image 2.1 and Qwen Image 2.1 Cache (tested on 0.37.0)

• Custom nodes: comfyui_controlnet_aux - only for the depth map (Depth Anything V2); it can be replaced by a depth map loaded from a file

• Models (about 19.3 GB with the depth weights): qwen_image_2.1_int8_convrot.safetensors, qwen3vl_8b_int8_convrot.safetensors, qwen_image_2.1_vae_bf16.safetensors, Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r128.safetensors, depth_anything_v2_vitl.pth (fetched by the depth node)

• GPU: runs on 12 GB (RTX 3080 Ti) and 8 GB (RTX 3070 Ti) cards; 8 GB is the tight end. On 12 GB the full swap took 49-51 s and the hair-only mode 25 s.

• This workflow is built for Qwen-Image 2.1 Edit. The model weights are not included and are used under their own licence - check it for your use.

HONEST LIMITS

• One woman per frame in this version.

• The model works at about one megapixel; larger frames are resized back and come out softer than the source.

• You fill the pose sentence and the details list yourself (two questions for any vision model are included). An automatic classifier is not part of the workflow.

• The prompt kit contains explicit wording for adult scenes. Use it only with your own characters or with consent; adults only.