# Krea 2 Turbo Checkpoint-Format Benchmark

> This report is generated from the saved raw tensors, trajectories, telemetry, and metric tables. No result is inferred from filename or nominal bit width.

## Executive conclusion

- **Highest model fidelity:** BF16, because it is the unquantized reference checkpoint used for all paired comparisons.
- **Best quantized BF16 fidelity by the preregistered primary endpoint (LPIPS-Alex):** INT8 ConvRot.
- **Primary-endpoint separation from the next quantized format:** not statistically separated at α=0.05; treat the numerical ordering as a near-tie.
- **Smallest checkpoint:** INT4 ConvRot W4A4.
- **Ada performance warning:** MXFP8 and NVFP4 use fallback/dequantized matrix multiplication on this SM 8.9 GPU; their recorded speed must not be projected to Blackwell hardware.
- **INT4 execution note:** ConvRot W4A4 is eligible for the native SM 8.x signed-INT4 MMA path in this comfy-kitchen runtime.
- **GGUF execution note:** Q8_0 and Q4_K_M use the pinned ComfyUI-GGUF loader and its on-demand dequantized matrix-multiplication path, not ComfyUI's native safetensors quantization layouts.
- Automated results establish original-model fidelity, prompt alignment, and measurable artifact behavior. They do not establish universal human aesthetic preference.

### Decision table

| Format | File GiB | LPIPS↓ | DISTS↓ | DINO cosine↑ | Final latent rel-L2↓ | Weight SNR dB↑ | Sampling s↓ | Peak VRAM MiB↓ | Energy J↓ |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| BF16 | 24.478 | 0.000000 | 0.000000 | 1.000000 | 0.000000 | n/a | 25.783 | 15221.0 | 3656.0 |
| FP8 Scaled | 12.239 | 0.093701 | 0.042738 | 0.971012 | 0.223618 | 31.585775 | 19.503 | 14397.0 | 3013.5 |
| INT8 ConvRot | 12.566 | 0.041864 | 0.026796 | 0.983833 | 0.132571 | 41.157030 | 12.417 | 15165.0 | 1944.8 |
| MXFP8 | 12.603 | 0.071229 | 0.035132 | 0.979443 | 0.201900 | 31.604820 | 31.929 | 14750.0 | 4142.8 |
| NVFP4 | 7.147 | 0.205124 | 0.084436 | 0.934769 | 0.381361 | 20.600707 | 24.925 | 14459.0 | 3679.3 |
| INT4 ConvRot W4A4 | 6.426 | 0.314728 | 0.132647 | 0.911258 | 0.490868 | 15.715704 | 9.537 | 14623.0 | 1440.8 |
| GGUF Q8_0 | 12.765 | 0.042618 | 0.024559 | 0.990568 | 0.145688 | 44.813097 | 27.493 | 14267.0 | 3997.6 |
| GGUF Q4_K_M | 6.972 | 0.153252 | 0.066291 | 0.954597 | 0.340616 | 22.973296 | 31.407 | 14333.0 | 4487.2 |

The decision table is multidimensional by design. No weighted composite score is used.

## Three-format extension conclusion

- **Best fidelity among the three additions: GGUF Q8_0.** It records LPIPS 0.042618, DISTS 0.024559, DINO cosine 0.990568, and 44.81 dB reconstructed-weight SNR. It is also effectively tied with INT8 ConvRot on the primary endpoint (Holm-adjusted permutation p=0.9067), while leading INT8 on DISTS, DINO cosine, and weight SNR.
- **Best minimum-size/speed option among the three: INT4 ConvRot W4A4.** It is 6.426 GiB and sampled in 9.537 s on this RTX 4060 Ti, but it has the largest quality loss: LPIPS 0.314728, final-latent relative L2 0.490868, and 15.72 dB weight SNR.
- **Storage-quality compromise: GGUF Q4_K_M.** At 6.972 GiB it is only 8.5% larger than INT4 ConvRot, while reducing LPIPS by 51.3%. Its 31.407 s sampling time is 3.29× INT4's time because this Ada test uses ComfyUI-GGUF's dequantized execution path.
- **Q8_0 costs size and Ada runtime for fidelity.** Its 12.765 GiB checkpoint is 1.99× the INT4 file and its 27.493 s sampling time is 2.88× INT4's, but it preserves the BF16 result far more closely on every paired fidelity endpoint.
- The three extension formats ran in the same session, so their timing comparison is direct. Exact timing comparisons against the original five-format session remain caveated by the bridge test.

