Alibaba’s Qwen team has shipped image models faster than most people can keep track of: Qwen-Image-Edit 2509, Qwen-Image-Edit 2511, Qwen-Image 2512 and now Qwen-Image 2.1, plus Turbo and Lightning versions of each. The numbers look like versions of one model, but they are not. One of them is not an editor at all, one uses a completely different text encoder and VAE, and one comes with a non-commercial licence. This guide compares Qwen Image Edit 2509 vs 2511 vs 2512 vs 2.1 side by side, with the exact ComfyUI files and folders, VRAM needs, LoRA compatibility, and a clear answer on which one to use for your job.
- 2509 (Sep 2025) and 2511 (22 Dec 2025) are edit models, 20B, Apache 2.0, up to 3 reference images. Use 2511; keep 2509 only for LoRAs trained on it.
- 2512 (Dec 2025) is a text-to-image model, not an editor. It is the updated Qwen-Image generator.
- 2.1 (20 Sep 2026, Turbo on 9 Oct 2026) is one 7B model that both generates and edits, takes many more reference images, and fits 16 GB easily. But it uses the Qwen Research License (non-commercial).
- 2509, 2511 and 2512 share
qwen_2.5_vl_7b+qwen_image_vae. 2.1 needs its ownqwen3vl_8bencoder andqwen_image_2.1_vae. - Commercial work: Edit 2511 + Qwen-Image 2512. Personal work on 8-16 GB: Qwen-Image 2.1 (Turbo).
Quick Comparison Table
| Edit 2509 | Edit 2511 | Qwen-Image 2512 | Qwen-Image 2.1 | |
|---|---|---|---|---|
| Job | Editing | Editing | Text to image | Generation and editing in one model |
| Released | September 2025 | 22 December 2025 | December 2025 | 20 Sep 2026 (Turbo 9 Oct 2026) |
| Diffusion model size | 20B | 20B | 20B | 7B (32 single-stream DiT layers) |
| Licence | Apache 2.0 | Apache 2.0 | Apache 2.0 | Qwen Research License (non-commercial) |
| Text encoder | Qwen2.5-VL 7B | Qwen2.5-VL 7B | Qwen2.5-VL 7B | Qwen3-VL 8B |
| VAE | qwen_image_vae |
qwen_image_vae |
qwen_image_vae |
qwen_image_2.1_vae |
| Reference images | Up to 3 | Up to 3 | None (prompt only) | Many; the ComfyUI node opens inputs up to 16 |
| FP8 file size | 20.4-20.5 GB | 20.5 GB | 20.4 GB | bf16 is about 14.2 GB; INT8 and GGUF are smaller |
| Best for | Old 2509 LoRAs | Commercial edits, consistency | Commercial text to image | Personal use, 8-16 GB GPUs, many references |
The Qwen Image Family Explained (Timeline)
The four-digit numbers are dates, year-month: 2509 is September 2025, 2511 is November-December 2025 and 2512 is December 2025. They are updates to the original 20B Qwen-Image line released in August 2025, which split into two branches:
- Generation branch: Qwen-Image, then Qwen-Image 2512.
- Editing branch: Qwen-Image-Edit, then Edit 2509, then Edit 2511.
Qwen-Image 2.1 is a new generation, not a fifth update. It is a smaller 7B diffusion transformer with a newer Qwen3-VL 8B encoder, and it merges both branches into a single checkpoint: the Comfy docs describe it as handling “both generation and editing in a single checkpoint”. That is why 2.1 needs different files from everything before it.
Qwen-Image-Edit 2509
Edit 2509 was the release that made Qwen editing popular in ComfyUI. Its headline features were multi-image editing (combine up to three images, for example a person and a product, or two people) and better consistency than the first Qwen-Image-Edit. In ComfyUI you feed the images through the TextEncodeQwenImageEditPlus node, which has three image inputs.
Today there is little reason to start with 2509. The main exception is LoRAs: edit LoRAs are usually trained for one exact version, and a large library of community LoRAs was made for 2509 in late 2025. If a LoRA you rely on says “2509”, keep this model for it.
Qwen-Image-Edit 2511
Qwen released Edit 2511 on 22 December 2025. Its official notes list four improvements over 2509:
- “mitigate image drift, improved character consistency”: faces and products change less across edits, including single-person and two-person group edits.
