Hardware and Performance

Hardware and Performance

What this project targets

A 12 GB card. That is the design constraint, and it is why some models were removed rather than left in the list — full-precision Flux1-dev is 23.8 GB before its text encoders, which does not fit, so it is gone. Every preset the panel lists is one you can actually run.

• 8 GB — works, with the smaller stacks. Expect offloading on the heavier ones.
• 12 GB — the target. Everything runs; the largest stacks offload.
• 16 GB+ — comfortable across the board.

Photoshop 2024 or newer. Windows 11 with Photoshop 2025 is what has been verified end to end; macOS is untested and reports either way are welcome.

Measured timings

Real numbers from a 4070 Ti (12 GB), not estimates:

Preset	Resolution	Time
txt2img-flux2-klein	1024×1024, 4 steps, CFG 1	~12 s
unflatten-qwen-layered	4 layers at 640px	~2 min
Flux.2 dev (GGUF)	1024×1024	minutes, not seconds

The GGUF preset is slow on a 12 GB card by nature — it is a 20 GB quantised model with an 18 GB text encoder. It is included for people with headroom, not as a default.

"What will run well"

The panel's Setup screen ranks every preset against the VRAM ComfyUI reports for your card: Comfortable, Tight, Will offload, or Not known.

Two things worth understanding about that rating:

• It rates on the largest single file in a preset's stack, not the total, because that is what has to fit at once.
• "Will offload" means slower, not broken. A stack too big for the card still runs; it just moves weights in and out of VRAM while it works.

It is a comparison of published file sizes against reported VRAM — not a measurement of what a run actually consumes. A preset whose model size is unpublished is reported as unknown rather than guessed.

Making it faster

1. Use a distilled model. Klein at 4 steps and Z_image_Turbo at 8 steps are an order of magnitude faster than a full-step model, and for most work the difference in the result is small.
2. Generate smaller, then Upscale. The pixel upscaler costs 67 MB and a couple of seconds.
3. Start ComfyUI with --preview-method auto. It does not speed anything up, but you see the image forming instead of waiting blind.
4. Close other GPU work. ComfyUI, a browser with hardware acceleration, and Photoshop are all competing for the same card.

Ports

What	Default	Change it
ComfyUI	8190	Settings → Find ComfyUI Active Port finds a server on any port
Agent Bridge hub	8199	--port <n> on both bridge commands, and the same value in Setup

OpenLayer defaults to 8190 rather than ComfyUI's usual 8188 so it does not collide with another tool already using your server. You do not have to move your server — the port finder will locate it.