## Test system and immutable controls

- GPU: `NVIDIA GeForce RTX 4060 Ti; driver 610.74; 16380 MiB; compute capability 8.9`
- ComfyUI commit: `917faef771a2fd2f14f44af94f17da3d0b2803a3` (2026-07-12T09:43:30-07:00)
- PyTorch: `2.13.0+cu130`; CUDA runtime: `13.0`; comfy-kitchen: `0.2.18`
- Python: `3.13.12 (tags/v3.13.12:1cbe481, Feb  3 2026, 18:22:25) [MSC v.1944 64 bit (AMD64)]`
- Resolution: 1024×1024; batch size 1
- Sampler: euler; scheduler: simple; steps: 8; CFG: 1.0; denoise: 1.0
- Shared text encoder: `qwen3vl_4b_bf16.safetensors`; shared VAE: `qwen_image_vae.safetensors`
- Negative conditioning is a zeroed copy of positive conditioning. Prompt rewriting, LoRAs, previews, upscaling, and post-processing are disabled.
- Each prompt/seed pair uses the same initial-noise tensor SHA-256 across all 8 formats.

## Checkpoint structure

| Format | File | SHA-256 | GiB | Loader | Tensors | Stored dtypes/qtypes | Metadata |
| --- | --- | --- | --- | --- | --- | --- | --- |
| BF16 | krea2_turbo_bf16.safetensors | 78bbf8f4165eda19cea3cb06c78089221932a39e2eed8af9da741f942c47ffb3 | 24.478 | unet | 430 | BF16: 256<br>F32: 174 | none |
| FP8 Scaled | krea2_turbo_fp8_scaled.safetensors | eb4dd8c612cfd10f64f25b057e6e6bbcb5737c94a7372177e456dbf7579502f1 | 12.239 | unet | 686 | BF16: 174<br>F32: 256<br>F8_E4M3: 256 | _quantization_metadata, format |
| INT8 ConvRot | krea2_turbo_int8_convrot.safetensors | 8e4eeda70dd5037ab1ba2bef6b417f9f901e26093117cf397f741fc1fdaaf3f1 | 12.566 | unet | 878 | BF16: 76<br>F32: 354<br>I8: 224<br>U8: 224 | none |
| MXFP8 | krea2_turbo_mxfp8.safetensors | 4c09131a442c3a01e95f934b43c76c1e691b9f683694a7b68d6fc707d5ecfce3 | 12.603 | unet | 686 | BF16: 174<br>F8_E4M3: 256<br>U8: 256 | _quantization_metadata, format |
| NVFP4 | krea2_turbo_nvfp4.safetensors | 61527003b2d537055494d01bc8efe51d6e86e64192ba23e3721a5647231fe394 | 7.147 | unet | 942 | BF16: 174<br>F32: 256<br>F8_E4M3: 256<br>U8: 256 | _quantization_metadata, format |
| INT4 ConvRot W4A4 | krea2_turbo_int4_convrot.safetensors | 4c74e8fc00d121aeb4f75cdf066b70d33f6a59e36f143b0bfd53d6467ffc13bd | 6.426 | unet | 686 | BF16: 174<br>F32: 256<br>I8: 256 | _quantization_metadata, converted_by |
| GGUF Q8_0 | krea2_turbo-Q8_0.gguf | 1fa2da08a7a708827c2100d0af41c8371d79efd9f7c5490c23ebc65408536490 | 12.765 | gguf | 432 | F32: 167<br>Q8_0: 265 | general.architecture, general.file_type, general.quantization_version |
| GGUF Q4_K_M | krea2_turbo-Q4_K_M.gguf | bc12f539de7a7a6ddf9bac17ec6e5bfecd42c3517190f491131d8464a1034b40 | 6.972 | gguf | 432 | F16: 1<br>F32: 167<br>Q4_K: 201<br>Q5_K: 63 | general.architecture, general.file_type, general.quantization_version |

