How to run Qwen-Image locally.

Which Qwen-Image.

Four things go by the name. Three of them share one text encoder and one VAE.

• Qwen-Image, from 4 August 2025, is the original: a 20B model under Apache 2.0.
• Qwen-Image 2512, from 31 December 2025, is the “December update” with better human realism and text. Same size, same encoder, same VAE. Users say it lost some of the flat anime look of the August model.
• Lightning LoRAs from lightx2v cut either one to 8 or 4 steps at CFG 1. There are separate LoRAs for the original and for 2512.
• Qwen-Image 2.1, from 20 September 2026, is a new model: 7.1B, with a Qwen3-VL 8B encoder, its own VAE with transparency, and text to image plus editing from up to 10 reference images in one file. Native size is 2048 × 2048.

The editing sibling, Qwen-Image-Edit (2509 and 2511), is a separate model. It gets a short section under the files.

Files you need.

One model, one text encoder, one VAE. Qwen-Image and 2512 take Qwen2.5-VL 7B. Qwen-Image 2.1 takes Qwen3-VL 8B. They don’t swap.

Qwen-Image and 2512

• Model qwen_image_fp8_e4m3fn.safetensors ComfyUI/models/diffusion_models/ 2512: qwen_image_2512_fp8_e4m3fn.safetensors, same size. Full precision: 40.9 GB
  20.4 GB Download
• Text encoder qwen_2.5_vl_7b_fp8_scaled.safetensors ComfyUI/models/text_encoders/ full precision: qwen_2.5_vl_7b.safetensors, 16.6 GB
  9.4 GB Download
• VAE qwen_image_vae.safetensors ComfyUI/models/vae/
  0.3 GB Download

Lightning LoRAs

Use the LoRA made for your model, and match the step count to the file: lightx2v says 4-step LoRAs technically run at 8 steps, but the 8-step LoRA at 8 and the 4-step LoRA at 4 is the good choice.

• Qwen-Image, 8 steps Qwen-Image-Lightning-8steps-V2.0-bf16.safetensors ComfyUI/models/loras/
  0.9 GB Download
• Qwen-Image, 4 steps Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors ComfyUI/models/loras/
  0.9 GB Download
• 2512, 8 steps Qwen-Image-2512-Lightning-8steps-V1.0-bf16.safetensors ComfyUI/models/loras/
  0.9 GB Download
• 2512, 4 steps Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors ComfyUI/models/loras/
  0.9 GB Download

Each also comes as a 1.7 GB fp32 file. lightx2v also has merged 2512 checkpoints with the LoRA built in, 20.4 GB each.

Qwen-Image 2.1

• Model qwen_image_2.1_int8_convrot.safetensors ComfyUI/models/diffusion_models/ full precision: qwen_image_2.1_bf16.safetensors, 14.2 GB
  7.3 GB Download
• Text encoder qwen3vl_8b_int8_convrot.safetensors ComfyUI/models/text_encoders/ full precision: qwen3vl_8b_bf16.safetensors, 17.5 GB
  9.4 GB Download
• VAE qwen_image_2.1_vae_bf16.safetensors ComfyUI/models/vae/
  0.7 GB Download

The same repo has two qwen3.5_9b_qwen_image_2.1_pe files. They are prompt enhancers, language models for ComfyUI’s TextGenerate node, not text encoders. Loading one as the encoder gives garbage or a shape error. The int8 files run fast only on PyTorch built for CUDA 13.0 (cu130); on cu128 they were slower than a GGUF in Kijai’s test.

models/
├── diffusion_models/
│   ├── qwen_image_fp8_e4m3fn.safetensors
│   └── qwen_image_2.1_int8_convrot.safetensors
├── loras/
│   └── Qwen-Image-Lightning-8steps-V2.0-bf16.safetensors
├── text_encoders/
│   ├── qwen_2.5_vl_7b_fp8_scaled.safetensors
│   └── qwen3vl_8b_int8_convrot.safetensors
└── vae/
    ├── qwen_image_vae.safetensors
    └── qwen_image_2.1_vae_bf16.safetensors

Smaller files: GGUF

Qwen-Image and 2512 GGUFs load through the ComfyUI-GGUF nodes and go in diffusion_models. For the encoder, unsloth’s Qwen2.5-VL 7B GGUF is 4.7 GB at Q4_K_M and 8.1 GB at Q8. unsloth also has Qwen-Image 2.1 GGUFs (4.2 GB at Q4_K_M), but ComfyUI-GGUF’s main branch doesn’t load them yet.

city96 goes lower too: Q3_K_M is 9.7 GB and Q2_K 7.1 GB.

