tutorials/image/qwen/qwen-image

[giscus[bot]]

tutorials/image/qwen/qwen-image

Qwen-Image is a 20B parameter MMDiT (Multimodal Diffusion Transformer) model open-sourced under the Apache 2.0 license.

https://docs.comfy.org/tutorials/image/qwen/qwen-image

[yuesam]

Will there be an i2i workflow soon? Thanks

[comfyui-wiki]

You can just try using the Load Image node and VAE Encoder node, just like the SD1.5 i2i workflow.

Not tested yet, but the Qwen team said the image editing model is being trained currently and will be released soon.

[matze2001]

First: Thx for the fast integration! :)

After about halfway through the steps, the preview image in the sampler turns completely black, and the output is also black. I'm using a RTX 5090 and the fp8 scaled version. Other models like the Flux Krea continue to work even after the ComfyUI update. Any advice?

[joeyshelton]

How exactly did you do this? I've seen issues on other threads saying the same but I'm running on Ubuntu, not windows so I have no .bat files for startup. I looked in main.py which seems to be the startup entrypoint of comfy launch and there is no mention of sage attention.

[joeyshelton]

Okay I resolved my issue but it may not be related to what other users reported here. I will share what resolved it just in case someone sees a similar issue:
I noticed that a blank image would be generated any time my input prompt had parentheses.

I can confirm with the latest version of ComfyUI and this workflow, If I input the Positive Prompt:
A vibrant, warm neon-lit street scene in Hong Kong at the afternoon, with a mix of colorful Chinese and English signs glowing brightly
Generation works fine. If I have any parentheses in the prompt, i.e.:
A vibrant, warm neon-lit street scene in Hong Kong at the afternoon, with a mix of colorful Chinese (and English) signs glowing brightly

This results in an all black image.

[GreyHound55]

From what I've read ... SageAttention is causing issues with Black outputs .... Supposedly disabling that, corrects the issue

[matze2001]

After several attempts, it only worked when I started ComfyUI without sageattention:

qwen doesn't work:
--windows-standalone-build --use-sage-attention

works:
--windows-standalone-build --fast fp16_accumulation

[GreyHound55]

Yes 👍

[offmybach]

I have used exact workflow and models that you have listed and keep getting error: Prompt executed in 0.06 seconds
got prompt
0 models unloaded.
loaded partially 3986.666702365875 3986.666702270508 0
!!! Exception during processing !!! Input tensors must be in dtype of torch.float16 or torch.bfloat16
Traceback (most recent call last):
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 496, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 315, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 289, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\execution.py", line 277, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\nodes.py", line 74, in encode
return (clip.encode_from_tokens_scheduled(tokens), )
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd.py", line 170, in encode_from_tokens_scheduled
pooled_dict = self.encode_from_tokens(tokens, return_pooled=return_pooled, return_dict=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd.py", line 232, in encode_from_tokens
o = self.cond_stage_model.encode_token_weights(tokens)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\text_encoders\qwen_image.py", line 37, in encode_token_weights
out, pooled, extra = super().encode_token_weights(token_weight_pairs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd1_clip.py", line 683, in encode_token_weights
out = getattr(self, self.clip).encode_token_weights(token_weight_pairs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd1_clip.py", line 45, in encode_token_weights
o = self.encode(to_encode)
^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd1_clip.py", line 288, in encode
return self(tokens)
^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\sd1_clip.py", line 261, in forward
outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\text_encoders\llama.py", line 347, in forward
return self.model(input_ids, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\text_encoders\llama.py", line 318, in forward
x = layer(
^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\text_encoders\llama.py", line 199, in forward
x = self.self_attn(
^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\text_encoders\llama.py", line 163, in forward
output = optimized_attention(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\ComfyUI\comfy\ldm\modules\attention.py", line 451, in attention_pytorch
out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\sageattention\core.py", line 130, in sageattn
return sageattn_qk_int8_pv_fp16_triton(q, k, v, tensor_layout=tensor_layout, is_causal=is_causal, sm_scale=sm_scale, return_lse=return_lse)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch_dynamo\eval_frame.py", line 838, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "K:\ComfyUI_windows_portable_nvidia\ComfyUI_windows_portable\python_embeded\Lib\site-packages\sageattention\core.py", line 219, in sageattn_qk_int8_pv_fp16_triton
assert dtype in [torch.float16, torch.bfloat16], "Input tensors must be in dtype of torch.float16 or torch.bfloat16"
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: Input tensors must be in dtype of torch.float16 or torch.bfloat16

Prompt executed in 0.15 seconds

[GreyHound55]

Replaced the Qwen model of fp8_e4m3fn with the fp16 and definitely a much better output ! (but takes 4-6 times the processing time) (running on a 4090)

Does anyone know if there's a fp16 text encoder ?

