Qwen-Image 2.1 Image Edit in ComfyUI (2026): 8 to 16 GB Setup, Multi-Image Edits and Best Edit LoRAs

Qwen-Image 2.1 is not just a text-to-image model. It is a single 7B model that also edits images, and since its September release it has quietly become the best open image editor you can run at home: up to 10 reference images in one edit, painted marks to show exactly what to change, background removal with real transparency, and a Turbo version that edits in 8 steps. The Comfy-Org repo has passed 5.6 million downloads. Most guides only cover generation, so this one is all about editing: which files you need for 8, 12 and 16 GB cards, the exact ComfyUI folders, how the reference slots and mark-up editing work, the best settings, the edit LoRAs worth installing (4-step Lightning, camera angles, object removal), and how to stop the Qwen3-VL encoder from second-guessing ordinary creative edits.

What Qwen-Image 2.1 Can Edit

Older Qwen releases used separate models for generating (Qwen-Image 2512) and editing (Qwen-Image-Edit 2509 and 2511). Qwen-Image 2.1 merges both into one 7B diffusion transformer with a Qwen3-VL 8B encoder that can see your input images. In one edit it can:

• Change part of an image from a text instruction: swap a jacket, recolour a car, turn day into night, add rain, change a sign’s text.
• Combine up to 10 reference images: put a product from one photo into a scene from another, dress a character in a garment from a reference, build a group shot from several portraits, while keeping identities and products consistent.
• Edit only where you mark: draw a circle or paint over a region and name the mark’s colour in the prompt.
• Remove backgrounds and output a PNG with a real alpha channel, or extract a subject from a photo.
• Work at up to about 2K, its trained size. Above that, prompt adherence drops.

The licence is the Qwen Research License, which is non-commercial. For client or paid work, check the licence or use an Apache 2.0 editor such as the MOSS-Image Edit LoRA.

VRAM Picks

ComfyUI runs the encoder first and then moves it to system RAM, so the diffusion model and encoder do not have to fit in VRAM together. 32 GB of system RAM is enough for every option above. What really sets speed and memory is canvas size: Comfy’s own tests on an RTX 5090 show about 0.3 seconds per step at resolution 1024 but about 6 seconds per step for a full 3008×4000 phone photo.

All Files and ComfyUI Folders

Official files are in Comfy-Org/Qwen-Image-2.1 (not gated). GGUF files are in unsloth/Qwen-Image-2.1-GGUF.

Check the exact GGUF file names on the unsloth “Files” tab before downloading, since repos sometimes rename files. Folder questions: ComfyUI model folder cheat sheet. Which quant: GGUF quantization explained.

Install and Update

1. Update ComfyUI to the latest release (0.36.0 or newer). The Qwen Image 2.1 nodes (Text Encode Qwen Image 2.1, the edit subgraph and the KV cache node) are core nodes, so an old install shows red missing-node boxes. Portable: run update/update_comfyui.bat. Git install: git pull then pip install -r requirements.txt.
2. For GGUF files, install ComfyUI-GGUF:

   cd ComfyUI/custom_nodes
   git clone https://github.com/city96/ComfyUI-GGUF
   pip install --upgrade gguf

3. For a GGUF Heretic encoder, also install pottokao’s add-on, which lets Qwen3-VL GGUF encoders load with their vision tower (needed for edits):

   cd ComfyUI/custom_nodes
   git clone https://github.com/pottokao-dotcom/ComfyUI-GGUF-Qwen3VL-TE

   It adds no nodes. After a restart the console prints a [GGUF-Qwen3VL-TE] patch message.

Your First Edit, Step by Step

1. Open Workflow, Browse Templates and load Qwen Image 2.1 Turbo: Image Edit.
2. Check the loaders: Turbo INT8 model, qwen3vl_8b_int8_convrot encoder and the 2.1 VAE. For GGUF, replace the model loader with Unet Loader (GGUF) and connect its MODEL output in the same place.
3. In the first Load Image node (image_1), load the picture you want to edit.
4. Open the Image Edit subgraph and set resolution to 1024 while you test. The default 0 keeps your photo’s full size, which is slow for phone photos.
5. Write a direct instruction, for example: “Change the season to winter: snow on the roofs and street, bare trees, overcast sky. Keep the buildings and the red car unchanged.”
6. Run, then fix the seed and change one thing at a time.

Want the base model’s extra detail? Use the Qwen Image 2.1: Image Edit template (25 steps) with the base INT8 file. Simple local edits often look the same on Turbo; whole-frame changes and hands benefit from the base model.

Multi-Image Edits (up to 10 References)

The Text Encode Qwen Image 2.1 node has inputs image_1 to image_16, and the edit templates wire up the first 10. The rule is simple:

• image_1 is the canvas: the image being edited, and the source of the output size.
• Every other slot is a reference you point to in the prompt by number.

