Open-source AI music has just become dramatically more practical. YuE2 is a local AI music generator and editor that can create full songs from text, work across genres and languages, transform reference audio into covers, and run on hardware with as little as 4 GB of VRAM using a smaller quantized model.
That is a major development for Canadian creators, agencies, startups, and in-house marketing teams. In an era where branded content must move at social-media speed, organizations want more creative output without surrendering their entire workflow, assets, and iteration process to a closed platform. YuE2 offers a compelling alternative: a locally runnable music workflow with unusually flexible editing controls.
The results are not merely novelty tracks. YuE2 can generate smoky jazz, 1930s boogie woogie, Spanish flamenco, Russian folktronica, Korean emo rock, cinematic instrumentals, and ambitious musical transformations based on existing reference audio. It also introduces a score-first approach that makes AI music generation more editable than the typical one-prompt experience.
For Canadian tech leaders tracking generative AI, this is a signal worth taking seriously. Music generation is moving from a cloud-only experiment into a workflow that can be configured, repeated, and integrated with other local and agentic AI tools.
Table of Contents
- YuE2 intro
- Jazz demo
- Boogie woogie
- Spanish flamenco
- Folktronica
- Korean emo rock
- Making cover songs
- Beethoven cover
- Minor cover
- Agent iterative editing
- How it works
- Benchmarks
- Luma
- How to install YuE2
- Text to music
- Instrumental only
- How to use reference audio
- License
YuE2 intro
YuE2 stands out because it combines three capabilities that are usually difficult to find together in open-source AI music tools: text-to-music generation, music editing, and cover-style generation from reference audio.
It is designed to produce complete music with vocals and instrumentation from a style description and lyrics. Rather than limiting users to a short loop or an instrumental sketch, the workflow can build structured songs, including verses, choruses, bridges, intros, and outros.
Accessibility is another important part of the story. The full BF16 checkpoint is approximately 7.8 GB, while a quantized version is approximately 3.96 GB and is intended for systems with less than 4 GB of VRAM. That does not mean every laptop becomes a production studio overnight, but it significantly lowers the hardware barrier for experimentation.
For organizations in Toronto, Montreal, Vancouver, Waterloo, Calgary, and beyond, local deployment can be particularly attractive. A locally managed AI workflow gives technical teams more control over model files, creative inputs, and pipeline configuration than an entirely browser-based service.
Jazz demo
A smoky jazz example demonstrates the kind of atmosphere YuE2 can derive from a well-written style prompt. The intended sound is intimate and nocturnal: clinking glasses, dim neon, quiet conversation, and a relaxed vocal delivery that sits naturally inside a late-night jazz setting.
This is where prompt specificity matters. A vague request for “jazz music” leaves enormous room for interpretation. A stronger brief can identify the mood, pacing, instrumentation, production feel, and emotional setting. Consider the difference between a generic prompt and a more operational creative direction:
- Generic: “Jazz song.”
- Specific: “Smoky late-night vocal jazz, brushed drums, upright bass, muted piano chords, relaxed tempo, intimate lounge atmosphere.”
For a business user, that distinction matters. A restaurant group developing ambience concepts, a Canadian retail brand making social content, or a GTA agency prototyping an audio identity needs outputs that align with a concrete brand feeling. YuE2 responds best when the prompt provides that creative framework.
Boogie woogie
YuE2 can also jump into a very different historical style: 1930s boogie woogie. The demo uses a shuffle-driven, danceable sound with an old-school rhythmic pulse and lyrics built around movement, doors opening, new floors, and the irresistible urge to dance.
The important takeaway is not simply that YuE2 can imitate an era. It can adjust the musical vocabulary around a genre request. Boogie woogie asks for a different rhythmic energy, instrumental character, and vocal attitude than smoky jazz. The model is being asked to interpret more than words. It is translating an aesthetic brief into a musical arrangement.
That is valuable for creative ideation. Businesses regularly need several directions before choosing one: retro, contemporary, elegant, playful, dramatic, or understated. Generating early audio concepts can help teams align on the emotional direction of a campaign before commissioning, licensing, or finalizing production work.
Spanish flamenco
Language support is another strength. A Spanish flamenco example shows YuE2 generating a song with Spanish lyrics and a regional, spirited character. The piece references Jaén, olive groves, neighbourhood life, pride, laughter, and cultural roots, all framed through an energetic flamenco-inspired style.
