ComfyUI INT8 image models cut runtime by about 20% and reach 2–3 seconds on a 4090
tisch_eins · x · 2026-07-27
This post compares INT8 versions of several image generation models inside ComfyUI, including Boogu-Image, Flux-2-Klein, and Z-Image-Turbo.
The author says the INT8 versions are faster than FP8 in their setup, saving about 20% of the runtime, while still producing equal or better quality than the FP8 models they tested. On a 4090, images reportedly take only 2–3 seconds.
The post includes example renders and links to the model files needed for each setup, making it a practical performance note for people tuning local image-generation pipelines.
More from Multimodal
- InVideo Agent One ranks #1 on an independent AI video-agent benchmark — aziz4ai · 2026-07-27
- Microsoft’s first text-to-image model scores 49% in a 192-prompt benchmark but lags on realism — dh7net · 2026-07-27
- Seedance 2.0 shows strong image-to-video motion transfer with a single prompt — miilesus · 2026-07-27
- A 480K-parameter latent model cleans up GPT Image 2 speckle artifacts — Parking_Baby_57 · 2026-07-27
- A Krea2 LoRA training trick keeps face detail at 512px by adding crops — More_Bid_2197 · 2026-07-27
- User tries to pair external audio with an LTX LoRA video workflow — itchplease · 2026-07-27