winrisef/gpt-generated-image-collection

SeedVR2 batch-0006 4K

163/200 images uploaded. Long edge: 3840 pixels; aspect ratio preserved. batch-0006 uses SeedVR2 7B mixed FP8; the existing batch 0001–0003 4K derivatives use FiDeSR. CPU offload is enabled (24 DiT blocks, VAE and tensors). VAE tiles are 512/64. Whole-image inference is attempted first; failures retry with 640-pixel input tiles, at least 128-pixel overlap, 4x inference, overlap blending and Lanczos resize. Per-image metadata records the actual pipeline and hashes. Original HF images are retained.

Historical FiDeSR descriptions and tables below refer to batches 0001–0003; this section records the current SeedVR2 batch.

SeedVR2 batch-0005 4K

200/200 images uploaded. Long edge: 3840 pixels; aspect ratio preserved. batch-0005 uses SeedVR2 7B mixed FP8; the existing batch 0001–0003 4K derivatives use FiDeSR. CPU offload is enabled (24 DiT blocks, VAE and tensors). VAE tiles are 512/64. Whole-image inference is attempted first; failures retry with 640-pixel input tiles, at least 128-pixel overlap, 4x inference, overlap blending and Lanczos resize. Per-image metadata records the actual pipeline and hashes. Original HF images are retained.

Historical FiDeSR descriptions and tables below refer to batches 0001–0003; this section records the current SeedVR2 batch.

SeedVR2 batch-0004 4K

200/200 images uploaded. Long edge: 3840 pixels; aspect ratio preserved. Batch 0004 uses SeedVR2 7B mixed FP8; the existing batch 0001–0003 4K derivatives use FiDeSR. CPU offload is enabled (24 DiT blocks, VAE and tensors). VAE tiles are 512/64. Whole-image inference is attempted first; failures retry with 640-pixel input tiles, at least 128-pixel overlap, 4x inference, overlap blending and Lanczos resize. Per-image metadata records the actual pipeline and hashes. Original HF images are retained.

Historical FiDeSR descriptions and tables below refer to batches 0001–0003; this section records the current SeedVR2 batch.

GPT-Generated Image Collection

The 4K images are FiDeSR (batches 0001–0003) or SeedVR2 (batch-0004, batch-0005, batch-0006) super-resolution derivatives, not native 4K originals. See model versions and processing pipeline.

1216 original PNG images and 1179 super-resolution derivatives (2395 PNG files). Exact generation-model versions are unverified and remain null. 中文说明.

Batch	Kind	Images	GiB
batch-0001	original	216	0.539
batch-0002	original	200	0.432
batch-0001-4K	upscaled	216	3.104
batch-0002-4K	upscaled	200	2.236
batch-0003	original	200	0.439
batch-0003-4K	upscaled	200	2.594
batch-0004	original	200	0.583
batch-0005	original	200	0.566
batch-0006	original	200	0.551
batch-0004-4K	upscaled	200	3.317
batch-0005-4K	upscaled	200	2.938
batch-0006-4K	upscaled	163	2.273

images/
  batch-0001/  (216 PNGs)
  batch-0002/  (200 PNGs)
  batch-0001-4K/  (216 PNGs)
  batch-0002-4K/  (200 PNGs)
  batch-0003/  (200 PNGs)
  batch-0003-4K/  (200 PNGs)
  batch-0004/  (200 PNGs)
  batch-0005/  (200 PNGs)
  batch-0006/  (200 PNGs)
  batch-0004-4K/  (200 PNGs)
  batch-0005-4K/  (200 PNGs)
  batch-0006-4K/  (163 PNGs)
metadata/
  batch-0001.jsonl
  batch-0001-prompts-original.json
  batch-0002.jsonl
  batch-0002-prompts-original.json
  batch-0001-4K.jsonl
  batch-0002-4K.jsonl
  batch-0003.jsonl
  batch-0003-prompts-original.json
  batch-0003-4K.jsonl
  batch-0004.jsonl
  batch-0004-prompts-original.json
  batch-0005.jsonl
  batch-0005-prompts-original.json
  batch-0006.jsonl
  batch-0006-prompts-original.json
  batch-0004-4K.jsonl
  batch-0005-4K.jsonl
  batch-0006-4K.jsonl
checksums/
  batch-0001.sha256
  batch-0002.sha256
  batch-0001-4K.sha256
  batch-0002-4K.sha256
  batch-0003.sha256
  batch-0003-4K.sha256
  batch-0004.sha256
  batch-0005.sha256
  batch-0006.sha256
  batch-0004-4K.sha256
  batch-0005-4K.sha256
  batch-0006-4K.sha256
catalog.json
README.md
README.zh-CN.md
UPSCALE.md

Index and provenance

The default dataset configuration loads original images; upscaled loads the derivatives. Each JSONL has one record per image, including image path, prompt/status, width, height, bytes and SHA-256. Upscaled records additionally identify source_image and source_sha256, original dimensions, processing_model, model_native_scale and effective x/y scale.

FiDeSR uses a native 4× model; these delivered files were resized to a 3840-pixel long edge. The -4K suffix does not mean exactly 3840×2160 or four times the source dimensions. They are generated derivatives, not ground-truth high-resolution originals; hallucination and local detail changes are possible. Original/derivative pairs must stay in the same split to prevent evaluation leakage.

1200 originals have recorded prompts; 16 have no recorded prompt (all in batch-0001). Missing prompts remain null. Prompts for numbered images 1–1200 are complete. Only batches 0001–0003 have corresponding FiDeSR derivatives; batch-0004 SeedVR2 derivatives are being added; batch-0005 contains originals; see catalog.json for exact counts.

Previously published images remain unchanged. Future releases may append new filenames and their index entries without replacing existing images. SHA-256 lists cover every PNG. Local scripts, runtime logs, lock files and credentials are excluded from publication. The license metadata is retained from the existing repository. This is a contributor-provided collection, not an independently verified image-quality benchmark.