The Workflow was setup to have a clean "GUI" showing only parameters that matter, so you might want to toggle off Link visibility, like in above screenshot.
V1.6 Ace Step 1.5. Turbo and SFT normal and XL model with Ollama. Text to Audio/Song (examples below)
updated the settings for XL models and added a 3rd System Prompt for tags to chose, with more descriptive song descriptions.
1.5 XL SFT pipeline now has an "Adaptive Projected Guidance" node and negative prompt.
** See below some tips which model and settings to start with.
V1.5 Ace Step 1.5. Turbo and SFT normal and XL model with Ollama:
setup to create up to 4 tracks in a run, 2x Ace1.5 and 2x Ace1.5 XL, each with Turbo and SFT model, to compare (can be individually switched on/off)
VAE changed to tiled Audio VAE decode, uses less Vram.
V1.2 Ace Step 1.5. Turbo and SFT model with Ollama Text to Audio/Song
small update to GUI, system prompts and SFT sampler "engine"
V1.0 Ace Step 1.5. Turbo and SFT model with Ollama Text to Audio/Song
Ace Step uses TAGS and LYRICS to create a song. These can be generated by Ollama or by own prompts.
Can use any Song, Artist as reference or any other description to generate tags and lyrics.
Will output up to two songs, one generated by Turbo model, the other by the SFT model (experimental).
Keyscales, bpm and song duration can be randomized.
able to use dynamic prompts.
creates suitable songtitle and filenames with Ollama.
Lora Loader included, hope to see some Loras soon!
Avoid sage attention in your comfyui starting parameters, avoid --lowvram setting, as this might force Texencoder to run very slow on CPU instead of GPU.
Download Files:
Ace Step 1.5 TURBO model: https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/tree/main/split_files/diffusion_models
Ace Step 1.5 SFT model: https://huggingface.co/ACE-Step/acestep-v15-sft/tree/main (download model.safetensor and rename it)
Ace Step 1.5 XL Models: https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/tree/main/split_files/diffusion_models
Ace Step 1.5 XL Model merges (i.e. merge of turbo with SFT): https://huggingface.co/Aryanne/acestep-v15-test-merges/tree/main/xl
Textencoder: https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/tree/main/split_files/text_encoders (Qwen_0.6b and Qwen_4b required, 1.7b is a smaller alternative to 4b)
VAE: https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/tree/main/split_files/vae
Ollama Models, required for tags, lyrics and songtitle, you can choose 1,2 or 3 different models, tags and lyrics might need a bigger model >7b, songtitle can use a smaller model:
https://ollama.com/mirage335/Llama-3-NeuralDaredevil-8B-abliterated-virtuoso (allround model, fast, usable for tags, lyrics and songtitle, recommended)
https://ollama.com/huihui_ai/qwen3-vl-abliterated (for tags and lyrics, able to use thinking)
https://ollama.com/artifish/llama3.2-uncensored (small and fast for songtitle and tags)
Alternative Turbo and SFT Models (normal, non XL) :
Turbo continuous: https://huggingface.co/ACE-Step/acestep-v15-turbo-continuous/tree/main
SFT-Shift1: https://huggingface.co/ACE-Step/acestep-v15-turbo-shift1/tree/main
SFT-Shift3: https://huggingface.co/ACE-Step/acestep-v15-turbo-shift3/tree/main
Merges of SFT, Turbo and Base model: https://huggingface.co/Aryanne/acestep-v15-test-merges/tree/main
GGUF Models "normal" and XL: https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/tree/main
Which models to start with ?
My current choice for normal model: Turbo-SFT merge_ta_0.5 & SFT-Shift1, using these settings:
Turbo-SFT_merge model with sampler: er_sde, scheduler: beta57 (or beta), 22 steps
SFT-Shift1 model with sampler euler, scheduler: normal, 138 steps
XL Model settings:
XL Turbo-SFT merge model: sampler: er_sde, scheduler: sgm_uniform, 40 steps
alternative: sampler: res_s2, scheduler beta57 (requires RES4LYF custom nodes)
XL SFT model: sampler: euler (or res_2s), scheduler: normal, 46 steps, CFG = 7.3, Adaptive Projected Guidance: eta = 1.05, norm_thresh= 1.3, momentum=0.0. Increase norm_thresh as the main parameter. These settings deliver "stabil" output for XL SFT,Base and their merges. The merges sound way better, pure SFT or Base introduce a lot of noise. I bypassed ModelSamplingAuraflow (see node next to model loader node). I think the base-turbo XL model merge fits well in that slot.
Disable "generate_audio_codes" in "TextEncodeAceStep" node to get different results, it works very well for many genres and reduces process time.
Ollama Model: Llama-3-NeuralDaredevil-8b-abliterated
More infos on models see thread below in discussion.
Save Location:
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ └── acestep_v1.5_turbo.safetensors
│ ├── 📂 text_encoders/
│ │ ├── qwen_0.6b_ace15.safetensors
│ │ └── qwen_4b_ace15.safetensors (or 1.7b)
│ └── 📂 vae/
│ └── ace_1.5_vae.safetensors
Custom Nodes used:
optional (use Beta57 scheduler for a bit more punch, requires RES4LYF): https://github.com/ClownsharkBatwing/RES4LYF
Examples various styles:
https://soundcloud.com/tele-joe/firestorms_unleashed_turbo_000
https://soundcloud.com/tele-joe/westsidevibesforever_funkhipho
https://soundcloud.com/tele-joe/cosmicdreamscapesynthpop_turbo
https://soundcloud.com/tele-joe/mindspin-synthpop_turbo_00001
https://soundcloud.com/tele-joe/wildreckoning_thrashmetal_turb
https://soundcloud.com/tele-joe/freeflight_metalcore_sft_00001
Ollama help:
Install Ollama from https://ollama.com/
download a model: Go to a model page, chose a model , then hit the copy button, i.e. https://ollama.com/mirage335/Llama-3-NeuralDaredevil-8B-abliterated-virtuoso
open terminal and paste the model name, i.e.: ollama run huihui_ai/qwen3-vl-abliterated
model will be downloaded and can be selected in green comfy node "Ollama Connectivity". Hit "Reconnect" to refresh.