Qwen-Image-Edit

Edit is its own model for changing an existing picture. It uses the same Qwen2.5-VL 7B encoder and Qwen-Image VAE as above, plus its own file from Comfy-Org’s Edit repo: qwen_image_edit_2511_fp8mixed.safetensors is the newest, 20.5 GB. In ComfyUI the picture goes in through the TextEncodeQwenImageEditPlus node, which takes the prompt and up to three images. If you use a GGUF encoder for Edit, it needs its matching mmproj file next to it, or the image input breaks. The ComfyUI Edit tutorial has the full walkthrough. Qwen-Image 2.1 can also edit, from reference images, with the files above.

What fits your computer.

At about 1 megapixel. The text encoder runs first and ComfyUI moves it out of the way before sampling, so the model file sets the limit. A model that doesn’t fit still runs with part of it in system RAM, only slower.

• 6 GBNo

  Even Qwen-Image Q2_K is 7.1 GB. Z-Image Turbo or Flux.2 Klein 4B are the better fit.

• 8 GBOffloads

  Only with most of the model in system RAM. Qwen-Image 2.1 int8 (7.3 GB) comes closest. Plan on 32 GB of RAM or more.

• 12 GBOffloads

  Qwen-Image as GGUF Q4_0 (11.9 GB) or Q3_K_M (9.7 GB), partly in RAM. Qwen-Image 2.1 int8 fits better.

• 16 GBFits

  Qwen-Image 2.1 int8 with the int8 encoder: an RTX 4060 Ti 16 GB peaked around 15 GB with no offloading. Qwen-Image as Q4_K_M or Q5_K_M GGUF, or fp8 with offloading and the Lightning LoRA.

• 24 GBFits

  Qwen-Image or 2512 in fp8. Comfy’s own timings are from a 24 GB RTX 4090D at 86% of its memory.

• Mac 24 GBSwaps

  Qwen-Image 2.1 swaps hard: one edit with three references at 1280 × 736 took 35 minutes on an M4 Pro with 24 GB.

• Mac 32 to 48 GBTight

  Qwen-Image as GGUF (13.1 to 21.8 GB) with a GGUF encoder, or Qwen-Image 2.1 in bf16, about 32 GB for the set. We found no measured Mac times to quote.

• Mac 64 GB+Fits

  Qwen-Image 2.1 in bf16 with room to spare, or Qwen-Image as Q8_0 GGUF.

• 16 GB graphics cards
• RTX 5060 Ti 16 GB
• 24 GB cards
• RTX 5090
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   Qwen-Image needs a ComfyUI from August 2025 or later; older ones say qwen_image is not in the list of CLIP types. Qwen-Image 2.1 needs ComfyUI 0.37.0 or the nightly build, for its TextEncodeQwenImage21 node. ComfyUI Desktop and the portable build can lag behind. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the files

   Model, text encoder and VAE from the lists above, plus a Lightning LoRA if you want 8 or 4 steps.

3. 3

   Put them in their folders

   Model in diffusion_models, encoder in text_encoders, VAE in vae, LoRA in loras. Restart ComfyUI so the loaders list them.

4. 4

   Open a Qwen template

   In the template browser, pick Qwen-Image: Text to Image, Qwen Image 2512 or the Qwen-Image 2.1 text to image template. The first two ship with a Lightning LoRA already wired in.

5. 5

   Check the loaders

   Load Diffusion Model gets the model. Load CLIP gets the encoder with type qwen_image. Load VAE gets the matching VAE. For Qwen-Image and 2512 there is a ModelSamplingAuraFlow node at shift 3.1; for 2.1 the prompt goes through TextEncodeQwenImage21 instead of a plain text encode.

