- Aug 17, 2016
- 583
- 3,026
With v3.029, the SEP picture translation engine now also has limited support for translating vertical text.
I say limited because you'll have to tilt your head 90° to read it.
Changelog since last changelog post:
(I noticed I didn't actually make one since v3)
3.029
SEP
Added support for vertical text.
Improved background fill further.
3.028
SEP
changed the solid background fill to more closely fit the detected text boxes.
3.027
Made project saving a lot more robust. It already made backups, but now it will make an additional verification.
SEP
Added options to allow replacing text with itself if someone wants to do that for whatever reason.
3.026
It will now convert file paths to UTF-8 before spawning ruby, fixing Windows system locale issues.
3.025
Improved SEP handling of solid backgrounds and added a font color detection option.
3.024
Added new logic to send scenes with only a single text box in their own bigger batches.
Fixed config.json, again. This time it should actually properly load everything even if it was only generated this session.
3.023
Fixed an issue that caused wrong tagging if the tool is launched without an already generated config.json
3.022
It will now only close the UI if all processes are finished.
ETA system now works for multiple requests at the same time.
3.021
Added compression to log files so they don't take up unreasonable amounts of space.
Added code comments and readme files to make it more obvious to someone besides me how the heck this project is structured.
3.020
It will now translate large scenes first and grouped scenes will actually be grouped to their limits.
The Batch Translator will now try to estimate how long the batch translation will take.
This is only a very rough estimate and will not work if multiple batches are sent at the same time.
Removed token summaries for every individual batch from the UI (They are still in the console), and instead added a final token summary for the entire batch translation.
3.019
More adjustments to the validation system.
3.018
Improved Batch Validation.
Fixed a pointless error showing up in the console.
3.017
Changed DSLR limit options and how batch translation works.
Fixed the options menu process sometimes not being correctly killed and just running in the backround endlessly.
3.016
Fixed new option merge breaking the Batch Translation button.
Improved Glossary LLM request system.
Removed some dead bloat from the Linux version.
3.015
Merged the Project Summary system into DSLR Glossary because it doesn't make sense to have 2 editors for the same context.
Changed default translator to DSLR.
Fixed some problems with the options menu.
3.014
Improved Project Summary to collect less useless information.
3.013
Project Summary: auto-generated at creation, view/edit button, persistence in .slrproj, context injection with unified budget
Unified context budget: summary/scene/glossary share one maxContextLength budget, trimmed by priority
Scene Graph panel: draggable, behind menus (z-8500), hidden during loading
SLR collapsible sections: fixed
GlossaryExtractor _getDslrEngine: fixed (same registry bug)
3.012
General cleanup of obsolete code.
3.011
DSLR
Added better reasoning support.
It will now show token progress of requests.
3.010
Fixed DSLR option presets and null parameters.
Fixed SEP jpg/jpeg support and added bmp, webp, tga and pcx support.
Improved handling of transparent images.
Fixed min font size option.
3.009
Added Scene Graph system.
It will now trace scenes across the files and build context for them.
Batches sent to a translator will now only include full scenes unless they cannot fit in given restraints.
It will automatically add a scene context block to those scenes.
You can limit how much context is added based on the context length on your model. If you do not set the option it will always send the full context.
Fixed some minor bugs and removed dead code.
3.008
Added new Raw window to SEP that shows where the text was collected from the original image and OCR confidence.
Added override commands to set text erasure method for specific rows.
Removed pointless tts function.
3.007
Improved Linux distro compatibility.
Added small documentation for SEP.
DSLR
Made "Response is too long" and "Response is too short" reject reasons more lenient.
3.006
Added aot-inpainting and LaMa options for text erasure to SEP.
Removed obsolete Blank project option.
3.005
Included first prototype of the SEP ONNX remake.
Fixed broken project creation in the windows version.
3.004
Fixed romaji and dictionary features.
Fixed some wrong colors.
Changed Raw tab.
