Qwen-Image-Edit Locally: Free Photoshop-Grade AI Editing in ComfyUI

Qwen-Image-Edit-2511 is the strongest image editor you can run locally for free, and you do not need a monster GPU: grab the Q4_K_M GGUF (13.2GB) for a 16GB card or Q4_0 (11.9GB) for 12GB, load ComfyUI's built-in image_qwen_image_edit_2511 template, and add the 850MB Lightning LoRA to cut 40 sampling steps to 4. The FP8 build is 20.5GB and wants a 24GB card; the full BF16 checkpoint is 40.9GB. Everything — model, quants, LoRA — is Apache 2.0. Total cost: $0.

That is the whole recommendation. The rest of this page is the detail: a VRAM-tier table built from the actual file sizes on Hugging Face (Qwen publishes no official VRAM numbers, so anyone quoting one is guessing — we show our working instead), the exact files and folders for both setup routes, and the limitations nobody puts in a launch thread. One scope note before we start: this is the editing model — you feed it images and tell it what to change. For generating images from scratch, see our Qwen-Image local guide; the two are siblings, not the same model.

---

What Qwen-Image-Edit-2511 Actually Is {#what-it-is}

A 20B-parameter, Apache 2.0 model that takes one or more input images plus a plain-language instruction and returns the edited result. Released December 23, 2025, it is the third snapshot of the Qwen-Image-Edit line and the one the local community has standardized on.

"Photoshop-grade" is a big claim, so here is what it concretely means: instead of masking, layering, and cloning, you write "replace the background with a rainy street at night" or "put the person from image 1 into the scene from image 2" and the model does the compositing — identity, lighting, and perspective included. Multi-image fusion (combining people or objects from separate photos into one coherent image) is the headline capability, per the official model card.

What changed versus the previous 2509 snapshot, straight from the model card:

• Character consistency in portrait edits — the person still looks like the same person after the edit.
• Multi-person consistency — group-photo synthesis from separate source images, previously the fastest way to get uncanny results.
• Popular community LoRA behaviors built in — relighting and novel-viewpoint generation ship in the base model now. (If you want to train your own on top, our image LoRA training guide covers the workflow.)
• Stronger geometric reasoning and industrial-design editing — construction lines, annotations, material swaps.
• "Mitigated image drift" — Qwen's own phrasing, and an honest admission that earlier snapshots subtly altered regions you never asked it to touch.

Adoption is easy to measure and worth quoting because it tells you where the tutorials, workflows, and bug fixes are: unsloth's GGUF repo alone shows ~220K downloads, lightx2v's Lightning LoRA repo showed 303K downloads in the last month, and the QwenLM/Qwen-Image GitHub repo sits at 8.2K stars (all Hugging Face/GitHub figures, August 2026). In the local image-editing niche, this is the model with gravity right now.

---

Which Build Fits Your GPU {#vram-requirements}

The short version: 12GB → Q4_0 GGUF (11.9GB). 16GB → Q4_K_M (13.2GB) or Q5_K_S (14.3GB). 24GB → FP8 (20.5GB) or Q8_0 (21.8GB). 40GB+ → BF16 (40.9GB). Below 12GB it technically runs, but you are trading real quality for it.

Qwen publishes no official VRAM table for this model. What is verifiable are the file sizes on the Hugging Face repos, so this table is built from those, plus one practical rule: ComfyUI can offload weights to system RAM when a file exceeds your VRAM, so a slightly-too-big file still runs — just slower. The fit judgments in the right column are ours.

File sizes: unsloth/Qwen-Image-Edit-2511-GGUF and Comfy-Org/Qwen-Image-Edit_ComfyUI, August 2026.

Two things the table hides. First, the text encoder: every route also needs the 9.38GB qwen_2.5_vl_7b_fp8_scaled encoder, but ComfyUI runs it before the diffusion model and releases it, so it does not stack on top of the numbers above — it does mean you want system RAM to hold it. 32GB system RAM is our practical floor, 64GB is comfortable — a rule of thumb, not an official spec. Second, edits at high resolution or with three input images push activation memory up; if you are borderline, be borderline with the smaller file.

Shopping for a card for image work? Our best GPU for image generation ranking covers the price-per-VRAM math, and if you already own a 16GB card, what else it can run is worth a look.

---

ComfyUI Setup: The Standard Route (FP8/BF16) {#comfyui-setup}

Four downloads into four folders, then load the bundled image_qwen_image_edit_2511 template from ComfyUI's template browser. No custom nodes required — support is native.

