Qwen-Image Model Files Explained: qwen_image_vae, qwen_2.5_vl_7b and Every File (2026)

If you searched for qwen_image_vae.safetensors or qwen_2.5_vl_7b_fp8_scaled.safetensors, you are probably setting up Qwen-Image in ComfyUI and trying to work out which files you actually need, where they go, and why the workflow still shows red nodes. This guide lists every official Qwen-Image file for ComfyUI in one place: the diffusion models, the text encoders and the VAE for Qwen-Image, Qwen-Image 2512, Qwen-Image-Edit 2509 and 2511, and the new Qwen-Image 2.1, which uses a different text encoder and VAE. You also get exact folder paths, file sizes, which version to pick for your VRAM, and the errors you get when the files are mixed up.

Which Qwen-Image Version Do You Have?

Alibaba’s Qwen team has released several image models under the Qwen-Image name. They fall into two families, and the family decides which text encoder and VAE you need:

The rule is simple: everything before 2.1 shares one text encoder and one VAE. Download those two files once and they work for Qwen-Image, 2512 and all three Edit versions. Qwen-Image 2.1 needs its own pair.

Folder Layout at a Glance

Here is what a complete setup looks like with the recommended fp8 files for the older family plus Qwen-Image 2.1:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   ├── qwen_image_2512_fp8_e4m3fn.safetensors        (20.4 GB)
    │   ├── qwen_image_edit_2511_fp8mixed.safetensors     (20.5 GB)
    │   └── qwen_image_2.1_int8_convrot.safetensors       (Qwen-Image 2.1)
    ├── text_encoders/
    │   ├── qwen_2.5_vl_7b_fp8_scaled.safetensors         (9.38 GB, older family)
    │   └── qwen3vl_8b_int8_convrot.safetensors           (Qwen-Image 2.1)
    ├── vae/
    │   ├── qwen_image_vae.safetensors                    (254 MB, older family)
    │   └── qwen_image_2.1_vae_bf16.safetensors           (Qwen-Image 2.1)
    └── loras/
        └── (Lightning / speed LoRAs, optional)

On the Windows portable build the full path starts with ComfyUI_windows_portable\ComfyUI\models\. On ComfyUI Desktop it is inside your install’s models folder. If you keep models elsewhere with extra_model_paths.yaml, the same subfolder names apply. My ComfyUI model folders cheat sheet covers every folder in detail.

qwen_image_vae.safetensors (and the 2.1 VAE)

The VAE turns the model’s latent output into pixels (and, for editing, your input image into latents). There is only one version of qwen_image_vae.safetensors, so you never need to choose a precision. Load it with the Load VAE node.

Several other models reuse the Qwen-Image VAE too, which is why you will see qwen_image_vae.safetensors in workflows for models that are not Qwen at all, such as Krea 2 and some Anima workflows. If a workflow asks for it, this is the same file.

qwen_2.5_vl_7b Text Encoder Files

Qwen-Image does not use CLIP or T5. Its “text encoder” is a full vision-language model, Qwen2.5-VL 7B, which reads your prompt (and, for Edit models, your input images). That is why the file is so big.

All three go in ComfyUI/models/text_encoders/ and are downloaded from the text_encoders folder of Comfy-Org/Qwen-Image_ComfyUI. Load them with the Load CLIP node and set its type to qwen_image.

The text encoder runs once per prompt and then ComfyUI can move it back to system RAM, so on a 16 GB card it is fine to use the 9.38 GB fp8 file even with a 20 GB diffusion model; ComfyUI swaps them. What you need is enough system RAM (32 GB minimum, 64 GB comfortable). If you run out, you will get the classic “Reconnecting” crash; see my ComfyUI Reconnecting fix.

Qwen-Image and Qwen-Image 2512 Diffusion Models

From Comfy-Org/Qwen-Image_ComfyUI, all into ComfyUI/models/diffusion_models/:

Unless you need the original release, use 2512; it is the newer of the two. Load it with the Load Diffusion Model node (weight dtype default). For a speed boost, the lightx2v Qwen-Image-Lightning LoRAs (for example Qwen-Image-Lightning-8steps-V1.0.safetensors in ComfyUI/models/loras/) cut generation to 8 or 4 steps.

Qwen-Image-Edit, 2509 and 2511 Files

From Comfy-Org/Qwen-Image-Edit_ComfyUI, all into ComfyUI/models/diffusion_models/. They use the same qwen_2.5_vl_7b text encoder and qwen_image_vae as above.

2509 vs 2511: both take up to three reference images with the TextEncodeQwenImageEditPlus node. 2511 is the newer one with better person and product consistency, so start there. Keep 2509 only if a LoRA you rely on was trained for it, since edit LoRAs are usually version-specific.

Qwen-Image 2.1 Files (New Encoder and VAE)

Qwen-Image 2.1 is a newer, smaller model and it ships with a different text encoder (Qwen3-VL 8B) and its own VAE. All files are in Comfy-Org/Qwen-Image-2.1:

For a full 2.1 walkthrough, including multi-image editing, see my Qwen-Image 2.1 edit guide; for faster generation, the Pruna 6-step LoRAs and the Qwen-Image 2.1 Turbo workflow for 16 GB.