3.003
Fixed non Japanese cell tagging.
3.002
Fixed a bug preventing glossary creation.
3.001
Added option to import translations from other slrpoj files.
Removed broken spreadsheet import option.
fixed a bug in the bracket preservation system.
Cleaned up a lot of regex.
3.000
Removed or refactored all leftover T++ code.
I changed the UI backend and rendering. Objects are now pre-rendered and cached, drastically improving performance.
De-duplication is now completely removed and as a result there is no SLRContextID anymore.
The parsers now collect string numbers and integers in new "num" tag rows. Allowing for easy font changes in MV/MZ.
The Prepare Project for Batch translation button was removed and is instead ran automatically on project creation.
The Fix Cells and Check for Errors button was removed it will instead now run it immediately during translation and only on the cells that were actually translated. This now also means it works on destinations other than [1].
Error codes will now not be put into the translation, but instead copied with the codes into a new Error[4] column.
Overwrite is now [5].
Collected plugins will no longer include strings that are modified by the plugins.js, since changing them would just break stuff for no reason.
It is now possible to insert ruby code at the start of Main in VX and VXAce Scripts to insert stuff like Font.default_size values.
Fixed OOM crashes for exporting extremely large VX/VXAce games.
Fixed inefficient parsing logic, VX/VXAce parsing is now 17x faster.
Implemented automatic logs of the last 3 sessions (saved to data>logs) that uncollapse all console log entries. So that you can read the console log even if the tool crashed so badly that it just closed itself. (Can be turned off in options)
Rebuilt the options menu and removed obsolete or useless options.
Added an automatic spellchecker. (Can be turned off.)
Replaced trans format with compressed much more complete slrproj.
Made project creation for certain titles more convenient.
Linux version is no longer an afterthought, it has ALL features of the windows version.
Removed a lot of obsolete code and files.
I say limited because you'll have to tilt your head 90° to read it.
Changelog since last changelog post:
(I noticed I didn't actually make one since v3)
3.029
SEP
Added support for vertical text.
Improved background fill further.
3.028
SEP
changed the solid background fill to more closely fit the detected text boxes.
3.027
Made project saving a lot more robust. It already made backups, but now it will make an additional verification.
SEP
Added options to allow replacing text with itself if someone wants to do that for whatever reason.
3.026
It will now convert file paths to UTF-8 before spawning ruby, fixing Windows system locale issues.
3.025
Improved SEP handling of solid backgrounds and added a font color detection option.
3.024
Added new logic to send scenes with only a single text box in their own bigger batches.
Fixed config.json, again. This time it should actually properly load everything even if it was only generated this session.
3.023
Fixed an issue that caused wrong tagging if the tool is launched without an already generated config.json
3.022
It will now only close the UI if all processes are finished.
ETA system now works for multiple requests at the same time.
3.021
Added compression to log files so they don't take up unreasonable amounts of space.
Added code comments and readme files to make it more obvious to someone besides me how the heck this project is structured.
3.020
It will now translate large scenes first and grouped scenes will actually be grouped to their limits.
The Batch Translator will now try to estimate how long the batch translation will take.
This is only a very rough estimate and will not work if multiple batches are sent at the same time.
Removed token summaries for every individual batch from the UI (They are still in the console), and instead added a final token summary for the entire batch translation.
3.019
More adjustments to the validation system.
3.018
Improved Batch Validation.
Fixed a pointless error showing up in the console.
3.017
Changed DSLR limit options and how batch translation works.
Fixed the options menu process sometimes not being correctly killed and just running in the backround endlessly.
3.016
Fixed new option merge breaking the Batch Translation button.
Improved Glossary LLM request system.
Removed some dead bloat from the Linux version.
3.015
Merged the Project Summary system into DSLR Glossary because it doesn't make sense to have 2 editors for the same context.
Changed default translator to DSLR.
Fixed some problems with the options menu.
3.014
Improved Project Summary to collect less useless information.
