Inpainting lets you repaint just one part of an image and leave the rest alone. You can remove a stray object, change a shirt colour, fix a messed-up hand, swap a sky or add something that was never there.
In ComfyUI it comes down to three things: an image, a mask that marks the area to change, and a model that fills that area so it blends with everything around it.
This guide covers the official basic inpainting workflow, how to draw masks in the Mask Editor, the two inpainting nodes and when to use each, the Flux.1 Fill dev method for the best quality, the settings that matter, and how to fix the usual problems. Steps follow the official ComfyUI docs as of November 2026.
- Quick start: download
512-inpainting-ema.safetensorsintoComfyUI/models/checkpoints/, load the official SD inpaint workflow, upload an image and press Run. - Mask: right-click
Load Image> Open in MaskEditor, paint over the area, click Save. - Two nodes:
VAE Encode (for Inpainting)for classic SD inpainting (denoise 1.0),InpaintModelConditioningfor Flux Fill and partial-denoise edits. - grow_mask_by: default 6. Raise it to 10 to 16 if you see hard seams.
- Best quality: Flux.1 Fill dev (
flux1-fill-dev.safetensorsinmodels/diffusion_models/). - Prompt only what goes in the mask, not the whole picture.
How Inpainting Works
A normal text-to-image run starts from pure noise. Inpainting starts from your picture instead. The mask tells ComfyUI which pixels it may change.
Inside the mask, the model generates new content from your prompt. Outside the mask, the original pixels are kept. The model also looks at the surrounding area, so lighting, perspective and texture match.
| Piece | Job |
|---|---|
| Image | The picture you want to fix or change |
| Mask | White (painted) area = change it, everything else = keep it |
| Prompt | Describes what should appear inside the mask |
| Inpainting node | Combines image, mask and VAE into a latent the sampler can work on |
| Model | Fills the masked area. Dedicated inpainting models blend best |
Inpainting is closely related to outpainting, which extends an image beyond its borders, and to image to image, which repaints the whole picture at a chosen strength.
Which Model to Use
| Model | File and folder | Best for | VRAM |
|---|---|---|---|
| SD 2 inpainting (official tutorial) | 512-inpainting-ema.safetensors in models/checkpoints/ |
Learning, quick fixes, low VRAM | Low |
| Flux.1 Fill dev | flux1-fill-dev.safetensors in models/diffusion_models/ |
Realistic edits, best blending and prompt following | High (fp8 or GGUF to save memory) |
| Any regular checkpoint (SD 1.5, SDXL) | Your usual checkpoint | Small edits in the same style as your image | Low to medium |
| Qwen-Image Edit and similar edit models | See our guides | Edits by instruction without a mask | High |
The official tutorial compares the inpainting model against the standard v1-5-pruned-emaonly-fp16.safetensors and notes that the inpainting model gives better results and more natural transitions, because it was trained for exactly this job.
Not sure where a file goes? Our ComfyUI model folders cheat sheet lists every folder.
Method 1: The Official Basic Workflow
- Update ComfyUI so you have the current nodes and templates. See our ComfyUI update guide. New users can start with how to install ComfyUI.
- Download the model
512-inpainting-ema.safetensorsfrom the Comfy-Orgstable_diffusion_2.1_repackagedpage on Hugging Face and put it inComfyUI/models/checkpoints/. - Load the workflow. Download the official
sd1.5_inpaint.pngworkflow image from the Comfy-Org example_workflows repo and drag it onto the canvas, or open it with Workflows > Open (Ctrl + O). - Check Load Checkpoint is set to
512-inpainting-ema.safetensors. - Upload your image to the
Load Imagenode. The tutorial’s sample image already has a transparent area, and that transparent area is treated as the mask. - Write a prompt for what should appear in the masked area, plus a negative prompt.
- Click Run or press Ctrl + Enter.
- Try other seeds in the
KSampleruntil you get a result you like.