This is the route for 24GB+ cards (FP8) or workstation GPUs (BF16). New to ComfyUI entirely? Start with our ComfyUI complete guide first — this section assumes a working, updated install (the 2511 nodes need a current build; update via the manager or git pull if the template is missing).

The files, per the official ComfyUI tutorial:

Download sources: the diffusion model from Comfy-Org/Qwen-Image-Edit_ComfyUI, the encoder and VAE from Comfy-Org repackage repos, and the Lightning LoRA from lightx2v/Qwen-Image-Edit-2511-Lightning. The URLs below are the exact ones the official tutorial links (yes, the encoder really lives in a HunyuanVideo repackage repo — Comfy-Org reuses it across models):

cd ComfyUI/models

wget -P diffusion_models https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/resolve/main/split_files/diffusion_models/qwen_image_edit_2511_fp8mixed.safetensors

wget -P text_encoders https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors

wget -P vae https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors

wget -P loras https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning/resolve/main/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors

Then in ComfyUI: Workflow → Browse Templates → search "Qwen" → image_qwen_image_edit_2511. Load your image — or several, since multi-image fusion is a headline capability of the model — type your instruction, queue. The template ships preconfigured; with the Lightning LoRA enabled it samples in 4 steps, and without it the reference configuration is 40 steps with a true-CFG scale of 4.0 (the setting Qwen's own diffusers example uses). Our advice: leave the template's sampler settings alone until you have a reason not to.

One practical note on the two checkpoint choices: fp8mixed at 20.5GB is what almost everyone should run. BF16 at 40.9GB is the reference, but on anything below a 48GB card it will be offloading constantly, and half the point of a 4-step LoRA is lost if every step swaps weights over PCIe.

---

The GGUF Route: 8-16GB Cards {#gguf-route}

Install city96's ComfyUI-GGUF custom node, drop a .gguf file into ComfyUI/models/unet/, and swap the template's diffusion-model loader for the GGUF Unet loader. Same encoder, VAE, and LoRA as above.

GGUF quantization is what makes a 40.9GB model realistic on a 12GB card. The community-standard files come from unsloth/Qwen-Image-Edit-2511-GGUF (built with city96's ComfyUI-GGUF tooling). The full menu, exact sizes from the repo:

Setup, assuming the standard route's encoder/VAE/LoRA are already in place:

# 1. Install the GGUF loader node
cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF

# 2. Put the quant where the loader looks (note: models/unet, not diffusion_models)
wget -P ../models/unet https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/resolve/main/qwen-image-edit-2511-Q4_K_M.gguf

Restart ComfyUI, load the same image_qwen_image_edit_2511 template, delete the "Load Diffusion Model" node, and replace it with the GGUF Unet loader (city96 files it under the "bootleg" category — that is the right one, not a knockoff). Wire it to the same connections, pick your .gguf, done. The node pack also ships GGUF variants of the CLIP/text-encoder loaders if you want to squeeze the encoder too, though with the encoder unloading after the encode step, the standard fp8 encoder is fine for most setups.

Honest advice on picking a quant: Q4 is the floor where results stay genuinely usable; below it, this particular model loses fine detail and edit precision fast — that is our judgment from testing quants of this class of model, not a published benchmark, so spend 20 minutes comparing Q4 against Q3 on your own images before committing a workflow to the smaller file. And if you are on a 12GB card wondering what else it can do on the generation side, our 12GB Stable Diffusion picks page covers that.

---

Honest Limitations {#limitations}

It is the best local image editor available, and it is also a 20B model with quarterly-snapshot churn, no official VRAM guidance, and edit drift that is "mitigated," not solved. Know what you are signing up for.

