latent-v3-blind-e254c16-9c7db490-47646604
completed —
Hypothesis
Blind to earlier coordinates, the decoder must take geometry from the latent. Pass, against latent-v2 on the 339 validation icons: (1) at most 15% of reconstructed paths span under half a unit (latent-v2 57%); (2) reconstruction pixel error improves on latent-v2's, paired, interval excluding zero, to 0.140 or less; (3) prior samples all distinct. The earlier criteria are kept below and still reported: A variational autoencoder over codec programs, trained on the 2,681 training icons and their compositions, keeps information in its 64-dimensional latent. Pass requires (1) reconstruction pixel error on the 339 validation icons below the error of the single prior-mean decode, paired, with the 95% interval on the reduction excluding zero; (2) at least 29 of 32 prior samples distinct. Reported without a criterion: reconstruction against the nearest training icon, interpolation and prior-sample sheets, and decode latency.
Visual output
Written result
Hypothesis. Blind to earlier coordinates, the decoder must take geometry from the latent. Pass, against latent-v2 on the 339 validation icons: (1) at most 15% of reconstructed paths span under half a unit (latent-v2 57%); (2) reconstruction pixel error improves on latent-v2's, paired, interval excluding zero, to 0.140 or less; (3) prior samples all distinct. The earlier criteria are kept below and still reported: A variational autoencoder over codec programs, trained on the 2,681 training icons and their compositions, keeps information in its 64-dimensional latent. Pass requires (1) reconstruction pixel error on the 339 validation icons below the error of the single prior-mean decode, paired, with the 95% interval on the reduction excluding zero; (2) at least 29 of 32 prior samples distinct. Reported without a criterion: reconstruction against the nearest training icon, interpolation and prior-sample sheets, and decode latency.
| measure | value |
|---|---|
| validation icons | 339 |
| reconstruction pixel error | 0.1733 [0.1667, 0.1802] |
| nearest training icon pixel error | 0.0897 |
| prior-mean decode pixel error (control) | 0.1932 |
| error reduction, own latent vs prior mean | 0.0198 [0.0156, 0.0242] |
| exact reconstructions | 0.000 |
| prior samples distinct / rendered | 32 / 32 of 32 |
| prior samples, median pixel distance to nearest training icon | 0.0958 |
| parameters | 10,114,082 |
| train seconds | 3423 |
| decode ms per icon (graph, batch 1) | 579.3 |
Sheets: reconstructions.png, interpolations.png, prior-samples.png.
State transitions
- running2026-09-29T12:22:11Z
- completed2026-09-29T13:20:52Z
Run record
Verbatim from runs/latent-v3-blind-e254c16-9c7db490-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
latent-v3-blind-e254c16-9c7db490-47646604
state
completed
planned_at
2026-09-29T12:22:11Z
hypothesis
Blind to earlier coordinates, the decoder must take geometry from the latent. Pass, against latent-v2 on the 339 validation icons: (1) at most 15% of reconstructed paths span under half a unit (latent-v2 57%); (2) reconstruction pixel error improves on latent-v2's, paired, interval excluding zero, to 0.140 or less; (3) prior samples all distinct. The earlier criteria are kept below and still reported: A variational autoencoder over codec programs, trained on the 2,681 training icons and their compositions, keeps information in its 64-dimensional latent. Pass requires (1) reconstruction pixel error on the 339 validation icons below the error of the single prior-mean decode, paired, with the 95% interval on the reduction excluding zero; (2) at least 29 of 32 prior samples distinct. Reported without a criterion: reconstruction against the nearest training icon, interpolation and prior-sample sheets, and decode latency.
parent_run
latent-v2-kl
git_commit
e254c16811d51ba88ae6d38b2d82f7161bfb1477
dirty_patch_sha256
config_sha256
9c7db490a2fe6b32…config
configs/latent/latent-v3-blind.yaml
config_resolved
model
image_size
144
d_model
256
heads
8
encoder_layers
2
decoder_layers
6
feedforward
1024
dropout
0.1
metric
true
fourier
10
order
path
latent
latent_dim
64
memory_tokens
16
encoder_layers
3
beta
0.1
beta_warmup_fraction
0.3
free_bits
0.5
input_dropout
0.25
blind_coordinates
true
training
steps
30000
batch_size
32
learning_rate
0.0005
weight_decay
0.01
warmup_steps
1000
seed
7001
eval_every
2000
eval_icons
64
train_icons
None
evaluate_on_train
false
augment_variants
0
augment_seed
9001
augment_mirror
0.5
augment_max_shift
48
augment_colour
0.5
augment_online
true
compose_probability
0.5
compose_parts
4
augment_original
0.1
loader_workers
12
extra_training
bf16
true
dataset
pilot_config
configs/learning/openmoji-g1-geometric-gate-v16.yaml
cache_sha256
476466042da98d72…train_icons
2681
evaluated_on
primary/validation
seed
7001
determinism
seeded; cuDNN and SDPA kernels not forced deterministic
model_parameters
10114082
command
.venv/bin/python -m mojidiff.learning.latent --config configs/latent/latent-v3-blind.yaml
outputs
run_dir
runs/latent-v3-blind-e254c16-9c7db490-47646604
checkpoints
/home/dev/.cache/mojidiff/runs/latent-v3-blind-e254c16-9c7db490-47646604
completed_at
2026-09-29T13:20:52Z
resource
device
NVIDIA GeForce RTX 4080
torch_version
2.14.0a0+4fdf77b940.nv26.08
cuda_version
13.4
python_version
3.12.3
peak_vram_gib
5.32373571395874
train_seconds
3422.7700097779743
inference_ms_per_icon
579.3265455286019
inference_p95_ms_per_icon
579.9213430145755
icons_per_second
1.726142203767925
decoder
fast_decode.GraphDecoder, float32, batch 1, from a latent
decoder_calls
459
excludes
rasterising the output SVG
note
measured while other GPU work may be running; see gpu load at the time
baselines
nearest_training_icon_pixel_error
0.08973124639576185
note
retrieval stores every training icon; the latent stores 64 numbers
result
icons
339
reconstruction_pixel_error
0.17333034218280716, 0.1666632787217345, 0.18015682739439176
nearest_training_icon_pixel_error
0.08973124639576185, 0.0848205749792233, 0.09466393787711656
prior_mean_decode_pixel_error
0.19317802360451677, 0.1890661415404978, 0.19739498853639517
latent_minus_prior_mean_error_reduction
0.019847681421709625, 0.015591162248654703, 0.024176434387389714
exact_reconstruction_rate
0.0
prior_samples
count
32
distinct
32
rendered
32
median_pixel_distance_to_nearest_training_icon
0.09575679898262024
selected_step
10000