Testing Krea 2 Turbo Quantization Formats
Merserk13 · reddit · 2026-07-13
The author tested various checkpoint formats for Krea 2 Turbo in ComfyUI: BF16, FP8 Scaled, INT8 ConvRot, MXFP8, and NVFP4.
Key Findings
- BF16: Best absolute fidelity, serves as the baseline
- INT8 ConvRot: Best overall performance among quantized formats, closest to BF16
- Speed/Quality Balance on RTX 4060 Ti: INT8 ConvRot is optimal
- NVFP4: Smallest file size but largest quality deviation
- MXFP8 / NVFP4: Experienced fallback/dequantization execution on Ada GPUs, so speed results cannot be directly extrapolated to Blackwell
Methodology
- 15 prompt groups covering portraits, hands, text, architecture, reflections, fog, materials, vector graphics, anime, watercolor, food, counting, and scientific diagrams
- 2 deterministic seeds per group
- 150 images total
- Fixed resolution at 1024×1024, 8 steps, CFG 1.0, Euler sampler
- Saved latents, denoising trajectories, and quantized weights for a detailed error audit
Recommendations
- For original release fidelity: use BF16
- Default quantization scheme: INT8 ConvRot
- MXFP4 is worth retesting on Blackwell
- FP8 Scaled is usable, but underperformed compared to INT8 in this test
Related event: Benchmarking Krea 2 Turbo Across Quantization Formats(3 posts)→
More from Multimodal
- Storyboard-first workflows are making AI dance videos and influencers more consistent — aftahi_ai · 2026-07-22
- Interactive video should be judged by responsiveness, not just frame quality — Soggy_Limit8864 · 2026-07-22
- Runpod MCP and Claude help spin up image and video generation workflows — 802high · 2026-07-22
- Midjourney prompt turns a bee into a glitching pixel explosion — michaelrabone · 2026-07-22
- A physics reward can improve video generation without creating a real physics engine — Dapper-Drawer4546 · 2026-07-22
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22