[Feature] Being able to reproduce previous generation with 0 alteration? (Seed for LM and seed for gen)

[StableInfo]

Hi,
In text to image, we can use seed , same model, same negative prompts, same parameters, and it let us generate the SAME music.
With the "think" mode , the LM interfering with the gen, we always never obtain the same music from the same parameters from a gen correct?
I was wondering if there was a way to give the "think" mode, a "seed" so we can predict how it will alter the music.

So if I use the same parameters, model, shift, seed for a gen, and a same "seed for LM", I would get THE SAME OUTPUT MUSIC.

I wonder if such feature can exist and if its too complicated or not.

[ServeurpersoCom]

An LLM is a causal model: each generated token depends on all previous tokens. Therefore, after a while, the calculation inevitably diverges due to tiny rounding errors or the thread scheduling in the hardware! However, the DiT flow matching is reproducible! That's why I didn't include the LLM seed.

Here are some technical explanations:

I separated the workflow into two parts because the project aims to be transparent about the ACE-Step architecture to the user.

By design, the prompts (Caption/Lyrics/Metadata) are common to both models because they were trained on the same data (text/music).

The "Inspire" and "Format" buttons only launch the thinking part of the LLM (CoT).

The "Compose" button launches phases 1 and 2 of the LLM inference process.

Phase 2 involves generating 5Hz audio codes. These codes are visible under "Advanced LM," and the indicator at the bottom (above the final "Synthesize" button) shows their presence.

The DiT is trained to understand its musical guidance codes, but it also uses the same prompt as the LLM!

The DiT generates 25Hz data which the VAE (which is actually a very efficient neural audio codec) decodes into final audio.

[ServeurpersoCom]

And the DiT is indeed reproducible, as long as you give it the same sources (prompt + FSQ, whether they come from the LLM or from a source music) and the same seed.

If you have a cool song in the list, simply click "Edit" to reload the prompt. When you click Synthesize again, the same song will be regenerated. You can then save it as a JSON file (using the "Save" button) to share it.

The JSON file stores the LLM code. This allows you to share the JSON file, so someone else should get the same song (with some minor differences if their hardware is different like CUDA vs. Vulkan, due to rounding errors).