-
Notifications
You must be signed in to change notification settings - Fork 5
Choosing a Model
The most common reason someone installs OpenLayer and then stops is that the model question looks like a 177 GB question. It isn't. You need one stack, and the smallest one is the best one to start with.
The README's Required files section has the exact filenames, folders and sizes. This page is the decision behind that list.
Three files. Apache-2.0, no licence page, no account. Every node it needs is core ComfyUI, so there is nothing extra to install in ComfyUI-Manager.
It gives you Text to Image, Image to Image, instruction editing, and Multi-Reference. Add
4x-UltraSharp.pth — 67 MB — and Upscale works too.
It is also the fastest thing in the panel: 1024×1024 in about 12 seconds on a 4070 Ti, at 4 steps with CFG 1. If a preset ever feels slow, it is almost certainly not this one.
| You want | Add | Cost on top |
|---|---|---|
| Sketch to Image | Z_image_Turbo stack + a ControlNet patch | ~14.7 GB |
| Better inpainting | Nothing — Klein already inpaints | 0 GB |
| Outpaint | The Flux Fill stack | 34.2 GB |
| Splitting a photo into layers | The Unflatten stack | ~30 GB |
| Live Painting | Any SD 1.5 checkpoint + an LCM LoRA | varies |
The sharing is the point. qwen_3_4b.safetensors (8 GB) is used by nine presets across four
tools — Klein and Z_image_Turbo both read it. So if you have Klein and then add Z_image_Turbo, it
costs 12.3 GB, not 20.7 GB. ae.safetensors is shared between Z_image_Turbo and Flux Fill the same
way.
That is why the panel's Setup screen is worth opening before you download anything: it counts each file once, checks what you already have, and tells you the real remaining total for your machine.
- Multi-Reference does not preserve a face. Clothing, props, setting and lighting carry across from a reference. A specific person comes back as a plausible stranger — confirmed against real photographs and dozens of generated ones. Compose scenes with it; do not try to cast someone in one.
- Unflatten needs a subject standing in front of a background. A close-up that fills the frame has no front and back to find, and comes back unchanged. This is about composition, not about whether the picture was photographed or generated.
Four files are non-commercial: flux1-dev-fp8, flux1-fill-dev, flux2-dev-Q4_K_M.gguf, and
mistral_3_small_flux2_fp8. They download without an account from the links the panel pins, but the
panel will not fetch them for you — read the licence before selling anything made with them.
Everything in the recommended starter path above is ungated.