### What the formats actually do

- **BF16 reference:** main matrix weights are BF16 while several norms/scales are stored as FP32. It preserves the published parameter values but exceeds the 16 GB VRAM capacity and therefore requires offloading.
- **FP8 Scaled:** E4M3 weights plus explicit tensor scales. The checkpoint marks 256 linear layers, including 96 linears configured for full-precision matrix multiplication and 160 eligible for quantized multiplication.
- **INT8 ConvRot:** 224 row-wise INT8 linear weights with group-size-256 online/offline Hadamard rotation. Sensitive input/output, conditioning, and text-fusion components remain higher precision.
- **MXFP8:** E4M3 values with E8M0 power-of-two block scales over 32-element blocks. It marks 256 linear layers; native fast multiplication requires SM 10.0 in this runtime.
- **NVFP4:** packed 4-bit floating-point values with per-tensor and 16-element block scales. It marks 256 linear layers; native fast multiplication likewise requires SM 10.0.
- **INT4 ConvRot W4A4:** 256 packed signed-INT4 linear weights with FP32 scales, ConvRot group size 256, quantization group size 64, and 174 BF16 tensors. ComfyUI 0.27 loads it through the native diffusion-model loader.
- **GGUF Q8_0:** third-party GGUF representation loaded by ComfyUI-GGUF. Its exact tensor qtype distribution is recorded in the manifest and its dequantized logical weights are audited against BF16.
- **GGUF Q4_K_M:** mixed K-quant GGUF representation using the same pinned loader. File size is not treated as proof of quality; reconstructed weights and matched outputs determine fidelity.

## Weight reconstruction fidelity

Every logical parameter was reconstructed with the same native comfy-kitchen layout used at inference and streamed against the BF16 safetensors reference. Aggregates are element weighted; unchanged high-precision tensors are included.

| Format | Logical tensors | Quantized tensors | MAE | RMSE | Relative L2 | Cosine | SNR dB | Max abs |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| BF16 | 430 | 0 | 0.000000 | 0.000000 | 0.000000 | 1.000000000 | n/a | 0.000000 |
| FP8 Scaled | 430 | 256 | 0.000722 | 0.001166 | 0.026346 | 0.999652901 | 31.586 | 0.187500 |
| INT8 ConvRot | 430 | 224 | 0.000277 | 0.000387 | 0.008753 | 0.999961693 | 41.157 | 0.091442 |
| MXFP8 | 430 | 256 | 0.000721 | 0.001163 | 0.026288 | 0.999654434 | 31.605 | 0.250000 |
| NVFP4 | 430 | 256 | 0.002923 | 0.004129 | 0.093318 | 0.995646025 | 20.601 | 0.531250 |
| INT4 ConvRot W4A4 | 430 | 256 | 0.005451 | 0.007245 | 0.163763 | 0.986854981 | 15.716 | 0.625000 |
| GGUF Q8_0 | 430 | 263 | 0.000201 | 0.000254 | 0.005746 | 0.999983494 | 44.813 | 0.022461 |
| GGUF Q4_K_M | 430 | 262 | 0.002456 | 0.003142 | 0.071013 | 0.997478661 | 22.973 | 0.144531 |

Full per-parameter results, including p50/p95/p99/p99.9 error samples and subsystem grouping, are in [`metrics/weight_parameters_all.csv`](metrics/weight_parameters_all.csv).

