I can’t provide an ETA, but we are working on it
1 Like
What we really, really, really need:
-
Shared Model Storage: Maintain a single, centralized model directory shared across all applications to eliminate duplicate downloads and storage.
-
Parallel Offline Downloads: Support downloading all required models through a single download link, using multi-threaded or parallel downloads instead of downloading models sequentially.
-
Resilient Installation: Design an installation architecture that handles network interruptions gracefully, resumes from the last successful checkpoint, and never restarts the entire process unnecessarily. (which it always does, it seems)
-
Lossless Model Compression: Support lossless compression of model files, either during transfer or on disk, with transparent decompression or streaming at runtime to minimize storage and bandwidth usage without altering model weights.
3 Likes
The other day, I had some fun testing Wonder 3.5 and Wonder 2 once again on a different poster. Well, it managed to capture everything cleanly even the details of the paper itself.
Look at this:
Poster (THE BELGIAN COAST) 1934
An example here on this improved section.
I did it in 2 passes: once at 4x, and then again at a bit over 2x, to reach 14,043 pixels on the longest side. I also created a variant using four passes with Wonder 2, and finished with a little over two passes using Wonder 3.5, resulting in 14,043 pixels on the longest edge. I then combined the two variants. The version using Wonder 2 captured certain finer details in the poster most notably the tennis racket held by the man on the right side of the poster.
As a quick aside, the first time I restored this poster, the most powerful model available at the time was Recover V1; it took me about 1 to 2 hours to complete the two passes. This time, using the two passes of the Wonder 3.5 variant, it took nearly 7 minutes—and the quality is better.
1 Like
The restoration of the lettering, among other things, deserves great recognition. For instance, the original “… EASTERN RAILWAY” was something of a puzzle to decipher; the result is excellent.
Yeah, I’m genuinely blown away by this extreme yet coherent restoration.
The removal of the original sharpening halos is phenomenal. It is now just two typos short of perfection.
What short typos?
“Published by the” became “Published , che” but to be fair without further knowledge it’s impossible to read these words in the original image. So the result is still impressive.
It also missed the LTD in JARROLD & SONS LTD.
Here is the source image. It’s the highest resolution I could find on the web.
And if you look closely, it’s definitely “Che”—not “The”—that’s written there. In any case, I see “Che”. But perhaps it’s “the”. It’s hard to tell because of the low resolution.
It would be “LTD”—I thought it was some kind of symbol next to the “L”.
A post was split to a new topic: Redefine | Understanding Workflow
