Video coming soon
ComfyUI Image Size Explained: Empty Latent Image, Aspect Ratio and Batch Size
What the empty latent is, width and height, why the same seed gives a new image at a new size, why bigger is not always better, why 1080 becomes 1088, aspect ratio, batch size, which empty latent node to use, and Resolution Pixaroma, with two deeper looks: what a latent really is, and where the steps of 16 come from. Made with Z-Image Turbo.
To open one: drag the .json file onto ComfyUI, or use Workflow > Open. The Pixaroma version needs the free Pixaroma nodes. New to ComfyUI? Install it first.
Practice tasks
The same tasks as the notes in the Practice workflow.
- 1
Connect
The blank canvas
The EmptySD3LatentImage node is not connected, so the KSampler has no canvas.
Drag its LATENT output to the latent_image input of the KSampler.You know it worked when you press Run and get the panda, with no red error.
- 2
Experiment
A tall image
Set width to 832 and height to 1216, and press Run.
You know it worked when you get a tall image, and a new panda: the same seed gives a new image at a new size.
- 3
Experiment
Steps of 16
Click the height value, type 1080, and press Enter.
You know it worked when it changes to 1088: this node only takes sizes in steps of 16.
- 4
Experiment
A batch
Set batch_size to 4, and press Run.
You know it worked when Save Image shows four different pandas from one Run. Run it again: the same seed gives the same four.
- 5
Save
Keep your work
Press Ctrl + S, type a name like My image size practice, and press Enter.
You know it worked when your workflow shows in the Workflows panel on the left.
Models
Download each file and put it in its folder.
| Model file | Size | Put it in | Download |
|---|---|---|---|
| Diffusion model pick onez_image_turbo_int8_convrot.safetensorsint8, the one used in the video | 6.2 GB | ComfyUI/models/diffusion_models/z-image | Hugging Face |
| Diffusion model pick onez_image_turbo_bf16.safetensorsbf16, full detail, needs more video memory | 12.3 GB | ComfyUI/models/diffusion_models/z-image | Hugging Face |
| Diffusion model pick onez_image_turbo_nvfp4.safetensorsnvfp4, for RTX 50 cards | 4.5 GB | ComfyUI/models/diffusion_models/z-image | Hugging Face |
| Text encoder pick oneqwen_3_4b_fp8_mixed.safetensorsfp8, the one used in the video | 5.6 GB | ComfyUI/models/text_encoders | Hugging Face |
| Text encoder pick oneqwen_3_4b.safetensorsfull detail | 8.0 GB | ComfyUI/models/text_encoders | Hugging Face |
| VAEae.safetensors | 0.3 GB | ComfyUI/models/vae | Hugging Face |
| Checkpoint (optional, Task 4)z-image-turbo-fp8-aio.safetensorsall three parts in one file | 10.3 GB | ComfyUI/models/checkpoints | Hugging Face |
After the download, press R in ComfyUI (or restart it), so the new files show up in the loaders.
Z-Image Turbo is free (Apache 2.0). The Comfy-Org files download without an account.
Chapters
Click a time to jump there on YouTube.