• The downloads are enormous. The sensible FP8 route is ~31GB of files (20.5 + 9.38 + 0.25 + 0.85); BF16 pushes past 51GB. Budget disk and bandwidth accordingly.
• It is slow without Lightning. 40 steps of a 20B diffusion model on a midrange card is a coffee break, not an iteration loop. The 4-step LoRA is what makes local use pleasant — and distilled 4-step output is not pixel-identical to the 40-step reference. lightx2v's claim is ~10x speedup with maintained quality; treat "maintained" as their claim and judge on your own images.
• Drift is reduced, not gone. Qwen's release notes for 2511 specifically call out "mitigated image drift." Our reading: expect regions you did not ask about to shift subtly on aggressive or repeated edits, especially at lower quants. Keep originals; compare before saving.
• No official hardware guidance exists. Every VRAM table for this model — including ours above — is derived, not published. File sizes are the only hard numbers.
• Snapshot churn is real. 2509 arrived September 22, 2025; 2511 on December 23, 2025. Filenames embed the snapshot, so workflows and tutorials go stale roughly quarterly (more in the next section).
• Mac users: tempered expectations. The official diffusers examples target CUDA with a CPU fallback — there is no Apple-specific path in Qwen's code. ComfyUI itself runs on Apple Silicon, but a 20B diffusion model on unified memory is a patience exercise. We would not buy a Mac for this workload.

None of these are reasons to skip it. They are reasons to run Q4/FP8 with the Lightning LoRA and keep your expectations calibrated to "excellent local tool," not "cloud service with zero tradeoffs."

---

Edit Model vs Generation Model vs FLUX.2 {#vs-alternatives}

Use Qwen-Image-Edit-2511 to change images that exist. Use a generation model — Qwen-Image, FLUX.2, or Z-Image Turbo — to create images that do not. The confusion is understandable because the names are nearly identical, but mixing them up wastes a 20-40GB download.

• You have photos and want to modify, combine, or restyle them → this page. Instruction-based editing, identity preservation, multi-image fusion.
• You want text-to-image with the best text rendering in the open-weight world (posters, UI mocks, signage, Chinese text) → Qwen-Image, the generation-side 20B sibling.
• You want the strongest general text-to-image aesthetics and a deep LoRA ecosystem → FLUX.2, the other heavyweight in local image AI right now.
• You want fast, light generation on modest hardware and can trade some quality for speed → Z-Image Turbo.

In practice the pairing we see most is a generation model plus this editor: generate with whatever you prefer, then fix, composite, and iterate with Qwen-Image-Edit. Since everything here runs in the same ComfyUI install, that is one workflow, not two tools.

---

Version Churn: Reading the Snapshot Numbers {#versions}

Qwen ships the Edit line as dated snapshots — 2509 (September 22, 2025), then 2511 (December 23, 2025) — and the generation line separately (Qwen-Image-2512 landed December 31, 2025). Expect a successor to 2511; the setup in this guide carries over, the filenames will not.

The number is the year-month of the training snapshot, and each release so far has been a drop-in replacement architecturally: same 20B scale, same encoder and VAE, same template structure — you swap one diffusion-model file. So when 2513-or-whatever lands, the fastest way to confirm what is current is ComfyUI's own template browser (Comfy-Org has shipped day-one templates for each snapshot so far) and the news section of the QwenLM/Qwen-Image GitHub repo. We will update this page when the Edit line moves; the VRAM math only changes if the parameter count does.

---

Verdict {#verdict}

If you have a 12GB+ GPU and any recurring need to edit or composite images, this is the easiest "install it this weekend" recommendation in local AI right now.

1. 16GB card (most readers): Q4_K_M GGUF + Lightning 4-step LoRA. ~24GB of downloads total, genuinely good results, pleasant iteration speed.
2. 12GB card: Q4_0. Same setup, one size down. This is the floor where we would say quality is still worth the effort.
3. 24GB card: FP8 fp8mixed. Near-reference quality without BF16's offloading pain.
4. 8GB card: Honestly — skip it, or accept Q2/Q3 as a toy. Put the effort into lighter generation models instead.
5. Everyone: take the free route end to end before considering any paid editing service. Apache 2.0 licensing, zero per-image cost, and your photos never leave your machine — for an editing workload, the privacy point alone is worth more than usual.

---

Sources {#sources}

• Qwen/Qwen-Image-Edit-2511 model card — parameters, license, capabilities, reference inference settings
• QwenLM/Qwen-Image on GitHub — release dates for 2509 / 2511 / 2512
• ComfyUI official Qwen-Image-Edit-2511 tutorial — file names, folders, template
• Comfy-Org/Qwen-Image-Edit_ComfyUI — BF16/FP8 checkpoint sizes
• unsloth/Qwen-Image-Edit-2511-GGUF — GGUF quant sizes and downloads
• lightx2v/Qwen-Image-Edit-2511-Lightning — Lightning LoRA files and speedup claim
• city96/ComfyUI-GGUF — GGUF loader node, folder conventions

All sizes and download counts read from the repos in August 2026.

---

FAQ {#faq}