- “integrated LoRA capabilities”: popular community LoRAs, such as lighting enhancement and new-viewpoint generation, are built into the base model.
- “enhanced industrial design generation”.
- “strengthened geometric reasoning ability”.
It uses the same text encoder, VAE and TextEncodeQwenImageEditPlus node as 2509, so you can switch a 2509 workflow to 2511 by changing only the diffusion model file. It remains Apache 2.0, which makes it the best Qwen editor for client and commercial work.
Qwen-Image 2512 (Text to Image)
This is the version people confuse most. Qwen-Image 2512 is not an edit model. It is the December 2025 update of the original Qwen-Image text-to-image model, known for strong text rendering in images. If you search for “Qwen Image Edit 2512”, what you want is either Edit 2511 (for editing) or Qwen-Image 2512 (for generating new images from a prompt). It shares the encoder and VAE with the edit models, so one set of support files covers all of them. Unless you need the original August 2025 Qwen-Image for an old LoRA, use 2512 for generation.
Qwen-Image 2.1 and 2.1 Turbo
Qwen-Image 2.1 arrived as open weights on 20 September 2026, and the Turbo checkpoint followed on 9 October 2026. What changed:
- One model for generation and editing. You no longer need separate 2512 and 2511 files.
- Smaller: a 7B diffusion transformer instead of 20B, so it fits 16 GB without offloading, and GGUF versions run on 8 GB.
- New encoder: Qwen3-VL 8B, which actually sees your input images.
- More references: the ComfyUI Text Encode Qwen Image 2.1 node opens new image inputs as you fill them, up to 16, and supports painted marks to show what to change and background removal with transparency.
- Turbo runs in 8 steps at CFG 1 instead of the base model’s 25-40 steps. It is a speed version, not a quality tier.
Turbo settings: 8 steps, CFG 1, euler, with the fixed schedule from the official template. Base settings: 25 steps (30 for hands and fine detail), CFG 1. Full setup details are in my Qwen-Image 2.1 edit guide for 8-16 GB and the Qwen-Image 2.1 Turbo workflow.
The Licence Difference That Matters
This is the single most important difference for anyone who sells images or works for clients:
- Edit 2509, Edit 2511 and Qwen-Image 2512: Apache 2.0. Commercial use is allowed.
- Qwen-Image 2.1 and Turbo: Qwen Research License Agreement. It allows royalty-free use for non-commercial purposes only and asks products built with it to show “Built with Qwen”. Commercial use needs a separate licence from Alibaba. Earlier models keep their Apache 2.0 licence; the change is not retroactive.
So for paid work, stay on 2511 and 2512 (or another Apache editor such as the MOSS-Image Edit LoRA). For personal projects, learning and testing, 2.1 is the better and lighter model. LoRAs have their own licences too, so check each one before publishing anything made with it.
Files and ComfyUI Folders for Each Version
Everything before 2.1 shares one text encoder and one VAE. Download them once and they work for 2509, 2511 and 2512:
| File | Size | Used by | Folder |
|---|---|---|---|
qwen_2.5_vl_7b_fp8_scaled.safetensors |
9.38 GB | 2509, 2511, 2512 (recommended encoder) | ComfyUI/models/text_encoders/ |
qwen_2.5_vl_7b_nvfp4.safetensors |
6.11 GB | Same, for RTX 50 cards | ComfyUI/models/text_encoders/ |
qwen_image_vae.safetensors |
254 MB | 2509, 2511, 2512 | ComfyUI/models/vae/ |
qwen_image_edit_2509_fp8mixed.safetensors |
20.5 GB | Edit 2509 | ComfyUI/models/diffusion_models/ |
qwen_image_edit_2511_fp8mixed.safetensors |
20.5 GB | Edit 2511 (recommended) | ComfyUI/models/diffusion_models/ |
qwen_image_2512_fp8_e4m3fn.safetensors |
20.4 GB | Qwen-Image 2512 | ComfyUI/models/diffusion_models/ |
Qwen-Image 2.1 needs its own set, all in the Comfy-Org/Qwen-Image-2.1 repo:
| File | Purpose | Folder |
|---|---|---|
qwen_image_2.1_turbo_int8_convrot.safetensors |
Turbo (8 steps), low memory | ComfyUI/models/diffusion_models/ |
qwen_image_2.1_int8_convrot.safetensors |
Base (25 steps), low memory | ComfyUI/models/diffusion_models/ |
qwen_image_2.1_bf16.safetensors |
Full precision (about 14.2 GB) | ComfyUI/models/diffusion_models/ |
qwen3vl_8b_int8_convrot.safetensors |
Qwen3-VL 8B text encoder | ComfyUI/models/text_encoders/ |
qwen_image_2.1_vae_bf16.safetensors |
2.1 VAE | ComfyUI/models/vae/ |
Never mix the two sets: the 2.1 model with the old encoder or VAE (or the reverse) gives errors or garbage. The full list, including GGUF and bf16 variants, is in my Qwen-Image model files guide. GGUF diffusion models load with Unet Loader (GGUF) from ComfyUI-GGUF; see my ComfyUI-GGUF guide. Update ComfyUI first: the 2.1 nodes are core nodes that only exist in recent releases (0.36.0 or newer), and the official templates (“Qwen Image 2.1: Image Edit”, “Qwen Image 2.1 Turbo: Image Edit”, “Remove Background: Qwen Image 2.1”) are the fastest way to start.
