Merserk/Krea-2-Turbo-Checkpoint-Format-Benchmark

Krea 2 Turbo ComfyUI Format Fidelity Benchmark

This release is a paired, deterministic comparison of eight Krea 2 Turbo checkpoint formats in ComfyUI: BF16, FP8 Scaled, INT8 ConvRot, MXFP8, NVFP4, INT4 ConvRot W4A4, GGUF Q8_0, and GGUF Q4_K_M. It contains 240 scored 1024×1024 images, saved float32 decoded tensors and final latents, every denoising trajectory, raw metric tables, telemetry, statistical comparisons, and reproduction code.

Main result

1. BF16 is the highest-fidelity reference because it is the unquantized published checkpoint.
2. INT8 ConvRot narrowly leads the preregistered LPIPS-Alex endpoint. GGUF Q8_0 is statistically tied with it on that endpoint (Holm-adjusted permutation p=0.9067) and leads INT8 on DISTS, DINO similarity, and reconstructed-weight SNR.
3. GGUF Q8_0 is the highest-fidelity of the three added formats.
4. MXFP8 ranks fourth in fidelity (LPIPS 0.071229) and halves checkpoint size versus BF16, but its native SM 10.0 fast path was unavailable on this Ada GPU, resulting in 31.929 s fallback sampling.
5. FP8 Scaled ranks fifth (LPIPS 0.093701) at 12.239 GiB and was the fastest of the original five formats at 19.503 s on this GPU.
6. GGUF Q4_K_M is a storage-quality compromise: 6.972 GiB and much closer to BF16 than INT4 ConvRot, but slow on this Ada GPU under ComfyUI-GGUF dequantized execution.
7. NVFP4 is 7.147 GiB with moderate fidelity loss (LPIPS 0.205124). Like MXFP8, its native fast multiplication requires SM 10.0; the recorded 24.925 s sampling time uses the fallback path.
8. INT4 ConvRot W4A4 is the smallest checkpoint and fastest of the three additions on this RTX 4060 Ti, but has the largest measured fidelity loss.

No weighted composite score is used. See the full technical report and tables/decision_table.csv.

Important performance limitation

The test GPU was an NVIDIA GeForce RTX 4060 Ti 16 GB (SM 8.9). MXFP8 and NVFP4 native fast matrix multiplication requires SM 10.0 in the tested runtime. GGUF Q8_0 and Q4_K_M use the pinned ComfyUI-GGUF on-demand dequantized matrix-multiplication path. These Ada timings must not be projected to native Blackwell execution.

The original five formats and the three additions were measured in separate sessions. Five unscored BF16 and INT8 ConvRot bridge repeats quantify drift. INT8 crossed the preregistered 5% drift threshold, so exact old-versus-new timing comparisons remain caveated; fidelity comparisons remain paired to the same saved BF16 prompt/seed outputs.

Dataset organization

• data/train/metadata.jsonl and data/train/images/ are the Dataset Viewer and loading source, using Hugging Face ImageFolder. The dataset card explicitly limits the train split to this tree so scientific JSON files elsewhere in the repository are not inferred as dataset splits. Each metadata row links an image to its prompt, seed, format, checkpoint provenance, raw scientific artifacts, and flattened metric values.
• raw/ contains decoded_float32.npy, final_latent_float32.npy, trajectory.npz, and capture metadata.json for every scored run.
• metrics/ contains raw per-image, per-parameter, paired-statistics, summary, trajectory, latency, and performance tables.
• comparison_sheets/ contains eight-format sheets, BF16-relative difference maps, and automatically selected detail crops.
• telemetry/ contains scored GPU/system telemetry plus the BF16/INT8 bridge telemetry.
• provenance/ contains sanitized environment, model, run, and release manifests.
• reproduction/ contains the ComfyUI workflows, custom capture node, benchmark driver, analyzers, model downloader, and exact instructions.

Load the image table locally:

from datasets import load_dataset

dataset = load_dataset("Merserk/Krea-2-Turbo-Checkpoint-Format-Benchmark", split="train")
print(dataset.num_rows)  # 240

Fixed inference controls

• 15 prompts × 2 deterministic seeds × 8 checkpoint formats = 240 scored images
• 1024×1024, batch size 1
• 8 steps, CFG 1.0, Euler sampler, simple scheduler, denoise 1.0
• Shared qwen3vl_4b_bf16.safetensors text encoder and qwen_image_vae.safetensors VAE
• Zeroed positive conditioning used as negative conditioning
• No prompt rewriting, LoRAs, previews, upscaling, or post-processing
• Identical initial-noise SHA-256 for every eight-format prompt/seed group

Reproduce or verify

Follow reproduction/README.md. A complete rerun downloads about 104 GiB of upstream model files and should have at least 150 GiB free for models, captures, analysis caches, and reports. The published benchmark was tested on Windows 11 with a 16 GB NVIDIA GPU; lower-memory configurations are not certified.

Validate this downloaded release:

python scripts/validate_release.py --root . --full

Upload a locally rebuilt copy:

hf auth login
python scripts/upload_to_huggingface.py OWNER/DATASET --root .

Licenses and responsible use

Generated images, data, and reports are released under CC BY 4.0. Benchmark scripts are MIT licensed. Model weights are not redistributed and remain governed by the Krea 2 Community License. Krea states that it does not claim intellectual-property rights over user-generated outputs, while users remain responsible for their prompts, outputs, and downstream use. Review the official Krea 2 Turbo model card and license before downloading or running the checkpoints.

The prompt suite avoids requests for real people, explicit content, unlawful activity, or protected logos. Generated images can still contain model errors, stereotypes, malformed text, or unintended resemblance. This benchmark evaluates checkpoint fidelity on one controlled campaign; it is not a universal ranking of human aesthetic preference or all hardware.

Citation

See CITATION.cff. The project attribution name is Krea 2 Turbo Formats Benchmark Contributors.