xkcd-artist (gnumanth/xkcd-artist)

Hybrid Fine-Tuned Diffusion UNet LoRA (sd_xkcd_lora) + 5-Head ExplainXKCD Scene, Pose & Speech-Bubble Planner (XkcdArtistForGeneration), trained on GCP NVIDIA L4 GPU (24 GB VRAM) over 891 ExplainXKCD stage directions and 226 speech-bubble-masked real XKCD comic panels.

• GitHub Repository: https://github.com/hemanth/xkcd-artist
• Live Interactive Studio: https://hemanth.github.io/xkcd-artist/

Model Contents

File	Size	Description
model.safetensors	6.16 MB	XkcdArtistForGeneration FP32 weights (1,586,603 params)
lora_weights.safetensors	376 KB	Internal LoRA adapter weights (93,568 params)
onnx/model_quantized.onnx	1.72 MB	Dynamic Int8 quantized ONNX Runtime graph (3.58x compression)
sd_xkcd_lora/adapter_model.safetensors	6.60 MB	Stable Diffusion (segmind/tiny-sd) UNet LoRA (r=16, 1,726,464 trainable params) trained on speech-bubble-masked XKCD panels
sd_xkcd_lora/adapter_config.json	1 KB	PEFT / Diffusers LoRA config (attn1/attn2 to_q, to_k, to_v, to_out.0)
modeling_xkcd_artist.py	6 KB	Self-contained PyTorch model definition (PreTrainedModel)

Empirical Benchmark (B = 1,000 Bootstrap on Held-Out XKCD + ExplainXKCD)

Runtime	Artifact Size	Layout Acc (%)	Stage Prop Acc (%)	Kinematic Pose MAE	Bubble Clearance (%)	Abstention AUROC	Latency p50
PyTorch FP32 + Diffusion LoRA	6.16 MB (1.59M)	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]	0.0051	100.0%	1.0	0.629 ms
Browser ONNX Int8 (WASM SIMD)	1.72 MB (3.58x smaller)	100.0% [100.0, 100.0]	100.0% [100.0, 100.0]	0.0053	100.0%	1.0	0.621 ms

Usage (PyTorch / ONNX)

from xkcd_artist import xkcd_artist

comic = xkcd_artist(
    "Machine learning pile of linear algebra stir the pile",
    backend="onnx",  # 'onnx' | 'pytorch'
)

svg_markup = comic.to_svg()
png_bytes = comic.to_png("machine_learning_xkcd.png")

License

MIT © Hemanth HM