Removing a photobomber, a power line or a stray cup from a photo is one of the most common edits people make. On 10 November, Pruna AI released a dedicated Remove-Object LoRA for Qwen-Image 2.1 that does exactly this inside ComfyUI, with 6-step and 8-step fast versions and a ready workflow.
This guide covers the files you need and their exact folders, the small custom node that comes with it, how to mask, the right settings for each version, how to write the instruction, the known limits, and fixes when something stays behind.
- What: a rank-32 object-removal LoRA for Qwen-Image 2.1 editing, from Pruna AI (repo PrunaAI/P-Q-Image-2.1-Remove-Object-Comfy). Licence: Qwen Research License.
- Versions: 6-step, 8-step and base (40 steps). The fast ones pair with Pruna’s matching distillation LoRA.
- Settings: fast versions use adapter 0.7 plus distillation LoRA 1.0; base uses adapter 1.0 at 40 steps. No CFG. removal_strength 0.7.
- Extra node: copy the
precise_removal_outlinefolder intocustom_nodesand install its requirements (OpenCV). - Best input: photos of 1 MP or more, a clean mask over just the object, and a plain instruction such as “Remove the phone and the raised arm holding it.”
New to ComfyUI? My ComfyUI beginner guide covers install, models, nodes and fixes in one place.
What the Remove-Object LoRA Is
Qwen-Image 2.1 can already edit photos from a text instruction. You can ask it to remove things, but a general edit model sometimes changes too much, leaves a ghost of the object, or fills the gap badly.
Pruna trained a small adapter that teaches the model one job: take the object out and fill the space so it looks like it was never there. Key facts:
- Type: LoRA, rank 32, for Qwen-Image 2.1 editing (image-to-image).
- Released: 10 November 2026.
- Repos: PrunaAI/P-Q-Image-2.1-Remove-Object (original) and -Comfy (ComfyUI files, node and workflows).
- Licence: Qwen Research License, the same as the base model. Read it before any commercial use.
- Scope: removal only. For other edits it does not beat the base model, so switch it off for those.
If you have not used Qwen-Image 2.1 editing before, start with our Qwen-Image 2.1 Edit guide.
6-Step vs 8-Step vs Base
| Version | Steps | Adapter strength | Also needs | Use it when |
|---|---|---|---|---|
| 6-step | 6 | 0.7 | Pruna 6-step distillation LoRA at 1.0 | You want the fastest result |
| 8-step | 8 | 0.7 | Pruna 8-step distillation LoRA at 1.0 | Try it if the 6-step fill looks too smooth |
| Base | 40 | 1.0 | Nothing extra | Hard scenes, maximum quality; no ready workflow |
Start with either fast version, and switch to 8 steps or base if the fill looks too smooth. The 6-step distillation LoRA is the same one covered in our Pruna 6-step LoRA guide.
Files and Folders
Base model files (Comfy-Org/Qwen-Image-2.1)
| File | Folder |
|---|---|
qwen_image_2.1_int8_convrot.safetensors |
ComfyUI/models/diffusion_models/ |
qwen3vl_8b_int8_convrot.safetensors |
ComfyUI/models/text_encoders/ |
qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors |
ComfyUI/models/text_encoders/ |
qwen_image_2.1_vae_bf16.safetensors |
ComfyUI/models/vae/ |
Note: the workflow switches the prompt-enhancer encoder (pe_i2i) off, but the file must still be present or the loader shows red. Low on VRAM? See our Qwen-Image 2.1 GGUF guide for 8-12 GB.
Remove-Object adapters (PrunaAI/P-Q-Image-2.1-Remove-Object-Comfy)
| File | Folder |
|---|---|
p_q_image_2.1_remove_object_6_steps_v0.1_comfyui_bf16.safetensors |
ComfyUI/models/loras/ |
p_q_image_2.1_remove_object_8_steps_v0.1_comfyui_bf16.safetensors |
ComfyUI/models/loras/ |
p_q_image_2.1_remove_object_base_v0.1_comfyui_bf16.safetensors |
ComfyUI/models/loras/ |
Distillation LoRAs (PrunaAI/P-Q-Image-2.1-Comfy)
| File | Folder | Pairs with |
|---|---|---|
p_q_image_2.1_6_steps_v0.2_comfyui_bf16.safetensors |
ComfyUI/models/loras/ |
6-step remove adapter |
p_q_image_2.1_8_steps_v0.1_comfyui_bf16.safetensors |
ComfyUI/models/loras/ |
8-step remove adapter |
You only need the files for the version you plan to use. Download with the Hugging Face CLI from your ComfyUI/models/loras/ folder, for example:
hf download PrunaAI/P-Q-Image-2.1-Remove-Object-Comfy p_q_image_2.1_remove_object_8_steps_v0.1_comfyui_bf16.safetensors --local-dir .
hf download PrunaAI/P-Q-Image-2.1-Comfy p_q_image_2.1_8_steps_v0.1_comfyui_bf16.safetensors --local-dir .If a file lands in a sub-folder, move it up into loras/ and refresh ComfyUI.
