AceStep 1.5 SFT Model Loader

Every AceStep 1.5 SFT workflow starts here. This is the boring, essential node: it loads the four pieces of a music-generation stack - the diffusion model, two text encoders, and the audio VAE - and hands you the three wires (model, clip, vae) that the rest of the pack runs on. No knobs to fiddle, no guidance magic. If you set up the model files right, this node just works and you never think about it again.

What it loads

AceStep 1.5 isn't one model file. It's a DiT diffusion model, a Qwen3-0.6B encoder that handles captions and lyrics, a bigger Qwen3 LLM (1.7B or 4B) that generates the semantic audio codes, and a VAE that acts as the audio codec. That's why the loader has four dropdowns instead of one:

• diffusion_model - the AceStep 1.5 diffusion model (.safetensors)
• text_encoder_1 - the Qwen3-0.6B caption/lyric encoder
• text_encoder_2 - the Qwen3 1.7B or 4B LLM for audio codes
• vae_name - the AceStep 1.5 audio VAE

Under the hood it loads the diffusion model, then calls ComfyUI's CLIP loader with both text encoders together and CLIPType.ACE, then builds a VAE from the safetensors. The dropdowns are populated straight from your model folders, so the names you see are whatever files you actually placed there.

Where the files go

This is the part beginners stumble on: the four dropdowns read from different folders.

ComfyUI/models/diffusion_models/     # the SFT (or merged SFT+Turbo) diffusion model
ComfyUI/models/text_encoders/        # qwen_0.6b_ace15 + qwen_1.7b_ace15 (or qwen_4b)
ComfyUI/models/vae/                  # ace_1.5_vae.safetensors

Drop the files in the matching folders, restart ComfyUI, and they appear in the dropdowns. The tooltips on each field even show the expected filename pattern (qwen_0.6b_ace15.safetensors, ace_1.5_vae.safetensors, and so on), which is a nice touch - you can't guess wrong if you follow the pattern.

Wiring it up

• model → the Generate node (or through a Lora Loader first)
• clip → the TextEncode node (or a Lora Loader)
• vae → the Generate node

That's the whole graph, top to bottom. There are no optional inputs and nothing to configure beyond the file selection - the real decision is which files you downloaded. The README's recommendation: use the merged SFT+Turbo model rather than the pure SFT release for better quality in practice, and grab the text encoders from the Comfy-Org ace_step_1.5_ComfyUI_files repo (the loader expects them in ComfyUI's split-file format, not the full checkpoint). Model links are in the README under "Download Model Files."

Troubleshooting

• A dropdown is empty → you've put a file in the wrong folder, or haven't restarted ComfyUI since adding it. Double-check the three paths above.
• "file not found" style errors on load → the loader uses get_full_path_or_raise, so it will loudly refuse anything it can't find. Usually a filename typo or a model sitting in the wrong subfolder.
• It loads but Generate gives garbage → you're probably using a non-AceStep diffusion model with this loader. The SFT pipeline is tuned for AceStep 1.5 specifically; the README is explicit that these nodes aren't tested with anything else.

Install the pack once (ComfyUI Manager → search "AceStep", or git clone https://github.com/jeankassio/ComfyUI-AceStep_SFT.git into custom_nodes/), place the models, restart, and you're generating.