Bad Image Quality generated by Qwen Image

[vagitablebirdcode]

Describe the bug

I installed the latest version of the diffusers library from the main branch, and used the inference code example in the application for inference, but the effect was very poor, as shown in the following figure:

However, the effects in the official application are as shown in the figure.

Reproduction

Firstly install the newest version of diffusers:
pip install git+https://github.com/huggingface/diffusers

Codes are as follows:

import torch.nn.functional as F

_orig_sdpa = F.scaled_dot_product_attention

def _sdpa_drop_enable_gqa(*args, **kwargs):
    kwargs.pop("enable_gqa", None)
    return _orig_sdpa(*args, **kwargs)

F.scaled_dot_product_attention = _sdpa_drop_enable_gqa


from diffusers import DiffusionPipeline
import torch

model_name = "Qwen/Qwen-Image"
torch_dtype = torch.bfloat16
device = "cuda"
pipe = DiffusionPipeline.from_pretrained(model_name, torch_dtype=torch_dtype)
pipe = pipe.to(device)
prompt = "一位身着淡雅水粉色交领襦裙的年轻女子背对镜头而坐，俯身专注地手持毛笔在素白宣纸上书写“通義千問”四个遒劲汉字。古色古香的室内陈设典雅考究，案头错落摆放着青瓷茶盏与鎏金香炉，一缕熏香轻盈升腾；柔和光线洒落肩头，勾勒出她衣裙的柔美质感与专注神情，仿佛凝固了一段宁静温润的旧时光。"

image = pipe(
    prompt=prompt,
    width=1024,
    height=1024,
    num_inference_steps=30,
    true_cfg_scale=4.0,
    guidance_scale=1.0
    generator=torch.Generator(device="cuda").manual_seed(42)
).images[0]
image.save("res.jpg")

Logs

System Info

- 🤗 Diffusers version: 0.35.0.dev0
- Platform: Linux-5.15.0-58-generic-x86_64-with-glibc2.35
- Running on Google Colab?: No
- Python version: 3.10.10
- PyTorch version (GPU?): 2.6.0 (True)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Huggingface_hub version: 0.34.4
- Transformers version: 4.55.2
- Accelerate version: 1.10.0
- PEFT version: not installed
- Bitsandbytes version: not installed
- Safetensors version: 0.5.3
- xFormers version: 0.0.22
- Accelerator: NA
- Using GPU in script?: <fill in>
- Using distributed or parallel set-up in script?: <fill in>

Who can help?

No response

[a-r-r-o-w]

If you use 50 steps instead of 30, do you get better results? I think lower steps results in worse quality with Qwen but I haven't extensively tested.

You can try using #12109 with UniPC for 30 step generation

cc @asomoza

[vagitablebirdcode]

If you use 50 steps instead of 30, do you get better results? I think lower steps results in worse quality with Qwen but I haven't extensively tested.

You can try using #12109 with UniPC for 30 step generation

cc @asomoza

Thank you for your reply. First, I indeed tried to take a larger step, but it still proved ineffective. The above implementation is based on:
https://huggingface.co/spaces/Qwen/Qwen-Image/blob/main/app.py
The comparison image used the same parameters, but the quality difference was very large.

I',m now try this: #12109.

[asomoza]

I think you have something wrong in your environment, just a quick test with quantization gives me a better image than yours.

Also can you try without the first part?

This is what I get, if you don't get this something similar, you have something in your env that is making the image worse.

[vagitablebirdcode]

Understood, thank you for your test and reminder, I will investigate the issues in the internal environment.

[spejamas]

@vagitablebirdcode Did you ever figure this out? I experience a similar quality problem, even with a fresh install in a very clean environment.

@asomoza Am I missing some important dependency? This is the most basic possible implementation I think.....

h100 pcie, 80gb vram
cuda 12.8.1
python 3.11
torch 2.8

python -m venv venv
source venv/bin/activate

pip install torch
pip install git+https://github.com/huggingface/diffusers
pip install transformers

python run.py

run.py:

import torch
from diffusers import QwenImagePipeline

pipe = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16)
pipe = pipe.to("cuda")

prompt = "A super cute, kawaii cat, half black and half orange. The cat is holding a martini glass. The background is white."

image = pipe(
    prompt=prompt,
    width=1024,
    height=1024,
    num_inference_steps=40,
    true_cfg_scale=4.5,
    generator=torch.Generator(device="cuda").manual_seed(1030)
).images[0]

image.save("flaming_garbage.png")

The resulting image is severely distorted--other outputs look more coherent, but still have some obvious quality issue compared to qwen-image's typical outputs.

[asomoza]

@spejamas are you using a clean environment, meaning you are creating and just installing those libraries? I can't think of any other issue that's happening here, I can't do anything myself to get those kind of bad images. I don't have access to an H100 though but I doubt that the hardware is the issue here.

This is what I get with your same exact prompt and settings:

It kind of make sense, since you're not specifying if you want a photo, cartoon, render, etc.

[spejamas]

@spejamas are you using a clean environment, meaning you are creating and just installing those libraries? I can't think of any other issue that's happening here, I can't do anything myself to get those kind of bad images. I don't have access to an H100 though but I doubt that the hardware is the issue here.

Yes, I am not installing anything else at all in the venv besides the above. I am as puzzled as you are. It's on a runpod gpu, which hasn't given me any trouble before, but who knows. I'm with you that it shouldn't be a hardware thing. Maybe it's a cuda thing on the runpod machine.

This is what I get with your same exact prompt and settings:

Thanks for the sanity check. This is definitely what it should look like. I will keep trying.

[spejamas]

Okay so it turns out, you must include a negative prompt, even an empty one.
@vagitablebirdcode

prompt = "A super cute, kawaii cat, half black and half orange. The cat is holding a martini glass. The background is white."

image = pipe(
    prompt=prompt,
    negative_prompt="",
    width=1024,
    height=1024,
    num_inference_steps=40,
    true_cfg_scale=4.5,
    generator=torch.Generator(device="cuda").manual_seed(1030)
).images[0]

prompt = "A super cute, kawaii cat, half black and half orange. The cat is holding a martini glass. The background is white."

image = pipe(
    prompt=prompt,
    width=1024,
    height=1024,
    num_inference_steps=40,
    true_cfg_scale=4.5,
    generator=torch.Generator(device="cuda").manual_seed(1030)
).images[0]

How silly of me.

Probably not the whole story still though.

EDIT: actually I think this is the whole story. After some more thorough sampling tests, simply passing an empty string for the negative prompt (instead of excluding that parameter) has fixed the quality problem I observed completely. I guess this is a noob error haha

[vagitablebirdcode]

@spejamas thanks for your answer, this also solve my problems! @asomoza I want to know what GPU device you are using, it seems that some GPU devices may cause image quality issues when not configured with a negative_prompt. But in other models, such as Flux, whether or not the parameter is used does not affect the quality of the results. Therefore, is it possible that the Qwen Image pipeline is lacking some checks or processing for the parameter?