6. 6

   Write a prompt and run

   Plain, detailed sentences work. The first run loads everything and takes longest.

Settings that work.

Qwen-Image and 2512

Steps

50

CFG

4

Sampler

euler

Scheduler

simple

Size

1328 × 1328

Shift

AuraFlow 3.1

Negative

yes

Latent

EmptySD3LatentImage

That is the model card: 50 steps at a true CFG of 4. Comfy’s first Qwen-Image template ran 20 steps at CFG 2.5, which is much faster and still good; its note says to try 50 for the original settings. The 2512 template uses 50 steps at CFG 4.

With a Lightning LoRA

Steps

8 or 4

CFG

1

Sampler

euler

Scheduler

simple

Steps match the LoRA’s name. Comfy’s current Qwen-Image template runs the 8-step LoRA at 8 steps, CFG 1, shift 3.1. Don’t combine Lightning with the community qwen_image_distill_full model: that one is already distilled, and runs at 15 steps and CFG 1 on its own.

The model card’s sizes. Comfy’s 2512 note uses 1472 × 1104 for 4:3, a multiple of 16. Users report anything from 768 to 2048 on a side works well.

Qwen-Image 2.1

Steps

25 to 40

CFG

1

Sampler

euler

Scheduler

simple

Size

1024 to 2048

Multiple of

32

Text encode

TextEncodeQwenImage21

Latent

EmptyLatentImage

The model card uses 40 steps at 2048 × 2048. Comfy’s template uses 25 steps at 1024 × 1024, which is quicker; users who saw faint banding fixed it with 40 steps at 2K. Keep CFG at 1 unless you add a negative prompt, then raise it. For a transparent PNG, wrap the prompt the way the model card does: start with “This is an RGBA image with transparency.” and end with “The image has alpha channel and the background is transparent.”

How fast.

Times after the first run, which loads the models (the 4090D took 94 s and 55 s on its first runs).

On an RTX 20-series card it is a different story: those GPUs have no bf16, fp16 overflows to black, so Qwen-Image runs in fp32 and one image took more than 10 minutes on a 2080 Ti.

When it goes wrong.

CLIPLoader ... Value not in list: type: 'qwen_image' not in [...]

ComfyUI is too old for Qwen-Image. Update it.

The image turns black partway through sampling

Start ComfyUI without --use-sage-attention, and don’t set the weight dtype to fp8_e4m3fn_fast. The same fix applies to black Lightning images.

Error while deserializing header: header too large

A broken download, usually the VAE. Download it again.

lora key not loaded, or blurry low-contrast Lightning images

Your ComfyUI is older than the LoRA support. Update it.

An error with a Qwen2.5-VL file from elsewhere

Wrong text encoder. Use Comfy-Org’s qwen_2.5_vl_7b_fp8_scaled or the full-precision one from the same repo.

TextEncodeQwenImage21 or QwenImage21Cache missing

Qwen-Image 2.1 needs ComfyUI 0.37.0 or the nightly build.

Given normalized_shape=[4096] ... got input of size [1, 338, 5120]

Qwen-Image 2.1 got the wrong encoder, often a qwen3.5_9b prompt-enhancer file. Use qwen3vl_8b_int8_convrot or qwen3vl_8b_bf16.

Qwen-Image 2.1 looks yellow, or a fine diamond grid shows on skin

For the tint, raise CFG and add a negative prompt. The grid comes from the 2.1 VAE; madebyollin’s texture-fix VAE is a drop-in replacement.

For the first, users VAE-encode a blank image and sample it at a denoise of about 0.95 instead of starting from an empty latent. The second has no official answer yet.

Questions.

Sources: Qwen-Image model card, Qwen-Image 2512 model card, Qwen-Image 2.1 model card, ComfyUI Qwen-Image tutorial, ComfyUI Qwen-Image 2.1 tutorial, Comfy Qwen-Image template [1], 2512 speed thread [2], Qwen-Image 2.1 on a 4060 Ti [3], Qwen-Image-Lightning, ComfyUI issue #10852, ComfyUI issue #16470, 2.1 on an M4 Pro.