The nodes in this workflow are: Load Checkpoint, Load Image, two text prompt nodes, VAE Encode (for Inpainting), KSampler, VAE Decode and Save Image.
Draw a Mask in the Mask Editor
For your own images you will usually paint the mask yourself.
- Upload the image to
Load Image. - Right-click the
Load Imagenode and choose Open in MaskEditor. - Set the brush size, opacity and hardness in the panel on the right.
- Paint over the area you want to change. Go slightly past the edges of the object.
- Use the eraser to fix mistakes.
- Click Save. The mask now flows from the node’s MASK output into the inpainting node.
Re-inpaint a result: right-click Save Image and choose Copy (Clipspace), then right-click Load Image and choose Paste (Clipspace). Open the Mask Editor again and paint the next area. This is how you fix an image one piece at a time.
| Mask tip | Why |
|---|---|
| Cover the whole object plus a small margin | Leftover edges of the old object show through otherwise |
| Include the shadow when removing an object | A floating shadow is a giveaway |
| Keep masks reasonably sized | A tiny mask gives the model no room to draw something new |
| Use a soft brush for organic shapes | Softer edges blend better on skin, hair and skies |
Need a precise mask around a person or product? A background removal model can make one automatically. See how to remove backgrounds locally.
VAE Encode for Inpainting vs InpaintModelConditioning
ComfyUI has two built-in ways to feed an inpainting job to the sampler.
| VAE Encode (for Inpainting) | InpaintModelConditioning | |
|---|---|---|
| Inputs | pixels, vae, mask, grow_mask_by |
positive, negative, vae, pixels, mask (current builds also have a noise_mask toggle) |
| Outputs | latent |
positive, negative, latent |
| What it does | Blanks the masked area before encoding, so the model must invent it from scratch | Passes the original image and mask to the model as conditioning, and keeps the original pixels in the latent |
| Denoise | Must be 1.0 | Works at 1.0 or lower, so you can make subtle changes |
| Best with | SD 1.5 and SD 2 inpainting checkpoints | Flux.1 Fill dev and inpainting models where you want to keep some of the original |
Simple rule: to replace something completely, either node works. To change something gently (recolour a shirt, refine a face), use InpaintModelConditioning with a denoise around 0.5 to 0.8, because VAE Encode (for Inpainting) throws away the original pixels under the mask.
Method 2: Flux.1 Fill Dev Inpainting
Flux.1 Fill dev is Black Forest Labs’ dedicated fill model. It understands the scene around the mask very well and follows prompts closely, so it is the best choice for realistic edits if your GPU can handle it.
Files you need
The model is gated: open black-forest-labs/FLUX.1-Fill-dev on Hugging Face, sign in and accept the licence first.
| File | Folder |
|---|---|
flux1-fill-dev.safetensors |
ComfyUI/models/diffusion_models/ |
clip_l.safetensors |
ComfyUI/models/text_encoders/ |
t5xxl_fp16.safetensors (or t5xxl_fp8_e4m3fn for less VRAM) |
ComfyUI/models/text_encoders/ |
ae.safetensors |
ComfyUI/models/vae/ |
If you already run Flux text-to-image, you only need the fill model. Our how to use Flux guide covers the shared files.
Steps
- Open Workflow > Browse Templates and search for Flux Fill. Pick the Flux.1 Fill dev inpaint template.
- Check the loaders:
Load Diffusion Model=flux1-fill-dev.safetensors,DualCLIPLoader= t5xxl + clip_l with type flux,Load VAE=ae.safetensors. - Upload your image and paint the mask in the Mask Editor.
- Describe what should appear in the masked area.
- Click Run.
The template routes the image and mask through InpaintModelConditioning and uses a high FluxGuidance value. That high value is normal for the Fill model, so leave it unless you have a reason to change it.
Inpainting With a Regular Model
You do not have to use a dedicated inpainting checkpoint. A regular SD 1.5 or SDXL model keeps the same style as images it generated.
