---
license: other
license_name: per-file
license_link: LICENSE
tags:
  - onnx
  - manga
  - webgpu
---

# Comic text ONNX models, half precision

Float16 conversions of three ONNX models for manga and comic text
processing, aimed at onnxruntime-web on the WebGPU execution provider. The weights are unchanged
apart from the precision cast; no retraining, pruning, or architecture edit.

These exist because WebGPU rewards half precision and, in one case, punishes
integer quantisation badly.

## Files

| file | converted from | licence |
|---|---|---|
| `detector-fp16.onnx` | [ogkalu/comic-text-and-bubble-detector](https://huggingface.co/ogkalu/comic-text-and-bubble-detector) `detector.onnx` | Apache-2.0 |
| `comic-text-detector-fp16.onnx` | [mayocream/comic-text-detector-onnx](https://huggingface.co/mayocream/comic-text-detector-onnx) `comic-text-detector.onnx` | GPL-3.0 |
| `aot-fp16.onnx` | [lemondouble/lemon-manga-translator](https://huggingface.co/lemondouble/lemon-manga-translator) `onnx/aot-inpainting/aot_folded.onnx` | GPL-3.0 |

Each file keeps the licence of the model it was converted from. They are
distributed together as an aggregate, which GPL-3.0 section 5 permits; the
GPL-3.0 files are not combined with the Apache-2.0 one.

## Measured

Chrome, Apple M3, onnxruntime-web on WebGPU, median of repeated runs at the
input size these are run at. "Shipping" is the artifact each was converted from,
except for detection, where it is the int8 build that was in use before.

| model | shipping | this file | agreement with float32 |
|---|---|---|---|
| detector | 1610 ms (int8) | **140 ms** | same 10 regions on a real page, worst box IoU 0.969 |
| comic-text-detector | 468 ms | **290 ms** | 0.006% of mask pixels land on the other side of the threshold |
| aot | 310 ms | **237 ms** | mean absolute difference 0.6/255, worst 15/255 |

The detector number is not a half precision result so much as an integer one.
WebGPU has no kernels for `ConvInteger` or `MatMulInteger`, so every quantised
node in the int8 build crosses back to the CPU. Its float32 build already runs
in 188 ms; half precision takes it to 140 ms.

A LaMa conversion is deliberately absent. It overflows in the fast Fourier
convolution branch and returns NaN, for a 16% gain that would not have been
worth it anyway.

## How they were made

`onnxconverter-common`'s `convert_float_to_float16` with `keep_io_types=True`,
followed by a repair pass. The converter retypes tensors but leaves each `Cast`
node's `to` attribute on the original dtype, so the graph declares float32 where
it now produces float16 and the session refuses to build. Setting `to` from the
inferred output type fixes 5 casts in the detector and 10 in AOT.

Inputs and outputs stay float32, so these are drop-in replacements: no caller
change beyond the file name.