r2s-full-v1-940f5d3-1974cf82-47646604
completed —
Hypothesis
Trained on the 2,681 family-disjoint training icons, the 8.9M-parameter render-to-program model transcribes held-out validation renders better than retrieving the closest training icon. Pass requires all three: (1) mean 72 px pixel error below the nearest-training-icon baseline with the paired 95% interval on the reduction excluding zero; (2) CLIP top-1 on Gate N's 32 validation icons at or above OmniSVG 4B zero-shot, 0.609; (3) median single-icon latency on the RTX 4080 under 500 ms.
Visual output
Written result
Hypothesis. Trained on the 2,681 family-disjoint training icons, the 8.9M-parameter render-to-program model transcribes held-out validation renders better than retrieving the closest training icon. Pass requires all three: (1) mean 72 px pixel error below the nearest-training-icon baseline with the paired 95% interval on the reduction excluding zero; (2) CLIP top-1 on Gate N's 32 validation icons at or above OmniSVG 4B zero-shot, 0.609; (3) median single-icon latency on the RTX 4080 under 500 ms.
Result
| measure | value |
|---|---|
| icons evaluated | 339 (primary/validation) |
| model pixel error, mean [95% CI] | 0.1487 [0.1412, 0.1561] |
| model pixel error, median | 0.1565 |
| exact program rate | 0.003 |
| rendered rate | 1.000 |
| blank canvas pixel error | 0.1720 |
| nearest training icon pixel error | 0.0897 [0.0848, 0.0947] |
| error reduction vs nearest icon | -0.0590 [-0.0638, -0.0540] |
| icons where model beats nearest icon | 16 |
| CLIP top-1, model greedy (32 Gate N icons) | 0.219 |
| CLIP top-1, nearest training icon | 0.344 |
| CLIP top-1, OmniSVG 4B zero-shot (Gate N) | 0.609 |
Resources
| measure | value |
|---|---|
| device | NVIDIA GeForce RTX 4080 |
| parameters | 8,921,634 |
| train seconds | 1655 |
| peak VRAM GiB | 4.089363098144531 |
| latency ms/icon, batch 1, median | 629.4 |
| latency ms/icon, batch 1, p95 | 655.0 |
| throughput ms/icon, batch 64 | 47.6 |
| model calls per icon (decoding) | 500.0 |
| torch / CUDA | 2.14.0a0+4fdf77b940.nv26.08 / 13.4 |
Latency covers encoding and decoding to a validated token program; it excludes rasterising the SVG. Samples: samples.png, rows are reference, model, nearest training icon.
State transitions
- running2026-09-27T19:57:45Z
- completed2026-09-27T20:26:13Z
Run record
Verbatim from runs/r2s-full-v1-940f5d3-1974cf82-47646604/run.yaml, the record committed before launch.
schema_version
2
run_id
r2s-full-v1-940f5d3-1974cf82-47646604
state
completed
planned_at
2026-09-27T19:57:45Z
hypothesis
Trained on the 2,681 family-disjoint training icons, the 8.9M-parameter render-to-program model transcribes held-out validation renders better than retrieving the closest training icon. Pass requires all three: (1) mean 72 px pixel error below the nearest-training-icon baseline with the paired 95% interval on the reduction excluding zero; (2) CLIP top-1 on Gate N's 32 validation icons at or above OmniSVG 4B zero-shot, 0.609; (3) median single-icon latency on the RTX 4080 under 500 ms.
parent_run
r2s-overfit4
git_commit
940f5d39cfd21468836064e15eca5de3f8831f73
dirty_patch_sha256
ce4f9416fc926524…config
configs/render2svg/full-v1.yaml
config_sha256
1974cf82ca22aa14…config_resolved
model
image_size
144
d_model
256
heads
8
encoder_layers
2
decoder_layers
6
feedforward
1024
dropout
0.1
training
steps
20000
batch_size
32
learning_rate
0.0005
weight_decay
0.01
warmup_steps
500
seed
7001
eval_every
1000
eval_icons
64
train_icons
None
evaluate_on_train
false
augment_mirror
0.0
bf16
true
final_eval_icons
339
clip_icons
32
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
8921634
command
.venv/bin/python -m mojidiff.learning.render2svg --config configs/render2svg/full-v1.yaml
outputs
run_dir
runs/r2s-full-v1-940f5d3-1974cf82-47646604
checkpoints
/home/dev/.cache/mojidiff/runs/r2s-full-v1-940f5d3-1974cf82-47646604
notes
completed_at
2026-09-27T20:26:13Z
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
4.089363098144531
train_seconds
1654.9081625590334
inference_ms_per_icon
629.4160155230202
inference_p95_ms_per_icon
655.0014589447528
icons_per_second
1.5887743167276083
batched_ms_per_icon
47.61035209412512
batched_icons_per_call
64
excludes
rendering the output SVG to pixels
baselines
blank_canvas_pixel_error
0.1720480552069557
nearest_training_icon_pixel_error
0.08973124639576185
gate_n_omnisvg_zero_shot_top1
0.609375
nearest_training_icon_clip_top1
0.34375
result
icons
339
model_pixel_error
0.14872044521621922, 0.14124088256183873, 0.15610677177139445
model_pixel_error_median
0.15654301643371582
exact_program_rate
0.0029498525073746312
rendered_rate
1.0
blank_pixel_error
0.1720480552069557
selected_step
19000
nearest_training_icon_pixel_error
0.08973124639576185, 0.0848205749792233, 0.09466393787711656
model_minus_baseline_error_reduction
-0.058989198820457375, -0.06381041379237455, -0.053992881722259606
icons_model_beats_baseline
16
clip_retrieval
model_greedy
top1_rate
0.21875
top5_rate
0.34375
mean_rank
12.75
clip_to_reference_mean
0.8145000487565994
chance_top1
0.03125
icons
32
nearest_training_icon
top1_rate
0.34375
top5_rate
0.59375
mean_rank
7.40625
clip_to_reference_mean
0.8630629815161228
chance_top1
0.03125
icons
32
gate_n_omnisvg_zero_shot_top1
0.609375
gate_n_omnisvg_best_of_12_top1
0.828125
model_calls_per_icon
500