# CafeAVA | Universal Pose Transfer — Qwen Image 2.1

Two ComfyUI workflows for transferring reference characters into a selected one-person or two-person pose. Choose your own pose references and identity images. No pose gallery assets, example photos, checkpoints, LoRAs, or upscaler model weights are bundled.

## Included files

- CafeAVA_Universal_Pose_Transfer_Simple_Public_Layout.json — streamlined 13-node workflow, one Qwen generation pass, optional LoRAs and Qwen cache.

- CafeAVA_Universal_Pose_Transfer_Enhanced_Public_Layout.json — 25-node workflow with aspect-ratio / megapixel control, pose padding, reference-image resizing, optional upscale/sharpen/DLSS5 path, original/enhanced save outputs, and a visual comparer.

Both JSONs are meant to be loaded directly into ComfyUI; make sure your installed nodes and model filenames match your local environment. The displayed image filenames are placeholders, not packaged image files.

## Requirements

- A version of ComfyUI with Qwen Image 2.1 multi-image editing nodes TextEncodeQwenImage21), Qwen Image 2.1 compatible model, Qwen3-VL 8B text encoder, and compatible VAE.

- VNCCS PoseStudio V2 pose LoRA VNCCS_QI2_PoseStudioV2.safetensors) from MIUProject. This is a model-only LoRA loader, strength 1.0 in the distributed defaults.

- rgthree Power Lora Loader (optional LoRA slots) and the QwenImage21Cache node type provided by your installation. The JSON may flag unavailable custom nodes; install matching dependencies or adjust your own setup.

- Enhanced only: ResolutionSelector, ResizeAndPadImage, ImageScaleToTotalPixels, SplitImageWithAlpha, Comfyroll CR Upscale Image, its selected 4x-UltraSharpV2.safetensors upscaler, WAS Image Lucy Sharpen, DLSS5 enhancer and settings nodes, rgthree Image Comparer. Ensure these are installed before using the enhanced pathway. The selected model and weights are not included.

## Getting started

1. Load the chosen .json workflow in ComfyUI.

2. Image 1: select a pose reference diagram or photograph. Image 2: select Character 1. Image 3: select Character 2 for two-person scenes.

3. Replace the model, LoRA, VAE, encoder and upscaler dropdown placeholders with files actually installed on your system.

4. Edit the opening sentence of the prompt to assign which image reference occupies which figure in your pose. Left/right are only the shipped example, not a restriction.

5. Set seed, steps, CFG and any LoRA strengths yourself. Nothing is locked. Run and inspect the result.

### Character position matters

Some poses have Character 2 on the left, others on the right. In others, a person is behind/in front of, above/below, or seated on the lap of the other character. Describe the positions unambiguously; don't assume that the example left/right order fits every pose.

Example, man on left:

Replace the left mannequin in image1 with the man from image3 and the right mannequin with the woman from image2.

Example, man behind:

Replace the person standing behind in image1 with the man from image3 and the person in front with the woman from image2.

Example, man seated and woman on lap:

Replace the seated person with the man from image3 and the person sitting on his lap with the woman from image2. Include a sturdy chair supporting the man, with realistic contact and body positioning.

Example with props:

Recreate the two characters on a large white king-size bed with realistic white cotton sheets and pillows, keeping the relative pose and camera angle from image1.

Add a specific background such as a comfortable cafe, yoga studio, or ocean-view beach when desired. If a pose depends on a chair, sofa, yoga mat or bed, specify that item in the prompt.

### Single-person mode

Use a one-person pose reference and a single identity image in Image 2. Change the prompt to explicitly describe one person.

- Enhanced: in the local test, bypassing both the Image 3 Load Image node and its associated Character 2 Resize node allowed the workflow to pass the connection check without disconnecting wires. Final one-character visual results should be verified by the user.

Sample prompt:

Replace the mannequin in image1 with the person from image2. Preserve the reference pose, perspective, expression, hairstyle, skin tone, proportions and visible identity details. Render one coherent photorealistic person in the selected environment.

## Defaults and known test notes

- Both shipped workflows retain 25 steps / CFG 1 / Euler / simple / denoise 1.0 and PoseStudio V2 LoRA at strength 1.0. These are starting points, not requirements.

- Our local tests found BF16 produced the cleanest visual quality among the Qwen Image 2.1 model variants we tried. We also obtained successful pose/identity outputs with Turbo and an INT8 ConvRot variant. Results are system- and setting-dependent, and Turbo may benefit from its own recommended schedule rather than copying full-model steps.

- In another local test, adding three optional LoRAs together while using CFG 5 preserved pose and identities. Optional LoRA stack in that specific test: NSFW Qwen by TheseAlpacas V2 (1.0), Qwen-Image-2.1 NSFW Image Edit (0.8), and QWEN-2.1-Unchained-V1 (0.7); file names/availability may vary. These are not included or required. This test is reported as an observation, not a default recommendation. Keep sample media on Civitai.com suitable for its SFW audience.

- Enhanced: ResolutionSelector controls the native output aspect ratio and approximate megapixel count; its pose image is resized/padded, not cropped. TextEncodeQwenImage21 resolution 0 is intentional for the size-controlled workflow. Enhancement is post-processing and can subtly change texture/identity. Compare original (A) and enhanced (B).

- We have inspected link topology and stripped personal sample names, but actual execution in your installed ComfyUI with the final distributed JSONs and model paths is the ultimate compatibility test.

## Pose reference gallery and attribution

Get individual pose images directly from the creator's original website:

https://ai-pose-gallery.pages.dev/

The gallery credits ギア (Gia) as original creator and explicitly disallows redistribution/reposting of its image files and bulk downloading. None of its images are included here, and its pose diagrams should not be included in redistribution or preview collages. Use the original gallery site to select references individually.

Creators and tools credited:

- Qwen team — Qwen Image 2.1: https://huggingface.co/Qwen/Qwen-Image-2.1

- MIUProject — VNCCS PoseStudio for Qwen Image 2.1: https://huggingface.co/MIUProject/VNCCS_PoseStudio_QI2.1

- ComfyUI and the authors of installed custom nodes (rgthree, Comfyroll, WAS, DLSS5 Enhancer) used by these community workflows.

## Distribution

This pack contains only ComfyUI workflow JSONs and this guide. Third-party model weights and original third-party pose reference images must be obtained from their respective owners/sources. This pack is made for educational and lawful creative workflows.