3.013
Project Summary: auto-generated at creation, view/edit button, persistence in .slrproj, context injection with unified budget
Unified context budget: summary/scene/glossary share one maxContextLength budget, trimmed by priority
Scene Graph panel: draggable, behind menus (z-8500), hidden during loading
SLR collapsible sections: fixed
GlossaryExtractor _getDslrEngine: fixed (same registry bug)
3.012
General cleanup of obsolete code.
3.011
DSLR
Added better reasoning support.
It will now show token progress of requests.
3.010
Fixed DSLR option presets and null parameters.
Fixed SEP jpg/jpeg support and added bmp, webp, tga and pcx support.
Improved handling of transparent images.
Fixed min font size option.
3.009
Added Scene Graph system.
It will now trace scenes across the files and build context for them.
Batches sent to a translator will now only include full scenes unless they cannot fit in given restraints.
It will automatically add a scene context block to those scenes.
You can limit how much context is added based on the context length on your model. If you do not set the option it will always send the full context.
Fixed some minor bugs and removed dead code.
3.008
Added new Raw window to SEP that shows where the text was collected from the original image and OCR confidence.
Added override commands to set text erasure method for specific rows.
Removed pointless tts function.
3.007
Improved Linux distro compatibility.
Added small documentation for SEP.
DSLR
Made "Response is too long" and "Response is too short" reject reasons more lenient.
3.006
Added aot-inpainting and LaMa options for text erasure to SEP.
Removed obsolete Blank project option.
3.005
Included first prototype of the SEP ONNX remake.
Fixed broken project creation in the windows version.
3.004
Fixed romaji and dictionary features.
Fixed some wrong colors.
Changed Raw tab.
3.003
Fixed non Japanese cell tagging.
3.002
Fixed a bug preventing glossary creation.
3.001
Added option to import translations from other slrpoj files.
Removed broken spreadsheet import option.
fixed a bug in the bracket preservation system.
Cleaned up a lot of regex.
3.000
Removed or refactored all leftover T++ code.
I changed the UI backend and rendering. Objects are now pre-rendered and cached, drastically improving performance.
De-duplication is now completely removed and as a result there is no SLRContextID anymore.
The parsers now collect string numbers and integers in new "num" tag rows. Allowing for easy font changes in MV/MZ.
The Prepare Project for Batch translation button was removed and is instead ran automatically on project creation.
The Fix Cells and Check for Errors button was removed it will instead now run it immediately during translation and only on the cells that were actually translated. This now also means it works on destinations other than [1].
Error codes will now not be put into the translation, but instead copied with the codes into a new Error[4] column.
Overwrite is now [5].
Collected plugins will no longer include strings that are modified by the plugins.js, since changing them would just break stuff for no reason.
It is now possible to insert ruby code at the start of Main in VX and VXAce Scripts to insert stuff like Font.default_size values.
Fixed OOM crashes for exporting extremely large VX/VXAce games.
Fixed inefficient parsing logic, VX/VXAce parsing is now 17x faster.
Implemented automatic logs of the last 3 sessions (saved to data>logs) that uncollapse all console log entries. So that you can read the console log even if the tool crashed so badly that it just closed itself. (Can be turned off in options)
Rebuilt the options menu and removed obsolete or useless options.
Added an automatic spellchecker. (Can be turned off.)
Replaced trans format with compressed much more complete slrproj.
Made project creation for certain titles more convenient.
Linux version is no longer an afterthought, it has ALL features of the windows version.
Removed a lot of obsolete code and files.
Ragingpeanut
New member
- Feb 24, 2024
- 5
- 4
- Aug 17, 2016
- 583
- 3,026
I've released v3.038.
Most notable change, I've added new option to add as many primary endpoints as you want.
And instead of using the concurrency setting in General there's now a "weight" option for the primary endpoints that decides how many concurrent requests can be made to each endpoint.
