WaveCut/flux3-klein-pup-50k

FLUX 3 Klein PUP 50K

50,000 unchanged source image prompts paired with expansions from FLUX 3 Image Klein PUP in its original BF16 precision. A separate configuration contains 512 matched BF16/FP8 outputs. This is an evaluation corpus of model observations, not a set of verified ideal rewrites.

from datasets import load_dataset

ds = load_dataset("WaveCut/flux3-klein-pup-50k", split="evaluation")
comparison = load_dataset(
    "WaveCut/flux3-klein-pup-50k", "precision-comparison", split="evaluation"
)
print(ds[0]["original_prompt"])
print(ds[0]["enhanced_prompt"])

The dataset also contains 12,000 BF16 observations for image editing and bounding-box modes. The original 50,000 rows and 512-row precision comparison remain unchanged. The repository name retains 50k for compatibility.

Configuration	Rows	Input	Output
default	50,000	Original image prompt	Expanded generation prompt
precision-comparison	512	Matched original prompt	BF16 and FP8 outputs
i2i	6,000	Short edit + one reference image	Expanded edit instruction
i2i_bbox	3,000	Matched short edit + one reference image	Instruction + proposed source/target boxes
t2i_bbox	3,000	Short prompt + supplied layout	Caption + object table

All configurations use an evaluation split. Images are referenced through pinned source rows and hashes; image files are not included.

edits = load_dataset("WaveCut/flux3-klein-pup-50k", "i2i", split="evaluation")
boxes = load_dataset("WaveCut/flux3-klein-pup-50k", "t2i_bbox", split="evaluation")

Original 50K: sources and selection

Source	Selected prompts	Source license
DiffusionDB	48,168	CC0-1.0
PartiPrompts	1,632	Apache-2.0
DrawBench	200	Apache-2.0

All PartiPrompts and DrawBench rows are included. The DiffusionDB sample was selected from 2,000,000 metadata rows after length, content-score, contact and keyword filters. The eligible pool contained 1,207,864 unique prompts. DiffusionDB prompts have 1–40 whitespace-delimited words and at most 500 characters. Across the original 50,000-row configuration, 49,985 inputs have at most 40 words; 15 longer benchmark prompts remain as controls.

DiffusionDB selection uses deterministic minimum-hash samples with approximately equal counts across 13 heuristic topics and four word-length bins (1–4, 5–12, 13–24, 25–40). Topics are keyword heuristics, not human annotations. The sample is balanced for coverage and does not estimate the natural distribution of user requests. It includes natural-language requests, keyword lists, style prompts, misspellings and compositional benchmark prompts. It is predominantly English; DiffusionDB input language was not verified and is recorded as und.

Deduplication uses NFKC normalization, case folding and collapsed whitespace for identity only. Submitted source strings retain their original spelling, case and whitespace. original_prompt means the source text submitted to PUP; it does not establish that the source was written without AI assistance. Source revision pins, exact selected strata and filtering details are in manifest.json. Per-row provenance is in sources.

Original 50K: generation

The pinned checkpoint and system prompt t2i were used with thinking disabled, vLLM 0.31.0, BF16 weights and 256 concurrent sequences on one RTX PRO 6000 Blackwell 96 GB. Generation used temperature 0, presence penalty 1.5, max_tokens 3,584 and the official PUP stop tokens. Requests were queued in batches of 512 for the initial 12,000 inputs and 2,048 for the 38,000-input extension. Model outputs were stripped of leading/trailing whitespace; they were not edited, ranked or regenerated to remove failures.

The first 512 BF16 rows were generated in a control run with the same engine and decoding settings as the remaining rows. The precision comparison uses those same 512 source prompts, with an FP8_DYNAMIC checkpoint derived from the pinned original. Linear attention, visual modules, embeddings and lm_head remain BF16. Its first 64 inputs emphasize text, counting, color, spatial relations and misspellings; the other 448 follow deterministic hash order in the initial 12,000-input selection. All 512 controls were retained when the source corpus expanded to 50,000. This subset is not a random sample of all 50,000 prompts. Greedy outputs can still depend on runtime, hardware, batching and numerical precision.