VRAM: Which Version Fits Your GPU
| GPU | Edit 2511 / Qwen-Image 2512 (20B) | Qwen-Image 2.1 (7B) |
|---|---|---|
| 8 GB | GGUF Q4/Q5 diffusion model, GGUF or fp8 encoder, slow | GGUF Q4_K_M (4.2 GB) or Q5_K_M (5.39 GB), comfortable |
| 12 GB | GGUF Q5/Q6, or fp8 with offloading | INT8 model + INT8 encoder |
| 16 GB | fp8mixed (20.5 GB) partly offloaded, or NVFP4 on RTX 50 | INT8 fully on the GPU, or bf16 with the encoder offloaded |
| 24 GB+ | fp8 fully on the GPU | bf16, many large references |
In practice, 2.1 is the far easier model on consumer cards. The 20B models need about 20 GB just for FP8 weights, so on 16 GB they rely on offloading, which slows each step. If you keep running out of memory, my ComfyUI VRAM guide and RTX 50 setup guide (for NVFP4) help.
LoRA Compatibility and Speed LoRAs
- LoRAs are version-specific. A 2509 LoRA may work loosely on 2511 because they share an architecture, but results drift. A 2509, 2511 or 2512 LoRA does not work on 2.1, which is a different 7B model.
- Speed LoRAs: the lightx2v team publishes 4-step and 8-step Lightning LoRAs for the 20B Qwen-Image and Edit models, and
lightx2v/Qwen-Image-2.1-Lightning(4 steps, CFG 1, October 2026) for 2.1 base. The official 2.1 Turbo LoRA and the Pruna P-Q LoRAs give 8, 6 or 5 steps; see my Pruna 6-step guide. - Never stack two speed LoRAs, and never add a Lightning or Turbo LoRA to the 2.1 Turbo checkpoint, which is already distilled.
- Load LoRAs with LoraLoaderModelOnly between the model loader and the sampler; files go in
ComfyUI/models/loras/.
Which One Should You Use?
| Your situation | Use |
|---|---|
| Personal edits on 8-16 GB, want speed | Qwen-Image 2.1 Turbo |
| Personal edits, maximum detail | Qwen-Image 2.1 base, 25-30 steps |
| Combining more than three reference images | Qwen-Image 2.1 |
| Client or commercial editing | Qwen-Image-Edit 2511 (Apache 2.0) |
| Commercial text to image with strong text rendering | Qwen-Image 2512 (Apache 2.0) |
| A LoRA that was trained for 2509 | Qwen-Image-Edit 2509 |
| Uncensored prompt following | See my Heretic text encoder guide |
Errors and Fixes
Missing node: TextEncodeQwenImageEditPlus or Text Encode Qwen Image 2.1
Both are core ComfyUI nodes, not custom nodes, so ComfyUI Manager cannot install them. Update ComfyUI itself (the 2.1 nodes need 0.36.0 or newer): in the portable build run update/update_comfyui.bat, in a git install run git pull inside the ComfyUI folder, or use the desktop app’s update button. Restart and reload the workflow.