For Canadian organizations operating in bilingual and multilingual markets, this capability is highly relevant. Canada’s business environment is diverse, and communications often need to move beyond one language and one cultural register. A tool that can work with lyrics in different languages gives creative teams a broader starting point for experimentation.
Of course, responsible teams should still review every lyric, pronunciation, cultural reference, and final production decision. AI can accelerate concept generation, but brand accuracy and cultural sensitivity still require human direction.
Folktronica
YuE2’s range extends beyond familiar North American genres. The folktronica example in Russian combines electronic production with folk-oriented musical ideas. It highlights the model’s ability to work with a style that is already hybrid by nature.
This is a useful reminder that modern AI music tools should not be judged only by whether they can make pop, rap, or cinematic background music. The more serious question is whether the system can follow a nuanced blend of genre, instrumentation, language, and mood.
For creative technology teams, hybrid genres are where the practical opportunities become more interesting. A brand might want acoustic warmth with electronic momentum. A documentary pitch could need folk textures layered into a contemporary score. A startup launch may need something more distinctive than a standard stock-music cue. YuE2 makes those experiments more accessible.
Korean emo rock
The Korean emo rock demonstration adds another signal of flexibility. Emo rock is not defined only by distorted guitars. Its identity depends on emotional intensity, vocal delivery, arrangement choices, and a sense of rising tension or release.
Generating a track in Korean with a genre-specific emotional profile illustrates that YuE2 can accept both stylistic and linguistic requirements. It will not eliminate the need for musicians, producers, or language experts, but it gives those professionals a powerful starting point for drafts, references, mood exploration, and musical prototyping.
For Canadian media, gaming, marketing, and content businesses, this type of tool can help explore global sonic references much faster. The strategic advantage is speed of ideation, not a substitute for creative judgment.
Making cover songs
The most striking YuE2 capability is reference-audio generation. Instead of starting only with a text description, users can provide an audio clip and ask the system to use its melody or broader musical structure as a reference while generating a new interpretation.
In practical terms, this enables cover-style transformations. A familiar melody can be recast in another genre, paired with different lyrics, or reinterpreted with a changed mood and arrangement. YuE2 can use either:
- Melody only, which is the preferred option when the goal is to carry over the core melodic idea into a new cover-style output.
- Full-song reference, which takes more of the supplied source audio into account.
This is a significant workflow feature because musical identity often lives in the melody. If a model can retain that foundation while rebuilding the production around a new brief, creators gain a far more direct editing mechanism than starting from scratch.
However, the business case comes with an obvious caution. Existing songs may be protected by copyright and other rights. Organizations should only use reference audio where they have the necessary rights, permissions, or a clear legal basis. YuE2’s technical capability should never be confused with authorization to commercially exploit a particular song or recording.
Beethoven cover
A particularly memorable example takes a recognizable Beethoven piece and reframes it as theatrical hard rock. The new lyric concept is intentionally absurd: “Where’s my wallet?”
It is a playful demonstration, but it reveals something serious about YuE2’s creative mechanics. The model can preserve the familiarity of a musical source while swapping the genre, vocal framing, and lyrical content. Classical material can become theatrical rock. A traditional tune can become jazz-funk. A core melodic concept can become the anchor for a radically different production.
For companies, this suggests potential uses in internal creative exploration, pitch decks, concept testing, educational experiences, and rapid soundtrack mockups. The output may not always be the final asset, but it can accelerate the conversation around what the final asset should sound like.
Minor cover
Another example uses “Jingle Bells” as reference audio but changes the musical character into a minor version. This is a deceptively powerful demonstration because moving a well-known major-key song into a minor setting immediately changes its emotional meaning.
That type of edit would traditionally require musical knowledge, arrangement work, and production time. YuE2’s score-oriented process makes transformations like major-to-minor conversion more approachable within an AI workflow.
For Canadian businesses producing seasonal campaigns, this is an intriguing creative option. The same familiar musical language can be reshaped into comedy, suspense, drama, irony, or an unexpected premium aesthetic. Again, legal clearance is essential when using recognizable protected music, but the underlying capability is remarkable.
Agent iterative editing
YuE2 becomes even more compelling when paired with an AI agent such as GPT Astra or GLM. Rather than manually rebuilding every configuration, a user can direct an agent through natural-language instructions and ask for iterative changes.
A pop song can be generated, then revised through a sequence of creative directions:
- Keep the melody and lyrics, but make the harmony more jazzy.
- Use reharmonization to shift the musical feel.
- Remove the guitar because the result still is not sufficiently jazz-oriented.