## Paired image fidelity

BF16 is a behavioral reference, not a real photograph. Low LPIPS/DISTS and high DINO/SSIM indicate that a quantized checkpoint follows the BF16 generation trajectory more closely; they do not independently prove a more attractive image.

| Format | LPIPS Alex↓ | LPIPS VGG↓ | DISTS↓ | MS-SSIM↑ | DINO cosine↑ | SSIM RGB↑ | ΔE2000↓ |
| --- | --- | --- | --- | --- | --- | --- | --- |
| BF16 | 0.000000 | 0.000000 | 0.000000 | 1.000000 | 1.000000 | 1.000000 | 0.000000 |
| FP8 Scaled | 0.093701 | 0.095434 | 0.042738 | 0.869319 | 0.971012 | 0.874324 | 2.588693 |
| INT8 ConvRot | 0.041864 | 0.050027 | 0.026796 | 0.956164 | 0.983833 | 0.948027 | 1.234503 |
| MXFP8 | 0.071229 | 0.094774 | 0.035132 | 0.917225 | 0.979443 | 0.901653 | 2.643008 |
| NVFP4 | 0.205124 | 0.223690 | 0.084436 | 0.768737 | 0.934769 | 0.796323 | 6.313444 |
| INT4 ConvRot W4A4 | 0.314728 | 0.336843 | 0.132647 | 0.642767 | 0.911258 | 0.662332 | 9.150108 |
| GGUF Q8_0 | 0.042618 | 0.055029 | 0.024559 | 0.952685 | 0.990568 | 0.944028 | 1.862765 |
| GGUF Q4_K_M | 0.153252 | 0.188202 | 0.066291 | 0.798899 | 0.954597 | 0.825529 | 5.339750 |

Statistics use 50,000 prompt-cluster BCa bootstrap resamples. Quantized-format pairwise tests use all 2^15 exact prompt-level sign permutations with Holm correction across 21 pairs per metric.

## Prompt adherence and no-reference quality

| Format | CLIPScore↑ | PickScore↑ | HPSv2↑ | MUSIQ↑ | MANIQA↑ | TOPIQ-NR↑ | NIQE↓ | BRISQUE↓ |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| BF16 | 41.327330 | 22.496188 | 0.287720 | 71.562683 | 0.464503 | 0.596940 | 4.279002 | 25.846741 |
| FP8 Scaled | 41.257368 | 22.677390 | 0.285767 | 72.232819 | 0.460313 | 0.598193 | 4.447991 | 25.120178 |
| INT8 ConvRot | 41.762024 | 22.668381 | 0.288574 | 70.952457 | 0.458785 | 0.596355 | 4.369028 | 26.581604 |
| MXFP8 | 41.587543 | 22.493693 | 0.289673 | 72.054489 | 0.463579 | 0.594426 | 4.395214 | 25.064392 |
| NVFP4 | 41.326290 | 22.693707 | 0.285645 | 69.801792 | 0.479385 | 0.591738 | 4.430767 | 26.218567 |
| INT4 ConvRot W4A4 | 40.899118 | 22.446506 | 0.286499 | 72.287605 | 0.481754 | 0.602120 | 4.407180 | 23.496765 |
| GGUF Q8_0 | 41.225086 | 22.583970 | 0.286743 | 71.787231 | 0.462959 | 0.598738 | 4.274705 | 25.504883 |
| GGUF Q4_K_M | 40.991884 | 22.502078 | 0.285156 | 71.055096 | 0.456792 | 0.583141 | 4.305879 | 25.181091 |

These models have different training biases, particularly on illustration and typography. They are reported independently and are never averaged into a single score.

### OCR on prompts with required text

| Format | Images | Exact target fraction↑ | Median exact fraction↑ | Normalized edit distance↓ | OCR confidence↑ | Recognized lines |
| --- | --- | --- | --- | --- | --- | --- |
| BF16 | 6 | 0.611111 | 0.666667 | 0.138889 | 0.764390 | 4.67 |
| FP8 Scaled | 6 | 0.555556 | 0.500000 | 0.145062 | 0.781645 | 4.33 |
| INT8 ConvRot | 6 | 0.666667 | 0.666667 | 0.133333 | 0.763667 | 4.67 |
| MXFP8 | 6 | 0.611111 | 0.666667 | 0.157265 | 0.759152 | 4.50 |
| NVFP4 | 6 | 0.611111 | 0.666667 | 0.152381 | 0.780622 | 4.67 |
| INT4 ConvRot W4A4 | 6 | 0.722222 | 0.666667 | 0.113492 | 0.778992 | 4.67 |
| GGUF Q8_0 | 6 | 0.611111 | 0.666667 | 0.195757 | 0.766976 | 4.50 |
| GGUF Q4_K_M | 6 | 0.722222 | 0.833333 | 0.074311 | 0.791182 | 3.83 |

OCR is evaluated only on the six package/storefront/scientific-poster images per format. Exact target fraction is the share of requested phrases found verbatim after uppercase normalization.

