What happened
ComfyUI’s documentation describes ACE-Step 1.5 as an open-source music-generation model from ACE Studio and StepFun. The model accepts a text description and optional lyrics, then produces a complete song. ComfyUI says it runs the model natively, with outputs ranging from 10-second loops to compositions up to 10 minutes. The guide presents two ways to install its workflow and model files, recommending the all-in-one option for most users.
The recommended AIO package combines the models in one checkpoint, ace_step_1.5_turbo_aio.safetensors, stored in ComfyUI’s models/checkpoints directory. Its workflow is called “ACE-Step 1.5 Music Generation AIO.” Builders can find it by searching that name in the Template Library, download its JSON, open it in Comfy Cloud, and provide style tags and lyrics to generate a track.
ComfyUI also documents a split-model workflow for people who want the components separately. It uses the diffusion model acestep_v1.5_turbo.safetensors, either the 0.6B or 1.7B Qwen text encoder, and ace_1.5_vae.safetensors. The diffusion model belongs in the diffusion-model directory; the text encoders and VAE go in their respective ComfyUI model directories. The workflow takes a text prompt describing the music style plus optional lyrics.
The model combines a language-model planner with a Diffusion Transformer audio decoder. ComfyUI says the planner expands a simple query into a song blueprint containing structure, metadata, lyrics, and captions, after which the DiT renders the audio. The guide calls this Chain-of-Thought planning and says alignment relies on the model’s internal mechanisms rather than an external reward model or human preference data. It also says LoRA training on a few songs can capture a user’s style.
ComfyUI reports that a full song takes under two seconds on an A100 and under 10 seconds on an RTX 3090. It describes a configuration using less than 4 GB of VRAM in which the DiT runs without the language-model planner. The documentation reports a musical coherence score of 4.72 and support for more than 50 languages, naming English, Chinese, Japanese, Korean, Spanish, German, French, Portuguese, Italian, and Russian among those with strong support.
The guide separates current ComfyUI support from capabilities that ACE-Step 1.5 has elsewhere. Cover generation can take a song, new prompt, and lyrics to reimagine the track in another style. Repaint can regenerate a selected segment while retaining the surrounding audio. ComfyUI labels both as coming soon rather than available through this workflow. Local users are told to update ComfyUI; the guide says missing nodes can arise when users are not on the latest Nightly version or nodes fail to import at startup. Comfy Cloud updates after stable releases.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This does not yet fit a pattern New Runtime tracks. The guide makes one music model easier to operate inside ComfyUI, but it does not by itself demonstrate a broader shift in agent interfaces, runtime boundaries, or production workflows.
Why it matters
For a creator already using ComfyUI, the practical change is that music generation can sit beside other node-based media steps instead of requiring a separate hosted product or a custom integration. A team can begin with one checkpoint and an existing template, while the split setup offers a more explicit view of the individual files that make up the workflow.
That matters for prototyping pipelines where generated audio is one stage among images, video, editing, or other creative work. The guide’s local path also gives teams a way to measure generation time and output against their own machines and acceptance criteria before designing a workflow around it. The documented distinction between model capabilities and available ComfyUI features is consequential: cover and repaint should not be assumed to be usable in a ComfyUI-based process today.
Limits and uncertainty
ComfyUI reports the speed, duration, language-support, and coherence figures, but these are publisher claims rather than independently verified results. The page does not describe a reproducible benchmark protocol for the 4.72 coherence score or establish how the listed speeds vary by prompt, duration, batch size, software version, or GPU configuration.
The less-than-4-GB configuration disables the language-model planner. ComfyUI does not claim that this reduced configuration produces the same planning behavior or output quality as the full model. The guide also identifies cover generation and repaint as features not yet supported in ComfyUI, without a delivery date. Builders should test their own hardware, prompts, output quality, and workflow reliability before treating the reported timings as a capacity plan.
Skill to build now
Create a small ComfyUI evaluation workflow using the AIO checkpoint first: run the same style prompt and lyrics across several target durations, record generation time, VRAM use, and a human quality review, then repeat with the split workflow only if component-level control is needed. Keep cover and repaint out of the first production design because ComfyUI marks them as unavailable. This turns the guide’s speed claims into a decision based on the team’s actual GPU and quality bar.
Reader takeaway
ACE-Step 1.5 is documented as a native ComfyUI music workflow, with a simple single-checkpoint route and a modular alternative. Treat ComfyUI’s performance numbers as a useful starting point, validate them on your own hardware, and do not assume every ACE-Step capability is already exposed in the workflow.