- Generate another version using the revised specification.
This is the real promise of agentic creative AI. The first output is not the final answer. It is a working draft that can be evaluated, adjusted, and regenerated. That is closer to how human creative teams operate.
For a Canadian marketing department, this could reduce the friction between a creative brief and early prototypes. For a startup founder, it can make early brand exploration more feasible. For IT leaders, it shows how AI agents may become orchestration layers around specialized models, moving from isolated prompts toward repeatable multi-step workflows.
How it works
YuE2’s technical approach is a key reason it can support flexible editing. Rather than converting a prompt directly into a finished audio file with no editable middle layer, it first generates an editable ABC notation score.
This score contains the core musical instructions, including vocal notes, instrumental notes, key, and tempo. It functions like a basic form of sheet music, providing a musical backbone that the system then uses to generate the full audio.
That intermediate representation matters. It helps explain why YuE2 can more naturally support tasks such as:
- Creating cover-style songs from reference audio.
- Changing lyrics while retaining a musical foundation.
- Preserving or adapting melody.
- Changing a song’s key or emotional direction.
- Generating vocal and instrumental structures from a common score.
For technology leaders, this is an important architectural distinction. Systems with structured intermediate outputs can be easier to inspect, adjust, and build into larger workflows than systems that produce only a final opaque result.
Benchmarks
YuE2’s reported Song Quality Index results place it ahead of other open-source music models, including MiniMax Music and AStep 1.5. The benchmark comparison also places YuE2 above Suno V6 and Suno 5.5, two closed-model points of comparison.
There is one nuance worth noting: the same benchmark places Suno V5 slightly ahead of YuE2. Interestingly, that benchmark result suggests Suno V5 performs better than both Suno 5.5 and Suno V6 on this quality measure.
Benchmarks should never be treated as the whole story. Quality is subjective, and a single score cannot capture every business requirement, including language handling, latency, local deployment, editability, licensing, hardware needs, and consistency. Still, the result is notable because it shows an open-source option competing seriously with prominent closed services.
That is precisely why Canadian AI decision-makers should pay attention. The market is no longer divided neatly between powerful closed tools and weak local alternatives. Open models are rapidly narrowing the practical gap.
Luma
YuE2 is focused on music generation, while Luma represents a broader vision of agentic creative workspaces. Luma Agents are positioned to help users develop concepts, create visuals, and shape a project in one environment rather than forcing them to jump between disconnected tools.
The platform’s focus on motion, physics, and 3D space makes it especially relevant for visual content, product storytelling, branded creative, and campaign development. Instead of taking complete control of the project, the agent can be guided continuously as creative direction changes.
One of Luma’s more useful workflow ideas is the concept of reusable skills. A skill can package a repeatable set of instructions so that a workflow does not need to be recreated every time. For example, a team could build a repeatable process that turns a product image into influencer-style user-generated content, or one that places a product image into a water-drop visual effect.
For Canadian businesses, this is the operational side of generative AI. The question is not just whether a tool can create a compelling asset once. The question is whether a team can turn successful creative processes into reusable systems.
How to install YuE2
YuE2 can be run through its official repository using Python, but a more approachable option is to use ComfyUI. ComfyUI is a popular visual environment for running open-source image, video, audio, and other generative AI tools locally.
The installation workflow begins with updating ComfyUI to the latest version. In the ComfyUI root folder, open the update folder and run update_comfyui.bat. After the update finishes, start a fresh ComfyUI session.
Next, download the YuE2 workflow JSON file and drag it into the ComfyUI interface. This loads a pre-built workflow, avoiding the need to assemble every node from scratch.
If the workflow displays missing-model errors, download the required model assets:
- Audio encoder: approximately 1.4 GB, placed in
ComfyUI/models/audio_encoders. - Full BF16 checkpoint: approximately 7.8 GB, placed in
ComfyUI/models/checkpoints. - Quantized checkpoint: approximately 3.96 GB, intended for systems with less than 4 GB of VRAM, also placed in
ComfyUI/models/checkpoints.
After adding the files, refresh the model list with R, then select the downloaded checkpoint in the Load Checkpoint node. The workflow should then resolve the missing-model errors and become ready for configuration.
Text to music
The top section of the YuE2 ComfyUI workflow handles text-to-music generation. It uses two core inputs: a style prompt and lyrics.
The style prompt should describe the desired genre, pacing, instruments, atmosphere, and production direction. A prompt such as “future bass, modern, energetic, inspiring” provides a basic style target. More detailed descriptions can improve the specificity of the result.
