Summary
Veteran AI runs a structured eight-category evaluation of Krea 2 Turbo — the eight-step distilled image generation model released by Krea AI — inside ComfyUI, providing a practical guide to setting up and getting the most out of this open-source model.
The video opens with a clear explanation of where Krea 2 sits in the landscape: it is an open-source model emphasizing creative and aesthetic exploration, available in two variants — RAW (the undistilled base, suited for fine-tuning) and Turbo (the eight-step distilled version tested here). Setup uses ComfyUI’s new Template Library with the FP8 Turbo model packaged by Comfy-Org, a Qwen3 VL vision-language model as the text encoder, and Qwen Image VAE for decoding. The video also demonstrates the same workflow running on RunningHub for users without local hardware.
Tests cover raw image quality with complex multi-element scenes, short-versus-long prompt comparisons with a side-by-side fashion portrait test, the built-in Prompt Enhancer (which uses Qwen3 VL to expand brief prompts into structured image descriptions), subject-to-aesthetic switching, text rendering, and high-resolution composition. The Prompt Enhancer section is particularly detailed: the video walks through the system prompt engineering behind it, showing exactly how it expands a four-word input like “a lonely robot in a flower shop” into a full scene description with lighting, mood, and compositional direction. Style LoRAs are introduced but not the main focus. CFG is fixed at 1 throughout, with Euler sampling and the Simple scheduler.
📺 Source: Veteran AI · Published June 29, 2026
🏷️ Format: Benchmark Test
