Datasets:
license: apache-2.0
pretty_name: MOSS and Z-Image quantization benchmarks
language:
- pl
- en
task_categories:
- text-to-image
tags:
- benchmark
- quantization
- image-generation
MOSS-Image-SFT and Z-Image — quantization evidence
Single research archive for source-matched generative-image benchmarks: original MOSS-SFT DiT quantizations, original Qwen/VAE precision conversions, Polish glyphs, scene retention and independent Z-Image Turbo runs. Contains authentic PNGs, parameters, seeds, measured latency/allocator values, SHA-256 and diagnostic records. No speedup or text-correctness claim follows from file size or image similarity alone.
How to read these reference examples
These PNG/JPG images and descriptive measurements are reference examples, not a standardized leaderboard or dataset-loader API. Browse the gallery and associated results in Files and versions; reproducing the experiments is not required to try any checkpoint.
Some samples intentionally show misspellings, changed compositions and failed runs. They are not a certified-quality showcase.
License and provenance: This dataset's original curation, generated samples, associated results and metadata are released under Apache License 2.0 by the dataset publisher. The underlying MOSS-Image-SFT and Z-Image-Turbo source models use Apache-2.0. Any third-party content (including the one archived historical HF comparison sheet in historical_references/) is supplied as reference material; its pre-existing rights are not relicensed by this dataset. Users should verify rights before independently redistributing third-party imagery.
Results depend on the hardware, sampler, step count, seed, encoder and precision. Do not merge the original 50-step SFT, experimental 10-step DPM++2M and four-step student-adapter trials into one speed or image-quality ranking. PSNR, SSIM and LPIPS indicate divergence from a reference; they do not score Polish spelling, readability, aesthetics or compatibility with a new GPU.
All eleven component releases are experimental. See each model card for format-specific validation gaps and further-testing recommendations. Missing native-kernel validation means unverified on the recorded setup, not a universal incompatibility.
Checkpoint comparison
All eight DiT files derive from OpenMOSS-Team/MOSS-Image-SFT at fc5de0d56fc910cbca6f6625a70f5f582b26deb1 and preserve the original model (50-step reference). All use actual TorchAO checkpoint serialization, not a student/LoRA export.
| DiT checkpoint | Weight GiB | Quantized Linear | Reference evidence and testing status |
|---|---|---|---|
| INT8-WO-Core | 5.470 | 228/269 | 5 SFT50 pairs; broader text/visual testing recommended |
| FP8-WO-Core | 5.463 | 228/269 | 5 SFT50 pairs; broader text/visual testing recommended |
| FP8-WO-Max | 4.739 | 269/269 | 5 SFT50 pairs; broader visual testing recommended |
| FP8-W8A8-Row-Core | 5.464 | 228/269 | Original SFT50 image test unavailable; native runtime requires testing |
| FP8-W8A8-Row-Max | 4.739 | 269/269 | Original SFT50 image test unavailable; native runtime requires testing |
| FP8-W8A8-Tensor-Core | 5.459 | 228/269 | 4 original SFT50 images on SM120; additional tests required |
| NVFP4-Core | 3.706 | 228/269 | No validated native FP4 SFT50 images; native runtime requires testing |
| NVFP4-Max | 2.663 | 269/269 | No validated native FP4 SFT50 images; native runtime requires testing |
BF16 original DiT storage: 9.467 GiB. Storage size is not full pipeline size, resident GPU VRAM or execution throughput. W8A8 and NVFP4 native kernel support was missing in the recorded SM80 original-50-step benchmark, not on every GPU universally.
Independent original components: Qwen FP8-WO, VAE FP16, VAE BF16.
Original full SFT — 50-step, seed-matched smoke test
5 images at 1024px, CFG 3 and seed 42, original model sampler, CMP 170HX / SM80. Median latency and peak allocated CUDA are scope-specific; quality metrics measure distance to BF16.
| Checkpoint | Median s/image | Peak CUDA allocated GiB | SSIM vs BF16 ↑ | LPIPS ↓ |
|---|---|---|---|---|
| BF16 original | 53.176 | 12.542 | Reference | Reference |
| INT8-WO Core | 58.824 | 8.544 | 0.89499 | 0.05729 |
| FP8-WO Core | 63.361 | 8.538 | 0.84926 | 0.08454 |
| FP8-WO Max | 65.320 | 7.816 | 0.78998 | 0.10494 |
Further source-matched grids and absolute difference maps.
Independent experimental 10-step study
DPM++2M10, 1024px, CFG 3, seed 42; 16 real images per variant. Not the original 50-step SFT protocol and not the 4-step student adapter. Same CMP 170HX / SM80 environment.
| DiT | Median s/image | Median CUDA reserved GiB | SSIM vs BF16 ↑ | Learned LPIPS ↓ |
|---|---|---|---|---|
| Original BF16 | 11.029 | 14.156 | Reference | Reference |
| INT8-WO Core | 12.074 | 10.344 | 0.9097 | 0.053268 |
| FP8-WO Core | 12.991 | 10.490 | 0.8559 | 0.073368 |
On this setup the packed models saved reserved CUDA memory but did not beat BF16 on runtime. Other GPU/kernel choices need distinct measurements.
All 32 paired image scores · pretrained LPIPS · complete aggregate limitations.
Raw generation archive (other GPU and Z-Image)
runs/moss-image-20261021/ holds a separate RTX 5060 Ti experiment with original 50-step MOSS-SFT, four-step student adapters, two-step kernel probes, native/high-resolution generations, upscales, and independent Z-Image Turbo API outputs. These have different timing scopes, sampling policies and conditioning environments. They are not one homogeneous benchmark. Historical reference discrepancy on “RIVER BAKERY” also occurred with original BF16 under local runtime controls; it is not evidence that quantization alone caused the issue.
| Subset | Images | Purpose |
|---|---|---|
quality |
268 | Original generated outputs and parameters |
upscales |
157 | Upscaled images plus parent references |
diagnostics |
12 | Technical execution probes (not output-quality evidence) |
historical_references |
1 | Archived comparison reference (not current-device latency) |
runs/moss-image-20261021/records/measurements.jsonl preserves raw successful and failed attempts. quality/metadata.jsonl preserves image hashes, expected text, model/precision, steps/CFG, seed, GPU and timings where measured; missing values remain null.
Limitations: PSNR/SSIM/LPIPS measure similarity to a reference PNG, not correctness of Polish characters, naturalness, real-world image quality or independent blind human preference. Store original component hashes, exact sampling protocol and device details with any derived figure. This repository is private research evidence, not a validated public model-quality leaderboard.
Research suites and sanitized diagnostics. The 50-step prompt suite and style benchmark suite are stored once here, rather than repeated across eight checkpoint repos. A compact compatibility summary accompanies raw individual runs. Server filesystem prefixes in diagnostics were redacted; numeric results, prompts, SHA-256 image hashes and status fields remain. Dataset remains private pending separate image-rights review.