Errors like “mat1 and mat2 shapes cannot be multiplied”, or noisy garbage images
This almost always means files from the two families are mixed. 2509, 2511 and 2512 need qwen_2.5_vl_7b + qwen_image_vae; Qwen-Image 2.1 needs qwen3vl_8b + qwen_image_2.1_vae. Check all three loaders (model, text encoder, VAE) against the tables above.
Out of memory with Edit 2511 or 2512 on 16 GB
The fp8 files are about 20.5 GB, larger than a 16 GB card. Close other GPU apps, let ComfyUI offload (do not force everything to the GPU), switch to a Q4 or Q5 GGUF with ComfyUI-GGUF, or use the NVFP4 encoder on RTX 50 cards. If you only need personal edits, Qwen-Image 2.1 INT8 fits far more easily.
“lora key not loaded” warnings and the LoRA does nothing
The LoRA was trained for a different version. 20B LoRAs (2509, 2511, 2512) do not work on 2.1, and 2509 LoRAs may only partly work on 2511. Use a LoRA made for the exact model you loaded.
Blurry, washed-out or burnt results
Usually stacked speed LoRAs or the wrong steps and CFG. Use one speed LoRA at most, never add Lightning or Turbo LoRAs to the 2.1 Turbo checkpoint, and match its settings (Turbo: 8 steps, CFG 1; 2.1 base: 25-30 steps, CFG 1).
The edited face or product drifts away from the original
Move from 2509 to 2511 (better consistency), give the reference image at a good resolution, and describe only the change you want (“change the jacket to red, keep everything else the same”). With 2.1, painted marks on the image help target the edit.
Frequently Asked Questions
What is the difference between Qwen Image Edit 2509 and 2511?
2511, released on 22 December 2025, reduces image drift, improves character consistency, builds in popular LoRA abilities such as lighting and new viewpoints, and improves geometric reasoning. Both take up to three images and use the same encoder and VAE.
Is Qwen Image 2512 an edit model?
No. Qwen-Image 2512 is the December 2025 update of the Qwen-Image text-to-image model. For editing, use Qwen-Image-Edit 2511 or Qwen-Image 2.1.
Is Qwen-Image 2.1 better than Edit 2511?
For most personal use, yes: it is a single 7B model that generates and edits, accepts more reference images and fits 8-16 GB GPUs far more easily. But it is licensed for non-commercial use only, while 2511 is Apache 2.0.
Can I use Qwen-Image 2.1 commercially?
Not under its default licence. Qwen-Image 2.1 ships under the Qwen Research License, which allows non-commercial use only; commercial use requires a separate licence from Alibaba.
Do Qwen Image Edit 2509 and 2511 use the same text encoder?
Yes. 2509, 2511 and Qwen-Image 2512 all use qwen_2.5_vl_7b (fp8_scaled recommended) and qwen_image_vae.safetensors. Qwen-Image 2.1 uses qwen3vl_8b and its own 2.1 VAE.
Do 2511 LoRAs work with Qwen-Image 2.1?
No. Qwen-Image 2.1 is a different 7B architecture, so LoRAs trained on the 20B 2509, 2511 or 2512 models do not work on it. Use LoRAs made for 2.1.
Which Qwen image model is best for 8 GB VRAM?
Qwen-Image 2.1 as a GGUF Q4_K_M (4.2 GB) or Q5_K_M (5.39 GB) with a quantized encoder. The 20B models run on 8 GB only with heavy quantization and offloading.
Final Thoughts
The short version: 2509 and 2511 are the old 20B editors (use 2511), 2512 is the old 20B generator, and 2.1 replaces both with one lighter 7B model, at the cost of a non-commercial licence. Pick by licence first and VRAM second. Then follow the matching setup: the Qwen-Image 2.1 edit guide for 2.1, or the Qwen-Image files guide for 2511 and 2512. For other open image models, see the best open source text-to-image models.
Sources
- Qwen docs: Qwen-Image-Edit-2511 (release and improvements)
- Comfy: Qwen-Image 2.1 (unified model, references, licence)
- OrcaRouter: Qwen-Image 2.1 release dates, licence and file sizes
- Comfy-Org/Qwen-Image_ComfyUI (2509, 2511, 2512 files)
- Comfy-Org/Qwen-Image-2.1 (2.1 files)
- lightx2v/Qwen-Image-2.1-Lightning
