Problem
On slower machines, the "Clean Image" step is the heaviest bottleneck during batch
processing.
Here's each step in the translation pipeline:
| Step |
Cloud API |
Local Load |
| Detect Text |
✗ |
Medium |
| Recognize Text |
✓ (MS, Google, GPT) |
Heavy |
| Get Translations |
✓ (all API) |
None |
| Segment Text |
- |
Low |
| Clean Image |
✗ |
Very Heavy |
| Render |
- |
Low |
OCR and Translation already have cloud options, but Inpainting is local-only.
Suggestion
Add Replicate API as an inpainting backend option. LaMa model on Replicate runs on NVIDIA
L40S at ~$0.001/sec.
This would let users with slow PCs offload the heaviest step to cloud.
Problem
On slower machines, the "Clean Image" step is the heaviest bottleneck during batch
processing.
Here's each step in the translation pipeline:
OCR and Translation already have cloud options, but Inpainting is local-only.
Suggestion
Add Replicate API as an inpainting backend option. LaMa model on Replicate runs on NVIDIA
L40S at ~$0.001/sec.
This would let users with slow PCs offload the heaviest step to cloud.