Currently using, for the text encoder, the qwen_2.5_vl_7b_fp8_scaled

[AlexAI2025]

will this work with Mac M4?

[olafrv]

For those with low VRAM (8GB)...

This tutorial works perfectly (Minute 25):

• https://www.youtube.com/watch?v=gO6_pDYHPlw

Use city96 K_Q2 GGUF models and example workflow:

• https://huggingface.co/city96/Qwen-Image-gguf

Relevant file locations:

• "C:\ComfyUI\models\diffusion_models\qwen-image-Q2_K.gguf"
• "C:\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q2_K.gguf"
• "C:\ComfyUI\models\vae\qwen_image_vae.safetensors"

[vijayanmmm]

using comfyUI verison: ComfyUI v0.3.47
ComfyUI_frontend v1.23.4
Discord
ComfyOrg
ComfyUI_desktop v0.4.59
ComfyUI-Manager V3.30.4

i downloaded and used the sample workflow json file, in the Load CLIP node its mentioned qwen_image which is missing in my computer & showing error, how to add this? thanks in advance.

[hltchj]

same problem. Have you sloved it?

Prompt outputs failed validation:
CLIPLoader:

• Value not in list: type: 'qwen_image' not in ['stable_diffusion', 'stable_cascade', 'sd3', 'stable_audio', 'mochi', 'ltxv', 'pixart', 'cosmos', 'lumina2', 'wan', 'hidream', 'chroma', 'ace', 'omnigen2']

[organisedtoast]

Same problem!

[comfyui-wiki]

Did you update ComfyUI?

[yongshaochen]

all i use is the Qwen-Image_fp8 (20.4 GB) ,and the offical workflow,my computer is 4090,24G vram,120G ram,i9-14900k,i ran the defual workflow take me more than 20min everytime,why?

[Callisin]

Will this work with M2 Mac 32 gb ???

[GreyHound55]

Is SageAttn on or off ?

[Adreitz]

@GreyHound55 Not sure who you're responding to. I've never used SageAttention. My impression is that it is for speeding up inference at the cost of lower precision. Not really interested.

[GreyHound55]

Ok, yeah, it definitely speeds up Inference ... but is a problem with Qwen, producing black images

[Adreitz]

@GreyHound55 Okay, not the source of my issue, but thanks! I think it's probably a Mac-specific issue.

[GreyHound55]

Ok ... Good luck 👍

[webasylum]

Does QWEN need or benefit from a Refiner, such as the xd_sl_refiner in xdsl?

[Dave6602]

Thank you! Using Q4_K_S quant (not distilled) with sageattention, works well on 8 GB 4060 laptop. It's long generation time, but results are good. 1328x1328, 19 s per step

[LeoLisena]

Hi, I’m testing Qwen-Image in ComfyUI for img2img and noticed a consistent issue:

Using the official ComfyUI text2image workflow template → works great from pure noise latent: sharp, detailed, no artifacts.
When I modify it to use an input image (VAE-encoded latent) → results are often washed out / watercolor-like and sometimes have visual artifacts (patchy areas, broken structure).

This behavior does not occur with Flux under the same conditions.
Is this a known limitation, or is there a recommended workflow for Qwen-Image img2img starting from encoded latents?

[RunebinderOG]

The Inpaint model doesn't seem to be working for me. Downloaded the workflow and made it identical to the example workflow screen shot, used the input image on the page, masked the hair and prompted "red hair", but the output is identical to the input with the white hair.

Canny works fine and checked the console and there's no errors in there.

[comfyui-wiki]

@RunebinderOG Just change the prompt

@mald55 Yes, if you can't find the model_patches folder, you need to create one by yourself. Don't forget to update ComfyUI first.

[ToniVacaESAT]

Same problem. I've tried the "make your hair deep red" prompt and increased the ControlNet strength to 2, but it didn't work at all. All the files are in place.

[RunebinderOG]

@RunebinderOG Just change the prompt

@mald55 Yes, if you can't find the model_patches folder, you need to create one by yourself. Don't forget to update ComfyUI first.

The prompt is already asking the model to change the hair from white to red and should work, makes no sense to change the prompt. Tried it though and as expected the output is identical to the input

.

[gitmeAnV]

You need to use Mask output while inpainting, but you have not.