- For a complete replacement, use
VAE Encode (for Inpainting)at denoise 1.0 and a generousgrow_mask_byso the seam blends. - For a light touch-up, encode the image with a normal
VAE Encode, add the mask withSet Latent Noise Mask, and use a denoise of 0.4 to 0.7. - Expect weaker blending than a dedicated inpainting model, especially on large masks.
Want to keep a pose or outline while repainting? Combine the mask with a ControlNet. See our ComfyUI ControlNet guide.
Settings That Matter
| Setting | Where | Starting value | Effect |
|---|---|---|---|
| grow_mask_by | VAE Encode (for Inpainting) | 6 (default), range 0 to 64 | Expands the mask outward so the edge blends. 10 to 16 for softer transitions |
| denoise | KSampler | 1.0 to replace, 0.5 to 0.8 to adjust | How much the masked area changes |
| steps | KSampler | 20 to 30 | More steps can add detail |
| cfg | KSampler | Model default (around 7 for SD) | Higher follows the prompt harder but can look harsh |
| seed | KSampler | Randomize | Try several seeds; inpainting results vary a lot |
| Prompt | Text encode | Describe the masked area only | Prompting the whole scene can paint extra objects into the mask |
Recipes for Common Jobs
| Job | Mask | Prompt idea | Settings |
|---|---|---|---|
| Remove an object | Object plus its shadow | Describe the background, e.g. “empty wooden table” | Denoise 1.0, grow_mask_by 10+ |
| Change clothing colour | The garment | “red knitted sweater” | InpaintModelConditioning, denoise 0.6 to 0.8 |
| Replace the sky | Sky area | “dramatic sunset sky, orange clouds” | Denoise 1.0, soft brush |
| Fix a hand | The whole hand | “natural hand, five fingers, relaxed” | Denoise 0.7 to 1.0, several seeds |
| Add an object | Where it should go, a bit larger than the object | “a ceramic coffee mug” | Flux Fill works best |
| Fix faces automatically | Detected by a model | Your normal prompt | Use a FaceDetailer node |
For removing objects with an edit model and no hand-painted prompt, try our Qwen-Image remove object LoRA workflow.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Nothing changes | Mask not saved, or denoise too low | Save in the Mask Editor and check the MASK wire; raise denoise |
| Grey or blurry blob in the mask | Regular model with VAE Encode (for Inpainting) | Use an inpainting model, or switch to Set Latent Noise Mask |
| Visible seam or halo | grow_mask_by too small, hard mask edge | Raise grow_mask_by, paint with a softer brush |
| Old object still visible | Mask too tight | Paint past the edges and include the shadow |
| Wrong thing appears in the mask | Prompt describes the whole scene | Describe only what should fill the mask |
| Colours shift outside the mask | Repeated VAE encode/decode | Composite the result back onto the original with the mask |
| Flux Fill out of memory | Model too big for your GPU | Use the fp8 text encoder or a GGUF fill model |
| Still stuck | Many possible causes | See our ComfyUI inpainting not working fix |
FAQ
Do I need a special model for inpainting in ComfyUI?
No, but it helps a lot. Dedicated inpainting models such as 512-inpainting-ema.safetensors or Flux.1 Fill dev blend much better than regular checkpoints.
What does grow_mask_by do?
It expands your mask outward by that many pixels so the model can repaint a thin border of the original and hide the seam. The default is 6.
What denoise should I use for inpainting?
Use 1.0 to replace something completely. Use 0.5 to 0.8 with InpaintModelConditioning or Set Latent Noise Mask when you want to keep part of the original.
How do I open the Mask Editor?
Right-click the Load Image node and choose Open in MaskEditor. Paint, then click Save.
Is Flux Fill free to use?
You can download it free after accepting the licence on Hugging Face. FLUX.1 dev models use a non-commercial licence, so check the terms before using outputs commercially.
Can I inpaint the same image several times?
Yes. Copy the result with Copy (Clipspace) from Save Image, paste it into Load Image with Paste (Clipspace), and mask the next area.




