Dataset Viewer
Auto-converted to Parquet Duplicate
input
imagewidth (px)
600
6.1k
layer
images listlengths
2
2
run_id
stringclasses
6 values
prompt
stringclasses
1 value
auto_captioned
bool
1 class
negative_prompt
stringclasses
1 value
seed
int64
585M
1.65B
⌀
steps
int64
50
50
⌀
layer_count
int64
2
2
guidance_scale
int64
4
4
⌀
resolution
int64
640
1.02k
⌀
cfg_normalize
bool
1 class
use_en_prompt
bool
1 class
model_id
stringclasses
1 value
seconds
float64
49.4
163
⌀
timestamp
timestamp[s]date
2026-08-31 12:31:32
2026-11-10 22:04:25
⌀
a504ea39fc77
true
1,646,854,650
50
2
4
640
true
true
Qwen/Qwen-Image-Layered
49.9
2026-08-31T12:31:32
08b160d9aa37
true
584,782,526
50
2
4
1,024
true
true
Qwen/Qwen-Image-Layered
162.8
2026-08-31T13:01:41
76ead06bf077
null
null
null
null
null
2
null
null
null
null
null
null
null
5bea245490f5
true
1,589,667,785
50
2
4
640
true
true
Qwen/Qwen-Image-Layered
50.2
2026-11-10T22:04:25
f024be761689
null
null
null
null
null
2
null
null
null
null
null
null
null
43ed89378e26
true
1,083,959,147
50
2
4
640
true
true
Qwen/Qwen-Image-Layered
49.4
2026-08-31T14:40:44

Mitosis Generations

Every decomposition run of the Mitosis Space is logged here: the input image, the RGBA layers produced from it, and the settings used.

⚠️ Public log

The Space is public and every run is recorded in this public dataset, including images uploaded by visitors. Do not submit anything that cannot be shared publicly.

Layout

data/
  metadata.jsonl          # one JSON object per run
  <run_id>/
    input.png             # exactly what the model saw (RGB(A) normalized, transparency on white)
    layer_1.png .. N.png  # the RGBA layers produced

Each metadata.jsonl line:

Field Meaning
run_id Folder holding this run's images
steps Denoising steps
input_file_name / layer_file_names Paths relative to data/; the viewer turns these into the input and layer image columns
prompt, negative_prompt Text the user typed; prompt is empty when the model captioned the image itself
auto_captioned Whether the prompt was generated by the model
seed, layer_count, guidance_scale, resolution Generation settings (resolution is a target area: 640 or 1024)
cfg_normalize, use_en_prompt Toggles from Advanced Settings
model_id Model that produced the layers
seconds Pipeline runtime
timestamp UTC ISO-8601

Loading

from datasets import load_dataset

ds = load_dataset("Elfsong/mitosis-generations", split="train")
ds[0]["input"]   # PIL image that was decomposed
ds[0]["layer"]   # list of PIL images, one per layer

Layers are produced by Qwen-Image-Layered.

Downloads last month
75