Person, Garment and Pose

What goes in, what comes out

+

+

→

How the workflow works

Explained earlier:seed, guidance and steps · reference images and written roles (03) · outfit transfer (04) · decode and save

From this workflow on, nodes are named as they appear in ComfyUI.

1. Load the three references in order

   Three LoadImage nodes, top to bottom: identity (face and hair), outfit (the complete look) and pose (a photograph showing the body position you want). The prompt calls them image 1, image 2 and image 3, so keep the order.

   

   1. Image 1Identity: face and hair.
   2. Image 2Outfit: the complete look.
   3. Image 3Pose photo.

2. Detect, edit and reuse the pose

   

   1. DWPose EstimatorFinds the skeleton in the pose photo.
   2. Pose Input ToggleSwitches between detecting the pose again and reusing your edited pose.
   3. OpenPose StudioThe editor. Drag joints here, then press Apply.
   4. PreviewShows the skeleton that will be used.

   The pose is handled in a fixed sequence. Follow it in this order:

   1. Detect. Run the workflow once. DWPreprocessor analyses the pose photo and produces a skeleton: coloured lines and points for the head, shoulders, arms, hands, hips, legs and feet.
   2. Inspect. Look at the skeleton preview. Check that both arms and both legs were found and that left and right are not swapped.
   3. Edit. Open OpenPoseStudio and drag any joint to a new position, for example to move a hand to the hip.
   4. Apply. Press Apply in the editor so the edited skeleton is stored in the node.
   5. Select REUSE EDITED POSE. Set PoseInputToggle to REUSE EDITED POSE. If you leave it on detection, the next run detects the pose from the photo again and overwrites your edit.
   6. Generate again. Press Run. The result now follows the edited skeleton.

   Change: one joint. Keep fixed: identity, outfit, prompt and seed. Inspect: whether only the body position changed.

3. How the pose reaches the model

   The skeleton is drawn as a picture and given to the model as a third reference image, in the same way as the identity and outfit photos. Some workflows use a separate add-on model called a ControlNet to force a pose. This one does not. The skeleton is a visual reference that the model is asked to follow, so it guides the body without fixing it exactly. Hands and depth can still differ.

   Using the skeleton instead of the pose photo keeps that photo’s clothes, face and background out of the result. The prompt assigns the three roles in writing, as in 03, and describes the pose in words as well. Describing the hands helps where the skeleton is unclear.

4. Generate locally

   The same models and controls as 01A, on an 832 × 1216 image. The workflow uses the recommended 4 steps and CFG 1 with a fixed seed.

The second workflow: six cloud providers

The file B_Cloud_Model_Comparison.json is a separate, optional workflow. Its purpose is to send the same task to six paid image providers in one run, so you can compare how each handles identity, outfit and pose.

• Inputs: the same three photographs (identity, outfit, pose) and one shared written instruction that all six providers receive
• Outputs: six images, one per provider, each saved separately under a provider-specific filename
• Needs: a Comfy account with credits. No local model files. Every full run can charge for six images. The note inside the workflow suggests budgeting roughly 120 to 160 credits per complete comparison. Check the price badges, because prices change

Why this is not a controlled ranking

• Input preparation differs. Google and Kling receive resized copies with white borders, because their nodes need all three images to be the same size. The other four receive the original files
• Sizes and quality tiers differ. Each provider is set to its own roughly one-megapixel option, and OpenAI and xAI have a quality setting the others lack
• Seeds are not comparable. Seed 22 is fixed where supported, but a seed means something different in every model, and OpenAI and Kling do not promise repeatable results
• The pose input differs from the local workflow. The cloud providers receive the original pose photograph. The local workflow uses the extracted skeleton

To test one provider only, switch off the other five Save Image nodes before running. Select them and press Ctrl+M. This is called muting a node.

Local and cloud results

These images were saved from earlier runs, and some of the cloud images were made with earlier versions of the input files. Treat them as illustrations of what each route can produce, not as a controlled comparison or a ranking.

Check the result

• Identity: the face matches image 1
• Outfit: the garments keep their proportions, seams, colour and footwear from image 2
• Pose: the joints follow the skeleton, including hands and crossing legs, and not the clothes or setting of the pose photo
• Hands touch the clothes convincingly, with no crushed fabric or hands merging into a hip
• For the cloud comparison: note the charge and the time for each provider alongside the image

A face belongs to someoneUse an identity photo only with that person’s permission for this purpose. Use it responsibly.

Download the workflow

• A_Person_Garment_Pose.jsonLocal workflow with pose detection and editing
• B_Cloud_Model_Comparison.jsonSix-provider cloud comparison · uses Comfy account credits
• HERO_model_identity.png, L06C_dressed_anchor.png, Pose Reference Prompt.jpgThe example identity, outfit and pose images. Put them in ComfyUI’s input folder, or load your own

Click a filename to download it, or right-click it and choose Save link as. Keep the .json ending. Then drag the file onto the ComfyUI canvas, or use Workflow → Open.