DEV Community: Alex Creatune

A stem splitter needs two listening tests: soloed and recombined

Alex Creatune — Fri, 02 Oct 2026 05:10:48 +0000

I’m Alex, a member of the Creatune team. This article is a listening protocol for source separation, based on our public interface and its documented behavior. It is not a report of a new generation experiment. The screenshots show the logged-out interface; the sample players are provided by the website.

If you want to learn a bass line, a stem with some residual guitar may still be useful. If you need an exposed vocal for a new arrangement, that same amount of leakage may be unacceptable. “Clean separation” is too vague to be an acceptance criterion.

Start with the next musical task, then decide what to listen for.

1. Choose the smallest useful separation

The Creatune stem splitter currently offers a vocal/instrumental split, a four-part split into vocals, drums, bass and other, and a broader mode with up to twelve instrument stems. Accessing your music or running separation requires signing in. More stems do not automatically mean a better result for your task.

Next task	Starting output	Main question
Practise a bass phrase	Four stems	Can I follow note starts and pitch movement?
Rehearse singing	Vocal/instrumental split	Does the backing retain useful timing and harmony?
Inspect drum timing	Four stems	Are attacks still distinct enough to hear?
Build a new arrangement from a particular instrument	Broader separation, if available for that source	Is the required part present and usable when exposed?

Use audio you have permission to process. Keep an untouched copy of the source. Separation estimates overlapping sources from the mix; it does not recover the original studio multitrack files.

2. Pick three passages before generating

Do not evaluate only the quietest verse. Choose three short passages and write down their timestamps:

A sparse passage where the desired part is easy to recognise.
A dense passage where several instruments share the same range.
A transition containing a pickup, fill, breath or reverb tail.

These passages probe different weaknesses. Dense chords can reveal leakage; a drum fill can reveal softened attacks; a transition can expose an abrupt ambience change. The exact durations are your choice, and are not a product limit.

3. Solo first, then put the parts together

In your audio player or editor, compare each chosen passage in the following order:

Listen to the original and identify the target part.
Solo the relevant stem. Note bleed, missing notes, softened transients and unstable texture.
Listen to complementary stems. Is part of the target instrument assigned elsewhere?
Recombine the available stems without changing their alignment or adding processing. Compare with the original at a similar listening level.
Try the actual intended use: a practice loop, an exposed vocal or a rough remix.

A stem can sound strange in isolation but work in context. Conversely, a pleasing recombined mix does not prove that an individual stem will survive being exposed. Do not normalise every stem independently before the recombination check: that changes their balance and makes the comparison harder to interpret.

You can make these comparisons in an external editor; this protocol does not imply that the product includes a multitrack mixing workstation.

4. Use a task-specific acceptance matrix

Observation	Practice/transcription	Exposed rearrangement
Desired notes are clear, but another instrument leaks in	May be usable if it does not obscure the phrase	Review the exposed passage; leakage may conflict with new harmony
Percussion attacks are softened	Avoid drawing precise timing conclusions	Check whether the groove loses impact
Reverb tail fluctuates	Often tolerable for studying notes	Check the tail in the final mix, especially before silence
Target instrument partly appears in another stem	Listen to those stems together for reference	May require a different split or a different source
Recombined result sounds close to the original	Useful consistency check	Still inspect the individual part that will be exposed

“Accept” means suitable for the stated use, not artifact-free. If the target notes remain ambiguous, a higher stem count alone is not a solution. Try a cleaner source, a simpler split or a narrower musical objective.

5. Keep a small review record

Source and permission:
Intended task:
Separation mode:
Passages reviewed:
Solo observations:
Complementary-stem observations:
Recombined comparison:
Decision: usable / usable with limits / retry
What I must not infer from this output:

For example, a practice output may help you hear a melody without being reliable enough to measure the exact original attack times. Recording that boundary is more useful than giving the entire separation one unexplained quality score.

The practical goal is modest: identify which musical decisions the separated audio supports, and which decisions still require the original recording.

An Image Is Evidence, Not a Tempo: A Better Brief for Image-to-Music

Alex Creatune — Thu, 01 Oct 2026 04:19:48 +0000

Turning an image into music sounds like a direct conversion. It is better treated as an interpretation problem.

A photograph does not contain a hidden BPM, key, or instrument list. It gives us evidence: scale, texture, contrast, implied motion, emotional distance, and narrative focus. We still have to decide how those observations should behave in time.

This article presents a repeatable briefing method. It is a planning framework, not a report of a controlled generation test.