Example from the official template: image_1 is a portrait, image_2 is a denim shirt on a plain background, and the prompt is “Dress the person in <image1> in the shirt from <image2>, keep the pose, face and background.” The same pattern works for products (“place the bottle from <image2> on the table in <image1>, matching the lighting”), interiors (“use the sofa from <image2> and the lamp from <image3>”) and characters (“draw the knight from <image2> and the dragon from <image3> on the cliff in <image1>”).

Tips: crop references tightly around the subject, keep them under about 1500 px on the long side (every reference adds compute), and name each one by what it is as well as its number.

Paint-to-Edit: Marks and Masks

Qwen-Image 2.1 understands drawn marks, which is the most precise way to say “here”:

1. Right-click the image_1 Load Image node and choose Open in Mask Editor.
2. Use the Paint Pen (not the mask brush) to draw a circle or fill over the area, in a colour that does not appear in the photo, such as bright green.
3. Save. The strokes become part of the image itself.
4. Name the colour in the prompt: “Replace the lamp in the green circle with a tall brass floor lamp. Remove the green mark.”

If traces of the mark colour remain in the result, say what colour that area should be (“the wall behind it is white”).

Background Removal With Transparency

The Remove Background: Qwen Image 2.1 template uses the prompt “Remove the background, and output a PNG image” and returns an image with a real alpha channel. Two notes:

• Nodes that expect a normal 3-channel image will fail on the 4-channel output. Add SplitImageWithAlpha after the decode and use its RGB output for them.
• The mask convention is 1 = transparent. To rebuild a PNG, connect the mask to the alpha input of JoinImageWithAlpha.

For very fast, CPU-only cut-outs of simple product shots, see local background removal on CPU.

Best Settings

• On Turbo, do not raise the step count without changing the schedule. The sigmas are fixed for 8 steps.
• Small local edits can work in 4 to 8 steps even on the base model. Whole-frame rewrites need about 25.
• Keep custom_size off unless you need a different canvas. If you turn it on, keep it close to the resized image_1 size, or the content can shift.
• More on samplers: ComfyUI samplers and schedulers explained.

Fewer Refusals: Heretic Encoder and Prompt Enhancer

Qwen-Image 2.1 reads your instruction (and looks at your images) with Qwen3-VL 8B, a chat model trained to refuse. For editing, that shows up as edits that are quietly ignored or toned down: a horror make-up effect that comes out mild, a battle-damaged sci-fi prop that comes out clean, a dark fantasy scene that turns bright. pottokao’s Heretic encoder removes most of that refusal behaviour: 100/100 refusals down to 5/100 on its test set, with a KL divergence of 0.022, which means outputs barely change for normal prompts.

Files (from pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF), into ComfyUI/models/text_encoders/:

• qwen3vl_8b_heretic-Q4_K_M.gguf (5.03 GB) plus its mmproj file (1.16 GB). For editing the mmproj is not optional: it is the vision tower that lets the encoder see your input images. Load with CLIPLoaderGGUF, type qwen_image, with the Qwen3VL-TE add-on installed.
• Or a safetensors version, which loads in the stock Load CLIP node (type qwen_image) with no add-on and includes vision: the FP8 file (9.34 GB) in the same GGUF repo, or qwen3vl_8b_bf16_heretic.safetensors (17.5 GB) from pottokao/Qwen-Image-2.1-Text-Encoder-Heretic, which also links NVFP4 and W4A8 versions. Do not use the four model-0000x-of-00004 shards; they do not load in ComfyUI.

Select it in place of qwen3vl_8b_int8_convrot in the template’s encoder loader and rerun the same seed to compare.

Prompt enhancer. The templates can rewrite your instruction with a 9B prompt enhancer (refine_prompt, off by default). The official one refuses 98 of 100 test prompts in pottokao’s tests, and when it refuses or runs out of tokens, the edit comes back almost unchanged. Leave refine_prompt off and write precise instructions yourself, or use pottokao’s Heretic PE-I2I (refusals about 5/100) in llama.cpp or LM Studio and paste its rewritten prompt into ComfyUI. If you do use the built-in enhancer with thinking on, raise max_length well above 4096 (the model card uses 24000).

Full background on encoders for other models: Uncensored local AI image generation in ComfyUI.

Best Qwen-Image 2.1 Edit LoRAs

LoRAs go in ComfyUI/models/loras/ and load with LoraLoaderModelOnly between the model loader and the sampler. Most Qwen-Image 2.1 LoRAs work for both generation and editing because it is one model.

Never stack two speed LoRAs, and never put a Turbo or Lightning LoRA on the Turbo model, which is already distilled. Check each LoRA’s licence before you publish anything you make with it.