## Denoising trajectory

The sampler saved both current latent `x` and predicted denoised state `x0` after every one of the eight free-running steps. The trajectory tables include relative L2, cosine, spectrum-relative L2, non-finite counts, and final latent divergence. Because each format consumes its own preceding output, later-step error includes accumulated trajectory divergence and is not a teacher-forced layer error.

See [`metrics/trajectory.csv`](metrics/trajectory.csv) and [`metrics/image_core.csv`](metrics/image_core.csv).

## Artifact detectors

For every float decoded image the analysis records clipping, gradient distribution, flat-region banding proxies, chroma noise, Laplacian energy, high-frequency spectral power, 8/16/32-pixel grid-boundary energy, luminance, saturation, and BF16-relative edge/color drift. These diagnostics identify where to inspect; none is treated as an aesthetic score by itself.

## Runtime, memory, and energy on RTX 4060 Ti

| Format | End-to-end s | Sampling s | VAE s | Peak VRAM MiB | Peak process RAM GiB | Mean W | Energy J | Peak °C |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| BF16 | 27.154 | 25.783 | 0.360 | 15221.0 | 14.161 | 134.1 | 3656.0 | 71.0 |
| FP8 Scaled | 20.921 | 19.503 | 0.344 | 14397.0 | 2.278 | 143.8 | 3013.5 | 73.0 |
| INT8 ConvRot | 13.623 | 12.417 | 0.356 | 15165.0 | 2.309 | 141.7 | 1944.8 | 72.0 |
| MXFP8 | 33.244 | 31.929 | 0.345 | 14750.0 | 4.507 | 124.3 | 4142.8 | 69.0 |
| NVFP4 | 26.391 | 24.925 | 0.345 | 14459.0 | 3.730 | 139.0 | 3679.3 | 71.0 |
| INT4 ConvRot W4A4 | 10.919 | 9.537 | 0.347 | 14623.0 | 2.303 | 131.5 | 1440.8 | 71.0 |
| GGUF Q8_0 | 29.344 | 27.493 | 0.802 | 14267.0 | 5.498 | 136.3 | 3997.6 | 72.0 |
| GGUF Q4_K_M | 32.743 | 31.407 | 0.355 | 14333.0 | 2.314 | 136.4 | 4487.2 | 71.0 |

Latency summaries use 30 scored runs per format after three repeatability/warmup executions. GPU and system telemetry were sampled every 100 ms. Model-load and warmup records remain available in `runs.jsonl` and telemetry files but are excluded from steady-state medians.
The original five formats and the three extension formats were measured in separate sessions. Five unscored BF16 and INT8 ConvRot bridge repetitions quantify session drift; no timing is automatically rescaled.
Bridge verdict: **keep cross-session timing comparisons caveated** using a 5% median-drift threshold.

## Comparison sheets

- [Master contact sheet](comparison_sheets/contact_sheet_replicate0.png)
- [`comparison_sheets/full/`](comparison_sheets/full/) — labeled 8-format images
- [`comparison_sheets/details/`](comparison_sheets/details/) — three automatically selected high-detail crops per pair
- [`comparison_sheets/differences/`](comparison_sheets/differences/) — BF16-relative absolute-difference maps with a shared per-pair scale

Difference heatmaps show divergence, not necessarily defects: even a small early numerical shift can produce a coherent but compositionally different eight-step result.

## Prompt corpus

### p01_portrait — portrait_microdetail

Ultra-realistic close-up editorial portrait of an elderly woman by a north-facing window, natural unretouched skin with visible pores and fine wrinkles, individual silver-gray hairs and eyelashes, wet catchlights in both eyes, a dark indigo wool coat with clearly visible fabric weave, soft neutral background bokeh, restrained color grading, high micro-contrast only on the face.