The purpose is definitely not to abuse free limits by having 50 accounts for something.
*wink wink nudge nudge*
Changelog since last Changelog post:
3.038
Added option for alternate concurrent Primary endpoints.
You can now add as many endpoints as you want and give them a weight setting.
The weight setting is how many concurrent requests the endpoint allows and it will route requests based on that.
Meaning if you have 3 endpoints, the first with 2 weight and the other two with a weight of 1, it will make 4 requests at the same time using all three endpoints and the first twice.
Added option to clear all Error[4] cells, just type "ignoreallerrors" into the field for word wrapping.
Error checks now also happen during translations done with the right click menu or the shortcut.
Added 2 new error codes for screwed up commands.
Cells can now have multiple error codes instead of just the last.
Fixed a typo in the MTLTextProcessor module.
3.037
Reworked DSLR Fallback system.
Fixed some DSLRTranslationEngineWrapper patterns.
Fixed a very niche batch validation bug.
3.036
Improved DSLRGlossary server error handling.
3.035
Added new DSLRGlossary option to change the response format to be compatible with official DeepSeek. (Or other providers that do not like JSON schemas.)
Added "additionalProperties: false" to the JSON schema to support some very specific stupid endpoints.
Changed first DSLR preset to a Nano-GPT DeepSeek V4 Flash 731.
3.034
Fixed a typo in the DSLR chain meant to try to fix broken responses.
3.033
Added fixing of \[] commands in JSON Full Batch and the DSLRGlossary if the model screws up the JSON and for example thinks that \n[1] is a linebreak.
Fixed DSLRGlossary requests not using the custom command replacements.
3.032
Added the automatic response trimming of DSLR to the DSLRGlossary response parsing to try to fix LLM JSON fuckups.
3.031
SEP
It will now try to detect low contrast environments and add an outline to drawn text if it would otherwise be hard to read.
3.030
SEP
Improved support for text erasure on semi transparant images.
Improved text color detection.
A single space now deletes the original text without drawing new text.
Spaces at the start will now be correctly applied.
(Used to only work with IDSP)
Most notable change, I've added new option to add as many primary endpoints as you want.
And instead of using the concurrency setting in General there's now a "weight" option for the primary endpoints that decides how many concurrent requests can be made to each endpoint.
The purpose is definitely not to abuse free limits by having 50 accounts for something.
*wink wink nudge nudge*
Changelog since last Changelog post:
3.038
Added option for alternate concurrent Primary endpoints.
You can now add as many endpoints as you want and give them a weight setting.
The weight setting is how many concurrent requests the endpoint allows and it will route requests based on that.
Meaning if you have 3 endpoints, the first with 2 weight and the other two with a weight of 1, it will make 4 requests at the same time using all three endpoints and the first twice.
Added option to clear all Error[4] cells, just type "ignoreallerrors" into the field for word wrapping.
Error checks now also happen during translations done with the right click menu or the shortcut.
Added 2 new error codes for screwed up commands.
Cells can now have multiple error codes instead of just the last.
Fixed a typo in the MTLTextProcessor module.
3.037
Reworked DSLR Fallback system.
Fixed some DSLRTranslationEngineWrapper patterns.
Fixed a very niche batch validation bug.
3.036
Improved DSLRGlossary server error handling.
3.035
Added new DSLRGlossary option to change the response format to be compatible with official DeepSeek. (Or other providers that do not like JSON schemas.)
Added "additionalProperties: false" to the JSON schema to support some very specific stupid endpoints.
Changed first DSLR preset to a Nano-GPT DeepSeek V4 Flash 731.
3.034
Fixed a typo in the DSLR chain meant to try to fix broken responses.
3.033
Added fixing of \[] commands in JSON Full Batch and the DSLRGlossary if the model screws up the JSON and for example thinks that \n[1] is a linebreak.
Fixed DSLRGlossary requests not using the custom command replacements.