The lyric input accepts structural metadata. This means users can label parts of a song, such as:
- Intro
- Verse 1
- Verse 2
- Pre-chorus
- Chorus
- Bridge
- Outro
The Generate ABC node creates the notation score, which can be previewed before it is passed to the music generation stage. This provides visibility into the vocal and instrumental note structure behind the eventual audio.
Several generation settings are especially useful:
- Maximum duration: Sets the upper limit of the song length in seconds. A setting of 360 seconds allows up to six minutes, but short lyrics should use a shorter maximum to avoid excessive continuation.
- Seed: Identifies a specific generation. Keeping the same seed, prompt, and settings reproduces the same result. Changing or randomizing the seed creates a new variation.
- Steps: Higher step counts generally prioritize quality but take longer to generate. A default of 32 steps is a practical balance.
- CFG: Controls how literally the model follows the prompt. Increasing CFG can help when the genre direction or lyrics are not being followed closely enough.
- Sampler and scheduler: Generation algorithms that can generally remain at their default settings for initial testing.
Decoding method also depends on hardware. Systems with more than 12 GB of VRAM can use the regular decode method for faster processing. Systems with less available VRAM should use tiled decoding.
On a laptop equipped with an Nvidia RTX 5000 Ada GPU with 16 GB of VRAM, a sample generation took just over two minutes. Generation time will vary based on hardware, song duration, model choice, and settings.
Instrumental only
YuE2 can generate instrumental music, although the workflow still expects an entry in the lyric field. The practical workaround is to specify instrumental directions in square brackets.
For example, a style prompt could request “epic, cinematic, orchestral music for a battle scene, instrumental only.” The lyric field could then include a production instruction such as “[steady buildup of staccato strings and ethnic drums].”
This approach gives the model additional information about musical development even when there is no conventional vocal lyric. It is useful for background scores, concept reels, presentations, product videos, internal events, and early-stage game or film audio ideas.
For business teams, instrumental generation may be the most immediately practical entry point. It can provide custom mood references without requiring lyric writing, vocal identity decisions, or language localization.
How to use reference audio
The lower section of the ComfyUI workflow handles reference audio. To activate it, unbypass the reference-audio nodes, bypass the top Generate ABC section, and connect the reference-audio output into the Generate Music node.
Then select the downloaded audio encoder in the Load Audio Encoder node and upload an audio clip. YuE2 can use that clip to generate the ABC notation that drives the new music output.
The core process is straightforward:
- Upload a reference audio clip.
- Select whether to use melody only or the full song as the reference.
- Enter a new style prompt.
- Add the original lyrics or replace them with entirely new lyrics.
- Run the workflow to generate the transformed result.
For a cover-style result, melody-only reference is generally the most useful setting. It allows the core melodic idea to be carried into a new arrangement. A reference track can become jazz with casual piano, saxophone, double bass, and brushed drums, for example.
The lyrics do not have to remain the same. They can be rewritten, translated, or replaced with a different language. That makes YuE2 one of the more flexible open-source options for creatively reworking music through text direction and reference audio.
License
YuE2’s licensing deserves careful attention, especially for Canadian businesses considering commercial deployment.
The code and documentation are released under the Apache 2.0 licence, which has minimal restrictions. However, the model weights are covered separately by a Creative Commons non-commercial licence.
The non-commercial terms state that use must not be primarily intended for or directed toward commercial advantage or monetary compensation. In practical terms, organizations should not assume they can generate a track with YuE2 and monetize it on platforms such as Spotify, use it in paid advertising, or deploy it in a commercial product without first understanding the applicable licence terms and obtaining appropriate legal advice.
This does not diminish YuE2’s importance. It simply defines its current position. It is an exceptionally capable tool for research, education, experimentation, creative prototyping, and non-commercial use. For enterprise use, licensing review is not an afterthought. It is a required part of the evaluation process.
The bottom line: YuE2 is a major open-source AI music development. Its ability to generate full songs locally, work in multiple languages, create instrumental material, edit through a score-first process, and transform reference audio makes it a serious creative technology platform rather than a gimmick.
For Canada’s fast-moving AI ecosystem, the opportunity is clear. Teams that learn how to combine local models, agentic workflows, reusable creative systems, and responsible rights management will be better positioned to move quickly without losing control of their creative process. Is your organization prepared to turn generative AI from a one-off experiment into a repeatable creative advantage?
