Qwen-Image 2.1 is an open image model. It has about 393,000 neurons in its image layers. We found single neurons that change every image in the same way. Each neuron acts like a dial. Turn it up, and every scene becomes night, fluffy, or a red-chalk drawing. The scene itself stays the same.

The dials

Choose a dial. Move the slider to change the strength. Zero is the original image. Each dial passed a test on 4 prompts and 6 random seeds.

Hover over an image to see the original. Click an image to compare the two. Use the arrow keys to change the strength.

Neurons that did not work

We tested 34 neurons. 19 did not pass. Some failures looked good at first. These three show the most common traps.

How we found them

1. We pushed 2,720 random neurons one at a time, on 3 short prompts. We measured how much each push changed the image.
2. We tested each neuron at half, normal and double strength. Some neurons work only at one strength.
3. We looked at the strongest candidates by eye, on 4 prompts.
4. We rendered the good candidates on 8 short prompts and 8 detailed prompts at 1024 pixels.
5. We gave each neuron a seed test: 4 prompts, 6 random seeds, 24 images. A neuron passes only if it changes all 24 images in the same way and each image still shows its prompt.

Limits

We tested about 0.7% of the neurons. Eleven dials work on detailed prompts. Four dials work only on short prompts, because a detailed prompt fixes too much of the image. The names describe what we saw. They are not measurements. We tested each dial at one strength with 6 seeds. At high strength, most dials damage the image. The early layers decide what is in the picture. Single neurons there usually replace the prompt. The later layers decide how things look. Single neurons there act as dials.