Which Files to Download for Your VRAM

In all cases plan for at least 32 GB of system RAM, because ComfyUI moves the text encoder and parts of the model between RAM and VRAM. My how much VRAM for ComfyUI post explains how this offloading works on 16 GB cards.

GGUF Versions for 8-12 GB GPUs

If the fp8 files are too big, GGUF quantized versions shrink the diffusion model and text encoder a lot. You need the ComfyUI-GGUF custom node (search “ComfyUI-GGUF” in ComfyUI-Manager, or install manually):

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF
# portable build: run from the ComfyUI_windows_portable folder
.\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-GGUF\requirements.txt

• Diffusion model: GGUF files such as those in QuantStack/Qwen-Image-GGUF or QuantStack/Qwen-Image-Edit-2509-GGUF go in ComfyUI/models/unet/ (or diffusion_models/) and load with Unet Loader (GGUF).
• Text encoder: a Qwen2.5-VL-7B-Instruct GGUF (for example from Unsloth) goes in ComfyUI/models/text_encoders/ and loads with CLIPLoader (GGUF), type qwen_image. For Edit models, which read images, also put the matching mmproj GGUF file in the same folder, or image input will fail.
• VAE: still the normal qwen_image_vae.safetensors; there is no GGUF VAE.

Q4_K_M or Q5_K_M is the usual sweet spot. My GGUF quantization guide compares the quality of each level.

Which Loader Node Loads Which File

The easiest way to get every node right is to open the official template: Workflow > Browse Templates > Image and pick Qwen-Image, Qwen-Image-Edit or Qwen-Image 2.1. ComfyUI then shows a download dialog for any missing file.

Common Errors From the Wrong File

• Red “Value not in list” on a loader: the file is not in the folder that node reads, or the name differs. Check the folder from the table above, then press R in ComfyUI to refresh the lists.
• Black, grey or noisy output: usually the wrong VAE, for example qwen_image_vae with Qwen-Image 2.1 or the 2.1 VAE with the older models. Match the VAE to the model family.
• mat1 and mat2 shapes cannot be multiplied at the text encoder: the wrong encoder for the model (Qwen2.5-VL with 2.1, or Qwen3-VL with the older family), the wrong CLIP type, or, with GGUF Edit workflows, a missing or mismatched mmproj file.
• Pure noise with fp8 files: if you launch ComfyUI with --fast fp8_matrix_mult, some unscaled fp8 files break. Remove the flag or use the fp8mixed/int8 versions.
• NVFP4 file fails or is slow: NVFP4 is for RTX 50 series cards. On RTX 30/40, use fp8 or GGUF instead.
• Console closes / “Reconnecting” while loading: system RAM ran out. Increase the page file and use smaller files; see my Reconnecting fix guide.

More error messages and their fixes are in my ComfyUI errors and fixes list.

Frequently Asked Questions

Where do I put qwen_image_vae.safetensors?

In ComfyUI/models/vae/. Load it with the Load VAE node. It is 254 MB and is shared by Qwen-Image, Qwen-Image 2512 and all Qwen-Image-Edit versions.

Where do I put qwen_2.5_vl_7b_fp8_scaled.safetensors?

In ComfyUI/models/text_encoders/. Load it with the Load CLIP node and set the type to qwen_image.

Do Qwen-Image-Edit 2509 and 2511 use the same VAE and text encoder?

Yes. Both use qwen_image_vae.safetensors and the qwen_2.5_vl_7b text encoder, the same files as Qwen-Image and Qwen-Image 2512. Only the diffusion model file changes.

Does Qwen-Image 2.1 use qwen_image_vae?

No. Qwen-Image 2.1 uses its own qwen_image_2.1_vae_bf16.safetensors and a Qwen3-VL 8B text encoder such as qwen3vl_8b_int8_convrot.safetensors.

Which Qwen text encoder file should I download?

For the older family, qwen_2.5_vl_7b_fp8_scaled.safetensors (9.38 GB) suits most PCs. Use the nvfp4 version (6.11 GB) on RTX 50 series cards, or a GGUF version on 8 GB GPUs.

Can I use qwen_2.5_vl_7b for other models?

Some models built on Qwen-Image reuse it, and the official workflow will list it if so. Models with their own Qwen text encoders, such as Qwen-Image 2.1 or Krea 2, need the file named in their template.

Final Thoughts

Qwen-Image looks complicated because there are so many versions, but the file list comes down to two families. For Qwen-Image, 2512 and every Edit version, download qwen_2.5_vl_7b_fp8_scaled.safetensors and qwen_image_vae.safetensors once, then add whichever diffusion model you want. For Qwen-Image 2.1, download its own Qwen3-VL encoder and 2.1 VAE. Put each file in the folder shown above, open the official template, and the red nodes disappear.

Next: learn editing with the Qwen-Image 2.1 edit guide, compare it with other open models in my best open source text-to-image models roundup, and keep the model folders cheat sheet handy.

Sources

• Hugging Face: Comfy-Org/Qwen-Image_ComfyUI (file list and sizes)
• Hugging Face: Comfy-Org/Qwen-Image-Edit_ComfyUI
• Hugging Face: Comfy-Org/Qwen-Image-2.1
• ComfyUI docs: Qwen-Image native workflow
• Hugging Face: QuantStack/Qwen-Image-GGUF
• GitHub: city96/ComfyUI-GGUF
• BetterToken: Qwen Image 2.1 file sizes