Speridlabs tests direct pixel-by-pixel generation, finding no consistent image-quality improvement

Article

AI Image generation skips compression, but image-editing results are no better

Beating AI news flash: Spanish AI startup Speridlabs has tried a new approach to image generation. Today, many AI image-generation models first generate images from compressed image information and then reconstruct them. This process may lose detail. Speridlabs wanted to know whether results would improve if this step were skipped and every pixel generated directly.

To this end, the team trained the open-source model Iris-3B, and also modified another existing model into a version that generates pixels directly. They had these models restore low-resolution images and determine the relative distance of objects in photos, then compared them with the original approaches. The results showed none of the expected consistent improvements; some metrics were instead slightly worse. Models that generate pixels directly also exhibited slight checkerboard artifacts in some test results.

However, the study also demonstrated the feasibility of generating pixels directly.Iris-3B achieved a score of 0.540 in an image-generation evaluation, essentially on par with Qwen-Image's score of 0.539. However, removing the image-compression step did not give it a clear advantage in tasks with higher detail requirements.

Original link https://m.theblockbeats.info/flash/370987

Key points

Speridlabs is trying an image-generation approach that directly generates every pixel.
The team trained the open-source model Iris-3B, and converted another existing model into a version that generates pixels directly.
Tests included restoring low-resolution images and determining the relative distance of objects in photos.
The test results showed no expected consistent improvement, with some metrics slightly worse.
Some test results showed slight checkerboard artifacts.
Iris-3B It scored 0.540 in an image-generation evaluation, while Qwen-Image scored 0.539。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Some image-generation models first process compressed image information and then reconstruct images. Speridlabs tried skipping compression and generating pixels directly, but the current tests showed no consistent advantage.

Why it matters to readers

The study shows that pixel-by-pixel image generation is feasible, but removing the image-compression step did not produce a clear improvement in results.

It shows that a technical approach that sounds more direct does not necessarily perform better across all tasks; the specific test results still need to be considered.

Risks and unknowns

No paper, model weights, or test code were provided as evidence; the study details and results still require further verification.
The evidence does not specify 0.540 and 0.539 the specific evaluation names and scoring criteria corresponding to the scores.

Speridlabs tests direct pixel-by-pixel generation, finding no consistent image-quality improvement

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center