Edit Prompt Recipes

• Be an instruction, not a caption. “Replace the sky with a stormy sunset” works; “a stormy sunset” does not.
• Say what to keep. End with “Keep the person’s face, pose and clothing unchanged” or “Keep everything else identical.”
• Text in images: put the exact words in quotes: “Change the shop sign to read “NIGHT MARKET” in red neon.”
• Style transfer: “Redraw <image1> as a 1970s sci-fi paperback cover painting, keeping the composition.”
• Relighting: “Relight the scene as golden hour from the left, long warm shadows.”
• Old photo restoration (your own family photos): “Restore this photo: remove scratches and dust, fix fading, add natural colour, keep faces unchanged.”

Generate your starting images with Qwen-Image 2.1 Turbo on 16 GB, then animate the edited result with LTX-2.5 or Wan 2.2.

Use It Responsibly

A model that preserves identity across edits is powerful, which is exactly why some edits are off-limits. Edit your own photos, fictional characters, products and scenes. Do not edit images of real people without their clear consent, do not create sexualised or degrading edits of anyone, do not make fake documents, IDs or “evidence”, and label AI-edited images when you publish them. Many countries now have specific laws against non-consensual edited imagery. The Heretic encoder exists to stop false refusals on ordinary creative work, not to get around these rules.

Troubleshooting

• Red missing-node boxes: ComfyUI is too old. Update to the latest release or nightly and restart.
• The edit comes back almost identical to the input: the prompt enhancer refused or ran out of max_length. Turn refine_prompt off, or raise max_length. Also check the prompt is an instruction.
• “12288-wide” shape error with a GGUF encoder: the ComfyUI-GGUF-Qwen3VL-TE add-on is missing.
• Edits ignore the reference images: the GGUF encoder was loaded without its mmproj file, so it cannot see images.
• Very slow edits: resolution is 0 with a large phone photo. Set it to 1024, and shrink large references.
• Content shifts or the framing changes: custom_size is on with a size far from image_1. Turn it off.
• Burned or over-sharp output: CFG above 2, or two speed LoRAs stacked.
• Background removal output breaks the next node: add SplitImageWithAlpha and use the RGB output.
• More: ComfyUI errors and fixes.

FAQ

Is there a separate Qwen-Image 2.1 Edit model?

No. Qwen-Image 2.1 is one model for generation and editing. You edit by giving it input images in the edit templates. The older Qwen-Image-Edit 2509 and 2511 are separate, previous-generation models.

Can Qwen-Image 2.1 edit on 8 GB VRAM?

Yes, with the GGUF Q4_K_M (4.2 GB) or Q5_K_M diffusion model, a GGUF encoder, resolution 1024 and a few references.

How many reference images can I use?

Up to 10 in the official templates. The encoder node has 16 inputs, but the model is documented for 10.

Turbo or base for editing?

Turbo (8 steps) for most edits. The base model (25 to 30 steps) for whole-image changes, hands and fine detail, or with the 4-step Lightning LoRA for speed.

Do Qwen-Image-Edit 2511 LoRAs work on 2.1?

No. 2.1 is a new architecture. Use LoRAs made for Qwen-Image 2.1.

Can I use Qwen-Image 2.1 commercially?

Its Qwen Research License is non-commercial. Check the licence text before any paid use.

Final Thoughts

Qwen-Image 2.1 is the most capable open image editor you can run locally in 2026: one 7B model, 10 references, mark-up editing and transparent cut-outs. On 16 GB, load the Turbo Image Edit template with the INT8 files and you are done; on 8 to 12 GB, use the unsloth GGUF. Set resolution to 1024 while testing, write clear instructions that say what to keep, swap in the Heretic encoder (with its mmproj) if ordinary creative edits get toned down, and add the Lightning, angle or object-removal LoRAs for specific jobs.

Related guides:

• Best Qwen-Image 2.1 Turbo Workflow on 16 GB
• Pruna P-Q-Image-2.1: Qwen-Image 2.1 in 6 Steps
• Uncensored Local AI Image Generation in ComfyUI
• MOSS-Image Edit LoRA in ComfyUI
• LTX-2.5 in ComfyUI on 16 GB

Sources

• Qwen/Qwen-Image-2.1 model card
• ComfyUI docs: Qwen-Image 2.1 templates, edit settings, masks and background removal
• Comfy-Org/Qwen-Image-2.1 file list and unsloth/Qwen-Image-2.1-GGUF
• pottokao Heretic text encoder and Heretic PE-I2I model cards
• ComfyUI-GGUF-Qwen3VL-TE and city96/ComfyUI-GGUF
• LoRA model cards: lightx2v Lightning, akhaliq Multiple-Angles and Next-Scene, prithivMLmods Object-Remover/Mover and Natural-Exposure, e-n-v-y Fix v2.0
• Hugging Face model search for “qwen-image-2.1” and “Qwen-Image-Edit”, sorted by downloads and date (checked November 2, 2026)