If ComfyUI’s node graph is not your idea of fun, the good news is that dwpose works beautifully inside AUTOMATIC1111 too. The WebUI route hides the machinery behind a familiar panel: upload a reference photo, pick the dwpose preprocessor, pick a matching ControlNet model, and generate. This guide sets the whole thing up from scratch, covering the ControlNet extension, the one preprocessor name you need to remember, the model matching rule that trips up beginners, and the settings worth adjusting once your first pose-guided image comes out.
What You Need in Place
- A working AUTOMATIC1111 WebUI install.
- The sd-webui-controlnet extension by Mikubill, installed and loading without errors.
- At least one OpenPose ControlNet model file, matched to the Stable Diffusion version of your checkpoint: an SD 1.5 model for SD 1.5 checkpoints, an SDXL model for SDXL checkpoints.
- A reference photo with a clear person in it.
Notice what is not on the list: a separate dwpose download. The dwpose estimator ships inside the ControlNet extension as a preprocessor, models included in the extension’s download flow, so your only real shopping trip is for the ControlNet model itself.
Step 1: Install or Update the ControlNet Extension
In the WebUI, open the Extensions tab, go to Install from URL (or the Available list), and add Mikubill’s sd-webui-controlnet extension. Apply and restart the UI completely. If ControlNet is already installed from long ago, update it before following the rest of this guide: dwpose preprocessors arrived in later versions of the extension, and a stale install is the number one reason the preprocessor name simply does not appear in the dropdown. After the restart, scroll down in the txt2img or img2img tab and confirm you can see a ControlNet panel with units you can enable. That panel is your home for everything that follows.
Step 2: Put the OpenPose Model Where It Belongs
Download an OpenPose ControlNet model that matches your checkpoint family and place the file in the ControlNet extension’s models folder (the extension shows its expected models path, and its own documentation points to it). Back in the WebUI, hit the refresh icon next to the model dropdown in the ControlNet unit so the new file appears. This matching rule deserves repeating because it causes half of all “ControlNet does nothing” reports: the control model, the checkpoint and the generation settings all have to belong to the same Stable Diffusion generation. An SD 1.5 pose model cannot steer an SDXL checkpoint, no matter how nicely everything else is set.
Step 3: Load Your Reference into a ControlNet Unit
Expand ControlNet Unit 0, tick Enable, and upload your reference photo into the unit’s image area. Tick the option to allow the preprocessor preview if you want to see the skeleton before generating; you do want this, at least while learning, because the preview is your only chance to catch a bad estimate before it steers a whole batch. Leave the control type selection alone for a moment; choosing the preprocessor and model explicitly in the next step is more reliable than letting presets guess.
Step 4: Choose dw_openpose_full
In the preprocessor dropdown, select dw_openpose_full. That name unpacks neatly: DW for dwpose, OpenPose for the skeleton style it outputs, and full for the complete layout of body, hands and face. Sibling options exist for body-only estimates, which are slightly quicker and cleaner for distant figures, but full is the right default and the one this guide assumes. Set the preprocessor resolution to 512 to start; raise it when hands or faces look soft in the preview. Then choose your OpenPose model in the model dropdown beside it. Preprocessor makes the map, model consumes the map, and both halves of the pair are now in place.
Step 5: Set Control Weight and Generate
Set the control weight to around 1.0 for your first run. Write a prompt that describes your character, clothing, scene and style, while saying nothing about the pose itself, and generate. Your character should land in the reference’s stance, fingers and head angle included. From there, tune in small steps: weight a little lower loosens the pose and lets the model pose hands more naturally when the skeleton fights the anatomy; weight higher enforces the map strictly. The starting and ending control steps default sensibly; the pose is mostly decided early in generation, so an ending step below 1.0 rarely changes the stance, it mostly frees up fine detail.
Checking the Preview Like a Professional
Get into the habit of reading the dwpose preview before every serious batch. Check that the wrists connect to the correct hands (crossed arms are the classic mix-up), that finger fans follow the real fingers rather than inventing new ones, and that the face points tilt the way the head tilts. A skeleton that passes those three checks will steer generations faithfully. One that fails them needs a better reference photo or a higher preprocessor resolution, not a different seed. Fixing the map costs seconds; fixing thirty generated images costs your evening.
A1111 or ComfyUI for DWPose?
Honest answer: the estimation quality is identical, because both front ends run the same dwpose models through the same kind of preprocessor. The difference is ergonomics. A1111 wins on simplicity: one panel, one upload, two dropdowns. ComfyUI wins on flexibility: reusable skeleton images, video pipelines and batch automation are easier to build there. Plenty of users keep both, prototyping poses in A1111 and moving to ComfyUI when a project grows. Your dwpose skills transfer completely between them, settings and all.
Using More Than One Unit, and Saving Poses
The WebUI lets you enable several ControlNet units at once, and dwpose plays well with company. A common pairing is a dwpose unit controlling the body plus a second unit handling depth or a face-focused reference, each with its own weight. Keep the pose unit’s weight firm and let supporting units sit lower, or the instructions start arguing. Separately, remember that the preview skeleton is itself an image worth keeping. Right-click and save good skeletons into a poses folder. In later sessions you can upload the skeleton directly into a unit and set the preprocessor to none, skipping estimation entirely. Your favorite stances become permanent, reusable assets, and a series of images built on one saved map stays noticeably more consistent.
img2img Notes
Everything above works in img2img too, with one shift in thinking. In txt2img, ControlNet supplies the pose to a picture being invented from nothing. In img2img, your source image already has a composition, and the dwpose unit can either reinforce that same pose (extract the map from the input image itself) or impose a new one from a different reference. The second option feels like puppetry: keep the character and setting, swap the stance. Watch your denoising strength here; too low and the old pose haunts the result, too high and the character you wanted to keep drifts away. Middle values plus a firm control weight usually land the swap cleanly.
Conclusion
Using dwpose in AUTOMATIC1111 comes down to a short checklist: current ControlNet extension installed, an OpenPose control model matched to your checkpoint family, a reference uploaded to an enabled unit, dw_openpose_full chosen as the preprocessor, and control weight starting near 1.0. Add the habit of reading the skeleton preview before you generate, and pose control in the WebUI becomes a calm, predictable tool rather than a slot machine. Once the first copied pose lands cleanly, the settings article below will help you push quality further.
Read next: