Great potential and wonderful graphic. But has Some issue

[DreamFilmVFX]

As i said in the title this UI has a lot of cool potential but suffer of some critical issue. I'm using it right now as i write this post, so i can update this post as i find an issue.

For now i found these issues:

1 - Doesn't matter if the seed is randomized. it generates almost the same song or slight variaton of it. But all of them are very similar.
2 - When i select inference steps 8, or 32 dont's seems to change anything about time of generation (its seems to stick to the turbo model) and i have all the models already downloaded
3 - when the CoT (thinking mode) option is enabled, the inference fails. with this error

"Generation failed: Generation failed on API side" (in the interface) while in the CMD windows there is this log:

INFO: 127.0.0.1:63547 - "POST /query_result HTTP/1.1" 200 OK Generating: 0%| | 0/2 [00:00<?, ?steps/s] 2026-02-04 22:48:21.948 | ERROR | acestep.llm_inference:generate_with_stop_condition:1104 - Error in batch codes generation: INFO: 127.0.0.1:63547 - "POST /query_result HTTP/1.1" 200 OK.

EDIT: it seems to be and Api call error because now every inference doesn't working event if i de-select Thinking mode. Now it gives me always API error

Will be fantastic if i can select all the models directly in the UI, i have all the models for the generation and all the LLMs. But i cannot select it.

[fspecii]

Hi @DreamFilmVFX, thanks for the detailed feedback and for taking the time to document each issue! Really appreciate it. Let me go through them one by one.

1. Seeds produce similar results

This is actually a behavior of the ACE-Step model itself rather than a UI bug — the UI does correctly randomize and pass different seeds to the backend. A few things that can help get more variety:

• Try different style/caption descriptions — even small wording changes can shift the output significantly
• Experiment with guidance scale — lower values (5-7) give more creative/varied results, higher values (10+) stick closer to the prompt
• Use SDE inference method instead of ODE — SDE (stochastic) adds more randomness to the generation process
• Adjust the shift parameter — this can also influence output diversity

2. Inference steps not affecting generation time

This is likely because the turbo model is optimized for a specific step count and additional steps don't significantly change the output or timing. We're currently reviewing a PR (#19) that adds model selection directly in the UI — this would let you switch between base, sft, turbo, and other model variants, each of which responds differently to step count changes. Stay tuned for that!

3. CoT (Thinking mode) crashing generation

Good news — this has been fixed in recent updates! We've made several changes:

• CoT features are now explicitly disabled by default so they won't accidentally trigger LLM errors
• New LM Backend selector — you can choose between PT (~1.6 GB VRAM) and VLLM (~9.2 GB VRAM) in Advanced Settings. PT is now the default and much more stable
• Batch size defaults to 1 — reduces VRAM pressure

The error you saw where "every inference stopped working even after disabling Thinking mode" was likely because the failed CoT request left the backend in a bad state. A restart of the ACE-Step API should recover from that.

4. Model selection in the UI

Totally agree this would be great! As mentioned above, PR #19 includes exactly this feature with support for multiple model variants. We're reviewing it carefully and hope to have it merged soon.

What to do now

Please update to the latest version (git pull in your ace-step-ui folder, or reinstall via Pinokio) — issues 3 and the VRAM improvements are already live. And keep an eye out for model selection coming soon!

Thanks again for the great feedback!

[DreamFilmVFX]

Thank you for this kind answer. You're such good person. I wish to ask if you can help for a big issue. I'm using your wonderful interface for generating my cover version and remasters. But i don't know what is broken since two days on. doesn't work anymore.. Everytime i try to generate a cover using the same identical audio that i've use before (when it works perfectly) now don't. Let me paste here the same message that i sent to ace-step 1.5 support. Maybe you can help me further.

"i've generate various covers with success, but from unespected moment everything is changed. Now every cover i try to generate, it gives me a file with ton of silence, and sometimes a looped chunk of audio. Is terrible, i don't know how to do. I've already tried to resolve by myself trought chatgpt but every try is a failure. Chatgpt allucinating cause i suspect that nobody has the same problem of mine. Everything has worked flawlessly since 2 days ago. I've pasted the log to show what happen from the beginnging, and the resulting Mp3 for your reference.

sotto le stelle (v2).mp3

log.txt

[fspecii]

The normalization_db default was wrong (0.0 instead of -1.0), which could have caused similar-sounding results. This is now fixed. The model selector lets you choose between turbo/base/sft - turbo uses 8 steps, base/sft use more. If you're using the model selector and selecting a non-turbo model, the steps should now apply correctly.

[fspecii]

All three should be fixed now:\n1. Similar songs regardless of seed - normalization_db default was wrong (0.0 instead of -1.0), now fixed\n2. Inference steps ignored / stuck on turbo - model switching is now handled automatically, and the model selector works\n3. CoT/thinking failures - that was the old REST API path, gone since the Gradio migration