3.032
Added the automatic response trimming of DSLR to the DSLRGlossary response parsing to try to fix LLM JSON fuckups.
3.031
SEP
It will now try to detect low contrast environments and add an outline to drawn text if it would otherwise be hard to read.
3.030
SEP
Improved support for text erasure on semi transparant images.
Improved text color detection.
A single space now deletes the original text without drawing new text.
Spaces at the start will now be correctly applied.
(Used to only work with IDSP)
- Aug 17, 2016
- 583
- 3,026
I've released v3.039.
I've made the multiple primary endpoints system smarter and added options for odd providers.
Changelog:
3.039
Added new server error handling for multiple endpoints.
If there's a 401, 403, or 404 error that endpoint will be blacklisted for the session and not used.
If there's a 429 the endpoint will be given an escalating cooldown. (10s, 30s, 1m, 2m, 5m, 15m, 30m capped at 30m.)
The endpoint will be skipped during the cooldown, if it succeeds a request the colldown resets to 10s.
Added option (default off) to group primary endpoints by URL, in case a provider actually cares about concurrent connections from the same IP.
When endpoints are grouped only the highest weight setting is used for concurrency, but even if set to 1 they still rotate each batch to be used evenly.
If an error forces rotation it will wait a minimum of 5s (customizable) before making the next request to avoid hitting the same endpoint immediately with a different key, in case a specific provider doesn't like that.
Moved +Add Endpoint button location to make more sense.
I've made the multiple primary endpoints system smarter and added options for odd providers.
Changelog:
3.039
Added new server error handling for multiple endpoints.
If there's a 401, 403, or 404 error that endpoint will be blacklisted for the session and not used.
If there's a 429 the endpoint will be given an escalating cooldown. (10s, 30s, 1m, 2m, 5m, 15m, 30m capped at 30m.)
The endpoint will be skipped during the cooldown, if it succeeds a request the colldown resets to 10s.
Added option (default off) to group primary endpoints by URL, in case a provider actually cares about concurrent connections from the same IP.
When endpoints are grouped only the highest weight setting is used for concurrency, but even if set to 1 they still rotate each batch to be used evenly.
If an error forces rotation it will wait a minimum of 5s (customizable) before making the next request to avoid hitting the same endpoint immediately with a different key, in case a specific provider doesn't like that.
Moved +Add Endpoint button location to make more sense.
- Aug 17, 2016
- 583
- 3,026
I've released v3.041.
Added support for "/v1/responses" format endpoints.
I DO NOT RECOMMEND THAT ANYONE ACTUALLY USES THEM.
Not only are they strictly a worse format than /chat/completions even if they were working properly, but they are an outright scam right now.
Probably a scam that can be explained with gross incompetence rather than malice, but a scam nonetheless.
If you send a translation request for the single word: "アーリャ"
It will send you a response with the content "Alya", which is 2 tokens, and the claim that there have been "reasoning_tokens:0".
But will bill you for 116 tokens.
Because the model still reasoned, and it still bills you for it, while claiming that it didn't.
That means if you don't want to pay for reasoning, using that endpoint format means you'll pay up to ~5000% more for the translation while it pretends that everything worked as expected.
Isn't that fun?
Added support for "/v1/responses" format endpoints.
I DO NOT RECOMMEND THAT ANYONE ACTUALLY USES THEM.
Not only are they strictly a worse format than /chat/completions even if they were working properly, but they are an outright scam right now.
Probably a scam that can be explained with gross incompetence rather than malice, but a scam nonetheless.
If you send a translation request for the single word: "アーリャ"
It will send you a response with the content "Alya", which is 2 tokens, and the claim that there have been "reasoning_tokens:0".
But will bill you for 116 tokens.
Because the model still reasoned, and it still bills you for it, while claiming that it didn't.
That means if you don't want to pay for reasoning, using that endpoint format means you'll pay up to ~5000% more for the translation while it pretends that everything worked as expected.
Isn't that fun?