just the workflow for the Turbo AIO model of ACE-Step 1.5
(it may prompt you to download the model when you drag it into comfyui, im just not sure but you may have to download the turbo model from the native workflow template on comfyui, this is a different workflow if you dont like theirs.)
i found that the subgraph may be misleading in the native workflow for some people so i deconstructed it and labeled it properly for those who dont understand. example will generate an instrumental in the workflow.
Description
deconstructed workflow, thats it. no models included.
FAQ
Comments (11)
I get gibberish noise as result with AIO templates in Comfy and this wf...not sure what is wrong I got all installed properly and there are no errors reported in console. It works fast but produces noise and some sort of old MIDI kind of sounds.
did you ask AI why this is happening?
@realrebelai not sure if limited to AIO model or how much one can trust Chatgpt lately it is universal ignoramus. I will test non aio version tomorrow...
The default AIO template in ComfyUI assumes float16 / bf16 precision for inference.
ACE-Turbo AIO requires FP32 to decode audio properly.
On 3B FL / Blackwell GPUs, float16 causes noise/static.
The vocoder node in the default template may be misconfigured:
Often, it outputs 44.1 kHz stereo PCM by default.
ACE AIO expects 48kHz mono.
Lyrics conditioning node is not connected correctly in default template
This can sometimes push the output tensor to invalid shape → more noise.
So basically: the template runs without errors, but the tensors it feeds to the vocoder are wrong, producing static.
@erdelman73267 Update - ComfyUI version: 0.12.3, I had the same problem, now everything works fine
@erdelman73267 i think you need to just update comfy fr
@realrebelai the problem was that comfyui frontend did not automatically update, it had to be manually updated, also templates got updated, now it works fine.
@erdelman73267 ahhh ive ran into that issue with other updates, its probably a custom node set you installed thru git. for some reason that locks my update from forcing and stays on whatever update the nodes were installed on. weird but it just blocks frontend and dependancies from installing. i recommend keeping a log of your custom nodes and removing them during updates and placing them back in after!
@realrebelai yes actually several custom nodes installed thru git. updates everything but frontend nada, but it's easy to check in settings About section, the version of Frontend and update manually if needed. Chatgpt was full of BS :D
@erdelman73267 yea that shiii hallucinates every other day for me. Gemini wasted my daily veo credits when i asked for prompts. Took 8 times for it to give me 1 simple image prompt bro i was mindblown. It called itself stupid
Sorry for this wall of text; i cant upload images to chat i guess. ANyway i wanted to show that the default workflow does for me (6gb RTX 1000 laptop). It is hit/miss with the generation; sometimes it sounds amazing and sometimes its a little off. So thats why I am here. I will try your wf; I just wanted to show poster that on my 6gb card it does not produce stuff. also my startup includes: fp8_e4m3fn-text-enc
got prompt
model weight dtype torch.bfloat16, manual cast: None
model_type FLOW
VAE load device: cuda:0, offload device: cpu, dtype: torch.float16
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float8_e4m3fn
Requested to load ACE15TEModel_
loaded partially; 3827.68 MB usable, 3807.67 MB loaded, 297.00 MB offloaded, 20.00 MB buffer reserved, lowvram patches: 0
Error collecting metadata (pre-execution): tuple index out of range
Requested to load ACEStep15
loaded partially; 3789.55 MB usable, 3765.55 MB loaded, 799.80 MB offloaded, 24.00 MB buffer reserved, lowvram patches: 0
100%|█████████████████████████████████████████████████████████████████████████████████████| 8/8 [00:04<00:00, 1.95it/s]
Requested to load AudioOobleckVAE
0 models unloaded.
loaded completely; 321.70 MB loaded, full load: True
Prompt executed in 93.57 seconds
@fcukit update comfy