Original 50K: columns

• id: SHA-256 of the normalized identity string.
• original_prompt, enhanced_prompt: source input and observed PUP output.
• sources: source repository, revision, license, category and row/image identifier.
• word_count, topic_heuristic, input_kind, language: input metadata.
• precision, output_tokens, finish_reason: generation metadata.
• checks: mechanical flags. quoted_text_preserved is nullable and only checks literal substrings in inferred text-rendering requests. It does not establish whether the resulting image would display that text. periodic_tail detects an exactly repeating final 256-character tail with period at most 32.

The comparison configuration contains both outputs, token counts, finish reasons and checks. The full corpus is one evaluation split; no train/test separation or contamination exclusion is claimed.

Original 50K: audit and limitations

BF16: 50,000 rows, 50,000 nonempty, 20 token-limit finishes, 11 periodic tails. The literal-text heuristic applies to 70 inputs and flags 1 outputs. The 512-row FP8 comparison has 2 token-limit finishes. Failed outputs remain in the data. Counts are recorded in audit.json.

Every input ID, original string and source record was checked against the frozen input corpus. Parquet export was read back and compared with the source records. Mechanical checks do not measure intent preservation, factual accuracy, aesthetic quality or downstream image improvement. Heuristic flags can miss errors and produce false positives. Content filtering does not certify that all text is suitable for every audience.

A known semantic error is retained: the input photo of egypt in 7 0 0 ad digital painting was expanded as 700 BCE in both BF16 and FP8. Numeric/date fidelity therefore requires evaluation even when generation finishes normally.

No images were generated for this dataset. Public benchmark prompts may already occur in a model's training data. Treat this dataset as a reproducible input/output record and use independent human or downstream evaluation for quality claims.

Licensing and attribution

The compilation and added evaluation annotations are released under Apache-2.0. The original 50K source texts retain the CC0-1.0 or Apache-2.0 terms declared by their pinned Hub repositories; inspect sources and NOTICE. PUP declares Apache-2.0. No FLUX image-generation weights, images, or model weights are redistributed.

Editing and bounding-box extension

English inputs contain 1–12 whitespace-delimited words and at most 160 characters. Chinese inputs contain at most 48 characters. Source strings retain their exact spelling, case and whitespace. OmniEdit supplies several variants; selection uses its shortest existing variant, without rewriting it. Selection balances source task categories and length bins with a deterministic seed. This sample measures coverage, not the population frequency of editing requests.

Source	Unique editing inputs	Declared source terms
MagicBrush	2,800	CC BY 4.0
OmniEdit	1,600	MIT
GEdit-Bench	600	MIT
AnyEdit	1,000	CC BY 4.0

MagicBrush uses first-turn train/dev requests. OmniEdit uses dev and four fixed train shards. GEdit contributes 300 English and 300 Chinese requests, with possible shared scenes across languages. AnyEdit uses single-image validation tasks; visual_* tasks are excluded because their additional references are not supplied. The source category labels are retained. MagicBrush categories use a keyword heuristic. Half of the selected inputs from each source are also run in i2i_bbox. Thus the 9,000 editing observations contain 6,000 distinct input identities.

The 3,000 t2i_bbox inputs come from LayoutGPT NSR-1K: 895 spatial and 2,105 counting layouts, drawn from its train and validation files. Identical short text with a different supplied layout can appear more than once. Coordinates are converted from fractional [x, y, width, height] to integer [y_min, x_min, y_max, x_max] on a 0–1000 grid. Invalid or empty boxes are excluded. Original source coordinates remain available in source.source_boxes_xywh_json. These supplied layouts are source annotations; they are not human labels of PUP outputs.