### p02_hands_group — anatomy_occlusion_counting

Candid documentary photograph of exactly four adult friends seated around a small round wooden table, each person exchanging a different small object with another person, all eight hands fully visible, natural overlapping arms and fingers, ceramic cups and folded paper on the table, warm side lighting, realistic anatomy, medium-wide composition.

### p03_package_text — product_typography

Front-facing studio product photograph of a premium matte white beverage carton on a pale gray seamless background. The package has only three lines of crisp black printed text: ‘NORTHSTAR LAB’, ‘SERIES 08’, and ‘250 ml’. Add a thin cobalt-blue geometric border, a small silver cap, straight edges, subtle paper texture, and a soft contact shadow.

OCR targets: NORTHSTAR LAB, SERIES 08, 250 ML.

### p04_storefront_text — environment_typography_lowlight

Rainy nighttime street photograph of a small corner cafe, viewed straight on through wet air and reflections. Three clearly readable signs say ‘OPEN 24 HOURS’, ‘CAFE LUMEN’, and ‘7TH STREET’. Deep blue shadows, warm amber windows, red neon reflections on asphalt, fine rain streaks, realistic perspective and restrained cinematic contrast.

OCR targets: OPEN 24 HOURS, CAFE LUMEN, 7TH STREET.

### p05_architecture — geometry_repetition

Perfectly symmetrical photograph of a contemporary museum interior in strict one-point perspective, long rows of repeating concrete columns, hair-thin black railings, large glass panels, precisely aligned pale stone floor tiles, a distant centered doorway, diffuse skylight, clean straight verticals, fine architectural detail from foreground to background.

### p06_reflections — specular_transparency

Luxury studio still life of a polished chrome mechanical wristwatch beside a transparent rectangular perfume bottle on glossy black glass, tiny engraved numerals and gear details, controlled white strip reflections, realistic glass refraction and caustics, faint condensation, crisp edges, black background with smooth tonal falloff.

### p07_foliage_bird — high_frequency_natural

Backlit wildlife photograph of a small kingfisher perched among dense wet foliage, individually resolved blue and orange feathers, leaf veins, tiny droplets, thin crossing branches and spider silk, layered depth separation, soft forest background, natural green color variation, sharp subject without oversharpening.

### p08_fog_neon — gradients_lowlight

Empty foggy city street before dawn in monochromatic navy and cobalt blue, one distant magenta neon tube glowing through mist, extremely smooth sky and fog gradients, subtle shadow transitions, faint wet pavement reflections, preserved near-black detail, no crushed blacks, no visible banding, quiet minimalist composition.

### p09_material_macro — material_texture

Museum-style macro material study arranged as five adjacent samples: coarse woven natural linen, brushed stainless steel, cracked celadon ceramic glaze, porous wet black stone, and translucent handmade paper. Raking side light reveals microtexture while every material remains clearly distinct and neutrally colored.

### p10_vector_gradient — flat_graphics_banding

Minimalist Swiss-style vector poster with a warm ivory background, three perfectly aligned ultramarine circles, hair-thin black geometric construction lines, one large coral rectangle, and a perfectly smooth pale-yellow to orange gradient band. Precise spacing, clean flat fills, razor-sharp edges, no texture and no shadows.

### p11_retro_anime — line_art_stylized

Single frame from a refined 1980s retro-anime film: a young mechanic adjusting a tiny radio with both hands, expressive face, thin confident ink outlines, intricate patterned jacket, limited teal cream and burgundy palette, cel shading, detailed fingers and radio controls, clean line intersections, subtle analog film grain.

### p12_watercolor — soft_stylized_gradients

Delicate transparent watercolor landscape on cold-pressed cotton paper, pale dawn sky fading from cool gray-blue to warm peach, mist drifting between layered hills, translucent washes, granulating pigment, soft wet edges, sparse fine tree silhouettes and untouched white paper highlights.

### p13_food — commercial_complex_scene

High-end commercial food photograph of handmade mushroom ravioli in glossy brown butter sauce, visible steam, tiny sage leaves, grated cheese, crumbs and pepper, a transparent water glass, polished fork and folded linen, warm window light, shallow depth of field, appetizing natural color and detailed highlights.