[RunebinderOG]

You need to use Mask output while inpainting, but you have not.

Sometimes need someone to point out the obvious mistakes, thank you :)

[RamonGuthrie]

It seems that all Qwen-Image Controlnet and Control-Loras produce faint Cross-Hatch Patterns on the image. Are there plans to fix this issue?

[lmkyl]

Pause each time after clicking 'run', no matter intalled comfyui or portable edition (updated to newest by running the bat file 'update_comfyui_and_python_dependencies' )

New installed Windows 10 22H2 Aug with 32GB RAM
Newest NVIDIA studio driver installed on 4070Super with 12GB VRAM
All safetensors' sha256 match

portable comfyui shows:

To see the GUI go to: http://127.0.0.1:8188
got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Using scaled fp8: fp8 matrix mult: False, scale input: False
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load QwenImageTEModel_
loaded completely 9633.8 7909.737449645996 True

E:\AI\ComfyUI_windows_portable>pause
请按任意键继续. . .

[animalcolm]

I'm using the image union controlnet and am I dumb or is there no way o control the strength / in / out of the controlnet? am i missing something?

[davoodtaba-glitch]

Hey everyone! Hope you're doing well.

I've been experimenting with Qwen 2.1 and I'm having a bit of trouble with image quality. Whenever I use this prompt, the results tend to come out quite noisy, muddy, and a little overprocessed. I'm trying to achieve a cleaner, more detailed, photorealistic look.

Does anyone have any idea what might be causing this, or which settings/prompt changes could help improve the result?

I'd really appreciate any tips or suggestions. Thanks in advance!

Prompt: A colossal, otherworldly biomechanical sea creature rising dramatically from rain-soaked, muddy ground covered with delicate wildflowers during an intense thunderstorm. Its exterior is a textured dark teal-green metallic shell, reinforced with intricate rivets, layered armor plates, and design elements inspired by an ancient deep-sea vessel. Multiple vivid neon-green luminous sensors and circular glowing apertures are integrated into its head and limbs, casting an eerie emerald illumination across the surrounding wet ground.The creature's enormous open mouth reveals rows of rugged ivory-colored teeth as it unleashes a powerful roar. One gigantic armored limb extends toward the foreground, its heavy articulated claws pressing into the soft earth and scattering pink and yellow flowers. Torrential rain streams through flashes of lightning, reflecting across the creature's wet metallic surface.Mud droplets, flower petals, mist, and environmental debris fill the air around its movement. In the distance, a dark mountainous landscape appears as jagged silhouettes beneath a turbulent storm-filled sky. Low-angle wide composition, immense sense of scale, cinematic perspective, hyper-realistic materials and environmental textures, dramatic volumetric lighting, atmospheric depth, epic dark-fantasy creature design, photorealistic detail, high contrast, dynamic storm atmosphere.

[lin-bot23]

Hi @davoodtaba-glitch, thanks for sharing the prompt and the sample images. In most cases like this the prompt is not the culprit: it is the model/settings pairing. A few things worth checking:

• Which Qwen model and workflow are you running? This page documents the original Qwen-Image (20B) and its template image_qwen_image.json. If you are on Qwen-Image 2.1, it has its own page and workflow with its own defaults: https://docs.comfy.org/tutorials/image/qwen/qwen-image-2-1
• Match the sampler settings to the model variant. The page lists three options and they are not interchangeable: the base qwen_image_fp8_e4m3fn runs at the template's default step/CFG, the distilled checkpoint wants about 15 steps at CFG 1.0 (10 steps also works), and the 8-step lightx2v LoRA is a separate acceleration path. Running a distilled or accelerated model at a high CFG, or combining the distilled checkpoint with the 8-step LoRA, is a common cause of muddy, over-processed output.
• Confirm the model files. Load the files listed on the page (qwen_image_fp8_e4m3fn.safetensors, qwen_2.5_vl_7b_fp8_scaled.safetensors, qwen_image_vae.safetensors) without mixing quantizations; a truncated or mismatched safetensors also shows up as degraded detail.
• Prompt length. A very long, heavily decorated prompt can dilute the composition. Try trimming to subject + scene + style in a few sentences and see whether the structure sharpens.

If you share the exact diffusion model file, your KSampler settings (steps / CFG / sampler / scheduler) and which workflow JSON you loaded, I can narrow it down further.

On the refiner question from @webasylum: the official Qwen-Image workflow is single-pass and has no separate refiner/upscale stage. A refiner-style LoRA is a third-party add-on rather than part of the ComfyUI template, so you would only need one if you are chasing a specific look.