Install the Precise Removal Outline Node
The Comfy repo includes a small helper node called Precise Removal Outline (category Pruna). It turns your mask into a clean outline guide for the edit. It is not in the Manager list yet, so install it by hand:
- Download just the node and workflow folders (not the weights):
hf download PrunaAI/P-Q-Image-2.1-Remove-Object-Comfy --include "precise_removal_outline/*" "workflows/*" --local-dir pruna-remove - Copy the
pruna-remove/precise_removal_outline/folder intoComfyUI/custom_nodes/. - Install its requirements with ComfyUI’s own Python:
cd ComfyUI/custom_nodes/precise_removal_outline pip install -r requirements.txtFor the Windows portable build:
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt - Restart ComfyUI. Search the node list for “Precise Removal Outline” to check it loaded.
The main requirement is OpenCV. If you already have opencv-python from other nodes, it will just confirm it.
Load and Run the Workflow
- From the Comfy repo, drag
workflows/p_q_image_2.1_remove_object_v0.1_8step.json(or the6stepone) onto the ComfyUI canvas. - Check each loader: the diffusion model, both text encoders, the VAE, the remove-object adapter and the matching distillation LoRA.
- Load your photo in Load Image.
- Draw a mask over the object (next section).
- Type your instruction, for example Remove the coffee cup from the table.
- Click Run. The output keeps the same size as your input.
There is no ready workflow for the base adapter. To use it, open the 8-step workflow, bypass the distillation LoRA, set the adapter to 1.0 and raise steps to 40.
How to Mask
You have two options:
- MaskEditor: right-click the Load Image node and choose Open in MaskEditor. Paint over the object, then save.
- Mask file: load a black image with the object painted white.
Good mask habits:
- Cover the whole object, including its shadow or reflection if you want those gone too.
- Leave a small margin, but do not spill over things you want to keep. Overlap is the main cause of over-removal.
- For a person holding something, mask both the object and the hand or arm if you want both gone.
No mask at all? In crowded scenes the instruction alone sometimes works better, because a rough mask can cut into nearby objects. Try both if the first result is off.
Masks not saving or working? Our ComfyUI inpainting fix guide covers common mask problems.
Settings Reference
| Setting | 6 / 8-step | Base | Notes |
|---|---|---|---|
| Steps | 6 or 8 | 40 | Match the adapter |
| Remove adapter strength | 0.7 | 1.0 | At 0.5 some removals fail |
| Distillation LoRA | 1.0 | off | Must match the step count |
| CFG | none (1.0) | none (1.0) | No negative prompt needed |
| removal_strength | 0.7 | 0.7 | On the Precise Removal Outline node |
| Seed | fixed | fixed | Change it if part of the object stays |
| Resolution | same as input | same as input | Generated in multiples of 32; about 2 MP works well |
Writing the Instruction
Keep it short and name exactly what should go. Good examples:
- Remove the phone and the raised arm holding it.
- Remove the man in the red jacket in the background.
- Remove the power lines across the sky.
- Remove the text watermark in the bottom corner. (Only on your own images.)
Tips:
- Describe the object by colour or position when there are several similar things.
- Do not describe what to put back. The adapter fills from the scene around it.
- One object or group per run. For several objects, run again on the output.
Known Limits
Pruna lists these limits on the model card:
- Over-removal when the mask overlaps something you want to keep.
- Smooth fill or a faint grid in the filled area on some images.
- Low-res photos: inputs under 1 MP lose texture. Upscale small photos first; see our ComfyUI upscale guide.
- Low strength: at 0.5 some objects are not removed.
- Removal only: it does not improve other kinds of edits.
Troubleshooting
| Problem | Fix |
|---|---|
| Precise Removal Outline node is red | Copy the whole precise_removal_outline folder into custom_nodes, install requirements with ComfyUI’s Python, restart. |
| Text encoder loader red even though it is off | Download qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors. It must exist. |
| Part of the object stays | Change the seed, enlarge the mask a little, or try the base adapter at 40 steps. |
| Nearby things disappear too | Tighten the mask so it does not overlap them, or try without a mask. |
| Fill looks blurry or plastic | Use a larger input (1-2 MP), or use the 8-step or base version. |
| Faint grid pattern in the fill | A known limit. Change seed, or try the base adapter. |
| Output is very noisy or broken | The distillation LoRA does not match the step count, or CFG is above 1. |
| Nothing changes | Adapter strength too low (raise to 0.7 or 1.0), or the LoRA node is bypassed. |
| Out of memory | Use the int8 files listed above, close other apps, or use a GGUF setup. |
FAQ
Can I use it commercially?
It uses the Qwen Research License, like Qwen-Image 2.1. Read the licence terms before using results commercially.
Does it work with Qwen-Image Lightning LoRAs?
Pruna’s workflows pair it with Pruna’s own distillation LoRAs. Other speed LoRAs, such as Lightning v2 RL, are not covered by the card, so results may vary.
Do I need a mask?
A mask usually gives the cleanest result. In busy scenes the text instruction alone sometimes works better.
Can it remove people?
Yes, background people are a common use. Only edit photos you own or have permission to edit.
Why is my output the same size as the input?
That is by design. The workflow generates at multiples of 32 and returns the original size.
Is there a video version?
No, this LoRA is for images only.
Sources
- Hugging Face: PrunaAI/P-Q-Image-2.1-Remove-Object-Comfy (files, node, workflows, settings)
- Hugging Face: PrunaAI/P-Q-Image-2.1-Remove-Object (model card, limits, licence)
- ComfyUI Wiki: Pruna Remove-Object LoRA