### p14_spatial_counts — spatial_reasoning_counting

Surreal gallery installation containing exactly three red cubes beneath a clear glass arch, exactly two blue spheres behind the arch, and exactly one yellow cone in front, all six objects reflected in a shallow sheet of water, centered wide composition, neutral white room, physically consistent shadows and reflections.

### p15_scientific_poster — diagram_labels

Clean scientific cutaway poster of a compact greenhouse system on a white background, precise pipes, roots, vents and sunlight paths, four simple arrows with clearly readable uppercase labels ‘AIR’, ‘WATER’, ‘ROOTS’, and ‘LIGHT’, restrained green and blue palette, thin technical linework, balanced educational infographic layout.

OCR targets: AIR, WATER, ROOTS, LIGHT.

## Reproduction

```powershell
.\run_benchmark.ps1 preflight -Config .\config_extended.json
.\run_benchmark.ps1 weights -Config .\config_extended.json
.\run_benchmark.ps1 generate -Config .\config_extended.json
.\run_benchmark.ps1 analyze -Config .\config_extended.json
.\run_benchmark.ps1 sheets -Config .\config_extended.json
.\run_benchmark.ps1 report -Config .\config_extended.json
.\run_benchmark.ps1 verify -Config .\config_extended.json
```

The runner resumes only complete captures and rejects an existing result directory when the configuration hash differs. It never silently retries with different image settings.
The external capture sampler was also compared against ComfyUI's built-in KSampler on the INT8 checkpoint: final latent and decoded float image were bit-identical with zero maximum absolute error. Evidence is in `validation/sampler_equivalence.json`.

## Limitations

1. The corpus contains 30 paired prompt-seed observations per format, optimized for artifact sensitivity rather than broad population-level aesthetic preference.
2. BF16 is the original-checkpoint fidelity reference, not an external ground truth for semantic correctness.
3. Pixel metrics become conservative when quantization changes composition rather than producing a local defect.
4. No human preference test was requested; conclusions about aesthetics remain outside scope.
5. Performance is hardware- and loader-specific. Native safetensors quantization, ConvRot W4A4, and ComfyUI-GGUF follow different execution paths; storage and paired fidelity are more portable than timing.
6. The extension adds three formats in a later session. Bridge runs quantify drift, but a future single-session eight-format rerun remains the strongest timing design.

## Primary sources

- [Krea 2 Turbo model card and official inference settings](https://huggingface.co/krea/Krea-2-Turbo)
- [Comfy-Org Krea 2 repackaged checkpoints](https://huggingface.co/Comfy-Org/Krea-2)
- [Winnougan INT4 ConvRot Comfy models](https://huggingface.co/Winnougan/INT4-Convrot-Comfy-Models)
- [vantagewithai Krea 2 Turbo GGUF](https://huggingface.co/vantagewithai/Krea-2-Turbo-GGUF)
- [ComfyUI-GGUF](https://github.com/city96/ComfyUI-GGUF)
- [LPIPS paper](https://arxiv.org/abs/1801.03924)
- [DISTS paper](https://arxiv.org/abs/2004.07728)
- [DINOv2 paper](https://arxiv.org/abs/2304.07193)
- [HPS v2 paper](https://arxiv.org/abs/2306.09341)
- [MUSIQ paper](https://arxiv.org/abs/2108.05997)
- [MANIQA paper](https://arxiv.org/abs/2204.08958)
- [TOPIQ paper](https://arxiv.org/abs/2308.03060)

## Raw artifact index

- `manifest.json`: environment, dependency versions, checkpoint metadata, and hashes
- `run_matrix.json`: all 240 preregistered scored runs
- `runs.jsonl`: execution ledger including failures and timings
- `captures/`: float decoded images, final latents, trajectories, PNGs, and per-run metadata
- `metrics/`: all raw, summary, and inferential CSV/JSON tables
- `telemetry/`: 100 ms GPU and system telemetry
- `advanced_metric_cache/`: resumable per-model metric outputs