Sign In
homemodelsimagesvideos3D Modelsarticlescomicschallengesupdatesshop

Krea 2 Turbo Checkpoint Format Benchmark in ComfyUI: BF16 vs FP8 vs INT8 ConvRot vs MXFP8 vs NVFP4

1

MM744's Avatar

MM744

Jul 12, 2026

(Updated: 19 days ago)

comparative study
krea 2 turbocomparisonworkflownvfp4mxfp8

Krea 2 Turbo Checkpoint Format Benchmark in ComfyUI

BF16 vs FP8 Scaled vs INT8 ConvRot vs MXFP8 vs NVFP4 using 150 matched ComfyUI generations.

Which Krea 2 Turbo checkpoint format is actually best?

image.png

TL;DR

I ran a controlled ComfyUI benchmark of every official Krea 2 Turbo checkpoint format: BF16, FP8 Scaled, INT8 ConvRot, MXFP8, and NVFP4.

  • Best absolute fidelity: BF16. It is the unquantized reference.

  • Best quantized format: INT8 ConvRot. It stayed closest to BF16 across perceptual, semantic, latent, trajectory, and reconstructed-weight measurements.

  • Best measured speed/quality balance on my RTX 4060 Ti: INT8 ConvRot.

  • Smallest checkpoint: NVFP4 at 7.15 GiB, but it also produced the largest measured shift from BF16.

  • Hardware caveat: MXFP8 and NVFP4 used fallback/dequantized execution on this SM 8.9 GPU. Their measured speed should not be treated as a forecast for native Blackwell/SM 10.0 execution.

image.png

Plain-language recommendation by use case.

The short decision table

image.png

* MXFP8 and NVFP4 used fallback execution on Ada. Retest them on Blackwell before making a speed decision.

image.png

Four independent fidelity measurements. Lower is better for LPIPS, DISTS, and latent error; higher is better for DINO similarity.

image.png

Checkpoint size versus matched-image LPIPS.

image.png

Sampling time versus matched-image LPIPS, with the Ada fallback warning.

What I tested

This was not a comparison of five unrelated generations. Every five-format prompt/seed group used identical initial noise.

  • 15 prompts covering portraits, hands, text, architecture, reflections, foliage, fog, material detail, vector geometry, anime, watercolor, food, spatial counting, and a dense scientific diagram

  • 2 deterministic seeds per prompt

  • 5 checkpoint formats

  • 150 scored images in total

  • 1024 x 1024, 8 steps, CFG 1.0, Euler sampler, and simple scheduler

  • Shared Qwen3VL 4B BF16 text encoder and Qwen Image VAE

  • No LoRA, prompt rewriting, previews, upscaling, or post-processing

I also saved decoded float32 tensors, final latents, and all eight denoising trajectory states. The quantized parameters were reconstructed and audited against BF16.

image.png

Controlled benchmark design: 15 prompts x 2 seeds x 5 formats.

What the metrics mean

  • LPIPS and DISTS: How much the matched output changed perceptually from BF16. Lower is closer.

  • DINO cosine similarity: Whether high-level image features remained similar. Higher is closer.

  • Final latent relative L2: How far the diffusion state diverged before VAE decoding. Lower is closer.

  • Weight SNR: How accurately the quantized checkpoint reconstructs BF16 weights. Higher is better.

I did not combine these measurements into an arbitrary composite quality score. The primary conclusion uses the preregistered LPIPS endpoint, with independent metrics as supporting evidence.

Why INT8 ConvRot performed so well

INT8 ConvRot uses row-wise INT8 weights with group-size-256 Hadamard rotation while retaining sensitive projections and conditioning components at higher precision.

Its reconstructed-weight SNR was 41.16 dB, compared with approximately 31.6 dB for FP8 Scaled and MXFP8, and 20.60 dB for NVFP4. This advantage remained visible through the denoising trajectory and final images.

INT8 was not perfect - dense diagrams and vector geometry still changed - but it was consistently the closest quantized format overall.

image.png

Reconstructed-weight fidelity and signal-to-noise ratio.

Visual comparisons

Easy scenes can look nearly identical at normal feed size. Dense geometry, scientific diagrams, and vector edges make format-specific divergence easier to see.

image.png

Portrait comparison: subtle differences require matched seeds and close inspection.

image.png

Scientific-poster comparison: dense geometry exposes structural drift.

image.png

Scientific-poster detail crops.

image.png

Vector-gradient comparison and BF16-relative difference map.

Full prompt coverage

image.png

Prompt contact sheet 1: portraits, hands, packaging, storefronts, and architecture.

image.png

Prompt contact sheet 2: reflections, foliage, fog, materials, and vector design.

image.png

Prompt contact sheet 3: anime, watercolor, food, counting, and scientific diagrams.

Which format should you use?

  • BF16: Choose it when preserving the published checkpoint is the priority and storage/offloading cost is acceptable.

  • INT8 ConvRot: My default quantized choice on this tested ComfyUI/Ada stack.

  • MXFP8: Worth retesting on Blackwell. Its fidelity was better than FP8 Scaled, but its measured Ada speed came from a fallback path.

  • FP8 Scaled: Usable and approximately half the BF16 file size, but it did not beat INT8 in this campaign.

  • NVFP4: Choose it when minimum checkpoint size matters more than matching BF16. It deserves a separate native-Blackwell benchmark.

image.png

Final recommendation: choose by constraint, not by bit width alone.

Full research and reproducibility files

The free Hugging Face dataset contains all 150 images, raw tensors, denoising trajectories, metric tables, comparison sheets, model hashes, ComfyUI workflows, and scripts:

https://huggingface.co/datasets/Merserk/Krea-2-Turbo-Checkpoint-Format-Benchmark

What do you think I should test next: INT4, GGUF Q8_0, or another format? I included all scripts, workflows, and reproducibility files, so you can rerun the benchmark with your own model variant or hardware.

1