The extension uses the complete vision-language checkpoint in BF16, one RTX PRO 6000 Blackwell Server 96 GB, vLLM 0.31.0, 128 concurrent sequences and batches of 128. The pilot uses 16 requests per mode. Generation is greedy, thinking is disabled, and the output limit is 3,584 tokens. Presence penalty is 0 for editing and 1.5 for t2i_bbox. Official system prompts and scaffolds are preserved. Version pins, model/runtime settings, input hashes, image preprocessing and batch timings are recorded in manifest-modes.json.

i2i_bbox explicitly selects the official system prompt without supplying locked input boxes. This differs from the official CLI's automatic mode selection. Its source and target coordinates are model proposals. They are not ground truth and may contain localization, identity or edit errors.

raw_output preserves the complete decoded output. enhanced_prompt removes outer whitespace. For t2i_bbox, the official reconciliation helper also restores locked input coordinates and matching IDs. raw_checks and processed_checks separate model behavior from that postprocessing. Failed, truncated and malformed outputs remain in the dataset. No best-of-N selection or failure-based regeneration is performed. The English prose and verbatim original-language text requirements come from the official prompts; the corpus does not certify that every output follows them. If the generated table cannot be parsed, the official helper can append the locked input table after the malformed output. The result can have a parseable final table while retaining broken table text in its paragraph. The additional raw_contract_checks and processed_contract_checks flag table syntax in prose, unmatched element references, missing citations and repeated citations. These are mechanical indicators, not a complete grammar or semantic validator.

Additional columns include mode, input_id, category, length_bin, source, input_boxes, box_origin, reference_images, postprocess_error, pilot, input_word_count, output_paragraph_word_count and output_total_word_count. The total word count includes JSON syntax and descriptions in structured outputs; the paragraph count excludes a parseable final table. id includes the mode; input_id links matched editing observations. Editing identities include source image ID and normalized text, so equal instructions on different images remain distinct. Images can be retrieved from source.dataset, source.asset_revision, source.file, source.row_index and source.image_column. The reference metadata records original/prepared hashes, dimensions, byte counts and preprocessing policy. The target edited image is never supplied to PUP.

Mechanical flags check nonempty text, truncation, simple periodic tails, box format, cited IDs, locked coordinates and inferred quoted literals. They do not establish semantic correctness, localization accuracy, English-only prose or image quality. Quoted-literal checks only inspect double-quoted substrings and can produce false positives. Full counts are in audit-modes.json.

Mode	Nonempty	Token-limit finishes	Invalid raw box tables	Median input words	Median output paragraph words
i2i	6,000	3	n/a	6.0	183.0
i2i_bbox	3,000	4	36	6.0	61.0
t2i_bbox	3,000	0	59	8.0	54.0

No downstream images were generated or evaluated. A small semantic pilot checks whether the invocation works; it is not a human preference evaluation. Known pilot errors remain: the Chinese guitar edit both removes a bottle and places it in the subject's hand. A pan-resizing instruction requests one-fifth dimensions, while its proposed box spans are approximately 14% and 17% of the source spans. Valid JSON and preserved input boxes do not imply a consistent edit.

The source terms apply separately from the Apache-2.0 compilation license. MagicBrush and AnyEdit texts retain CC BY 4.0. NSR-1K retains its MIT repository terms and CC BY 4.0 for COCO-derived annotations. OmniEdit and GEdit-Bench declare MIT in their pinned dataset cards. GEdit-Bench supplies no separate license file; AnyEdit's prose license field is an unfilled template, while its YAML declares CC BY 4.0. These source declarations and their limitations are retained here. See NOTICE, OmniEdit's MIT license, LayoutGPT's MIT license, and the MIT permission terms for attribution and license text.

The extension rental cost was $1.294 (provider charges observed after verified deletion, including GPU, storage and traffic). The rental was deleted after local checksum verification. Timings include 1906.4 seconds of timed generation; preparation and model startup are separate.