Separate observation from interpretation

Start with two columns.

What is visible	One possible musical interpretation
A lone figure occupies a small part of a wide landscape	Sparse arrangement and a slow expansion of register
Hard diagonal shadows divide the frame	Short rhythmic accents or an uneven pulse
Repeated windows create a grid	Repeating pattern with small variations
Warm light appears behind a dark foreground	Gradual harmonic opening rather than immediate brightness

The right column is deliberately called one possible interpretation. A dark image can support tension, calm, intimacy, or humor. Color and composition constrain the conversation, but they do not determine a single correct soundtrack.

This distinction prevents a common prompt failure: describing only what is in the picture and expecting the system to invent the musical argument.

Inventory six visual signals

Before writing a music prompt, describe the image with six short notes.

1. Subject

What should receive attention: a person, object, environment, gesture, or absence?

2. Scale

Does the frame feel intimate, human-sized, architectural, or vast?

3. Contrast

Are light, color, and shapes softly blended or sharply separated?

4. Texture

Does the image feel polished, grainy, organic, metallic, delicate, or rough?

5. Implied motion

Even a still image can suggest stillness, drift, repetition, acceleration, impact, or instability.

6. Narrative distance

Are we observing from far away, standing inside the moment, or remembering it afterward?

Do not translate these notes into music yet. The goal of this pass is to agree on what the image is doing visually.

Candidate image 1: the logged-out public interface keeps the image reference and written musical direction together.

Convert the inventory into decisions that can be heard

Now map the observations into five musical dimensions.

Pace: slow, but with a visible pulse
Register: begin low and narrow; open upward in the second half
Density: two layers at first, then add one supporting texture
Articulation: soft attacks with occasional dry accents
Arc: restrained curiosity → wider, unresolved finish

This brief is more useful than a pile of mood adjectives because each line creates something a listener can check.

“Mysterious, cinematic, emotional” may describe taste. “Begin with a narrow low register, add a pulse after the opening, and finish wider without a triumphant cadence” describes behavior.

Give every observation a confidence level

Some visual signals are obvious; others are speculative. Mark them.

Decision	Confidence	Why
Keep the opening sparse	High	The subject is isolated with large negative space
Use metallic percussion	Medium	The architecture has hard reflective surfaces
Build toward hope	Low	The bright horizon could also read as distance or loss

Low-confidence choices are useful candidates for alternate versions. They should not silently become facts about the image.

Change one interpretation at a time

If the first direction does not fit, resist rewriting the whole prompt.

Create controlled alternatives:

Version A: keep the pulse, change the emotional arc
Version B: keep the arc, replace metallic texture with wood and breath
Version C: keep texture and arc, make the pacing less regular

This produces a decision trail. You can explain why one reading supports the visual better instead of choosing the most impressive track in isolation.

Review against the image, then without it

Use two listening passes.

Pass 1: with the image visible

Ask:

Does the entrance match the image’s first focal point?
Does the arrangement respect its amount of visual space?
Does the musical arc support the implied story rather than merely repeat the mood?
Is one detail over-interpreted?

Pass 2: without the image

Ask:

Does the track still have a coherent beginning, development, and finish?
Are changes audible, or do they exist only in the written prompt?
Does the arrangement become crowded because every visual detail received its own sound?

A soundtrack should support the image and remain musically legible.

Apply the method to the current tool boundary

The current Creatune Image to Music page accepts JPG, JPEG, PNG, and WebP images up to 50 MB. The image supplies visual context, while the written prompt defines mood, instrumentation, pace, and development.

Each 25-credit run returns one instrumental soundtrack. A still image has no timeline, so the model determines the duration. The result is audio; the image is not embedded and the workflow does not render a slideshow or video.

Those limits make the brief especially important. The upload is the reference, but the prompt still has to explain how the visual idea should unfold in time.

Candidate image 2: the public page explains what the image influences, what it does not decide, and the one-result workflow.

A compact review template

Visual evidence:
- subject:
- scale:
- contrast:
- texture:
- implied motion:
- narrative distance:

Musical decision:
- pace:
- register:
- density:
- articulation:
- arc:

Uncertain interpretation to test:

One variable to change next:

The aim is not to prove that an image “means” one piece of music. It is to make each interpretation specific enough to compare, revise, and discuss.

Disclosure: I’m Alex from the Creatune team. I used Creatune’s current Image to Music interface as the product example; the observation-to-brief method can be used with other visual soundtrack workflows as well.