I use ComfyUI for images and video on a 16 GB card and had never put it side by side with Forge on the same machine. So I measured ComfyUI, Forge Neo and the original Forge on an RTX 4070 Ti SUPER with the same files, seeds and settings, diffusers as a baseline, 36 paired images per model and the decision rule written down before the first run.
| Model | ComfyUI | Forge Neo | Original Forge | diffusers |
|---|---|---|---|---|
| SDXL (fits in 16 GB) | 6.71 s | 6.71 s | 6.76 s | 6.85 s |
| Z-Image Turbo bf16 (does not fit) | 6.75 s | 13.39 s | n/a | 15.81 s |
| Qwen-Image fp8 (does not fit) | 75.7 s | 96.0 s | n/a | n/a |
Median seconds per image, 36 images each.
- While the model fits, they tie. With SDXL the three tools land within 1 % of each other.
- When it does not fit, ComfyUI wins. Twice as fast with Z-Image Turbo, and 21 % less time per image with Qwen-Image fp8. Paired bootstrap intervals exclude a tie in both cases.
- The logs point to memory management. ComfyUI v0.41 reports dynamic VRAM loading and «0 models unloaded»; Forge Neo logs three «Moving model(s)» lines per image, which add up to a median of 7.2 s with Z-Image, close to the 6.6 s gap.
- Same seed, nearly the same image for ComfyUI and Forge Neo. The original Forge has no engine for Z-Image or Qwen-Image; its last commit on main is from June 2025.
- Wan 2.2 A14B, one clip per tool as a curiosity: 877 s against 932 s, not enough to call a winner.
The post also covers what broke before measuring (a ComfyUI cache that faked a 0.05 s first image, a Forge Neo refiner switch that counted noise instead of steps) and the threats to validity.
Read the full measurements, the pre-registered sheet and the scripts → https://efraingaray.com/en/blog/comfyui-vs-forge-16gb/
