Goal: one original character, tested end to end through an AI VTuber generator workflow, output as a five-asset starter kit. This is a build log, not a promo piece. Prompts, corrections, and failure points included.
Scope: still images only. Avatar, four expressions, a stream screen, a profile icon, and a channel banner. No rig, no tracking, no live streaming.
The complete VTuber Starter Kit, built from one persona
What You Actually Get from an AI VTuber Generator
Still images. Not a working streamer. This produces avatar art, four expressions, a stream screen, plus a profile icon and banner derived from those.
No rigged Live2D or 3D model, no face tracking, no real-time streaming. PixAI's own FAQ confirms this: you still need a rig in something like VTube Studio to go live.
It's still useful as a pipeline step. A VTuber avatar maker like this gets you a persona to test and post, and gives a rigger a visual reference before you commit budget to a full rig.
Here's the step-by-step build using an AI VTuber avatar generator.
Step 1: Persona Setup and First Generation with Tsubaki.3
Opened PixAI's generator, selected Tsubaki.3. Before touching the prompt box: one hairstyle, one outfit, two accessories, nothing else. Simple beats interesting. The avatar has to survive a dozen more generations after this.
Persona:
An original female VTuber with a chin-length silver-lavender bob with blunt bangs and light mint-green eyes, wearing an oversized zip-up hoodie with the hood down in navy blue with coral details, with white over-ear headphones resting around her neck and a small black crescent-moon hair clip in her bangs.
Worked from PixAI's own prompt guide, which asks for a plain description rather than a tag list.
Filled the same details into the base prompt template, generated the first direction with Prompt Helper switched off. Off, because it can translate or rewrite the prompt before generation, and I wanted the prompt below to match exactly what ran.
Prompt for Direction A
An original female VTuber character has a chin-length silver-lavender bob with blunt bangs and light mint-green eyes. The character wears an oversized zip-up hoodie with the hood down in navy blue with coral details, along with white over-ear headphones resting around her neck and a small black crescent-moon hair clip in her bangs. The character is alone in front of a plain light-gray background, facing forward with a neutral expression. Show a waist-up portrait at eye level, with both eyes, the full hairstyle, and the signature accessories clearly visible. Keep the hands outside the frame. Use soft, even lighting and an anime illustration style with clean line art and simple cel shading. The image contains no text or logos.
Tsubaki.3 selected, with the tested Direction A prompt and Prompt Helper switched off
Direction A is the base design. B swaps the hoodie for a sailor-style blouse. C keeps the hoodie, swaps the palette to burgundy and mustard-yellow.
Prompt for Direction B
An original female VTuber character has a chin-length silver-lavender bob with blunt bangs and light mint-green eyes. The character wears a navy blue sailor-style blouse with a wide square collar, with coral trim along the collar edges and a coral ribbon bow at the collar, along with white over-ear headphones resting around her neck and a small black crescent-moon hair clip in her bangs. The character is alone in front of a plain light-gray background, facing forward with a neutral expression. Show a waist-up portrait at eye level, using the same camera distance and framing as Direction A, with no skirt or lower torso visible. Both eyes, the full hairstyle, and the signature accessories should remain clearly visible. Keep the hands outside the frame. Use soft, even lighting and an anime illustration style with clean line art and simple cel shading. The image contains no text or logos.
Direction B took three attempts (two corrections) to land. Breakdown:
Attempt 1 rendered the blouse mostly white instead of navy, coral only on the collar trim, though the crop already matched Direction A's framing.
Attempt 2: rewrote the color clause to name the garment directly, asked for the crop to end above the waist. Fixed the blouse color, but the new crop overcorrected. It came out tighter than Direction A, with a strip of midriff showing.
Attempt 3: replaced "medium portrait" with "waist-up portrait," anchored the framing to Direction A. Result: navy blouse, coral trim and bow, framed like A.
Direction B across the two correction passes
Prompt for Direction C
An original female VTuber character has a chin-length silver-lavender bob with blunt bangs and light mint-green eyes. The character wears an oversized zip-up hoodie with the hood down in burgundy with mustard-yellow details, along with white over-ear headphones resting around her neck and a small black crescent-moon hair clip in her bangs. The character is alone in front of a plain light-gray background, facing forward with a neutral expression. Show a waist-up portrait at eye level, with both eyes, the full hairstyle, and the signature accessories clearly visible. Keep the hands outside the frame. Use soft, even lighting and an anime illustration style with clean line art and simple cel shading. The image contains no text or logos.
The three tested directions, compared side by side
A and C needed zero correction. C kept the hoodie design, only shifted the palette.
Across all three: hairstyle, eye color, and both signature accessories, the headphones and crescent-moon clip, held consistent. Picked Direction A as the master avatar. Clean on the first attempt, eyes read clearly, hoodie and accessories fully legible.
The master avatar, carried into every generation from here
Step 2: Reference Locking to Stop Character Drift
A good avatar only matters if it survives change. Loaded the master into Tsubaki.3 as a reference (@image1), changed one thing at a time.
The pattern: describe the requested change, use the reference to preserve everything else. Both tests below started from the master, not from each other's output.
Prompt for Pose test
@image1. Turn the character's shoulders slightly to a three-quarter angle while keeping the face looking toward the viewer. Keep the background and framing unchanged. Keep the character's appearance exactly as in the reference.
Interface note: with a reference loaded, the detail panel shows a rewritten version of the prompt. What's in this article is what I typed.
Master avatar loaded as the reference, with the pose test prompt visible
Shoulders turned three-quarter, face stayed toward viewer. Hairstyle, eye color, clip, headphones, hoodie all matched the master; background and framing unchanged. No correction needed.
Prompt for Background test
@image1. Change only the background to a plain muted-blue backdrop. Keep the pose and framing unchanged. Keep the character's appearance exactly as in the reference.
Background changed to muted blue, pose and framing held. Hairstyle, eyes, clip, headphones matched. The coral sleeve markings didn't: sharp branching lines on the master came back as soft blurred patches, some coral shifted toward purple.
Fix: added one clause naming the object directly.
@image1. Change only the background to a plain muted-blue backdrop. Keep the pose and framing unchanged. Keep the character's appearance exactly as in the reference, including the sharp coral branch-like markings along both sleeves.
Background test before and after adding the sleeve marking clarification
Retry brought back sharp coral markings on both sleeves. Naming the specific detail beat the generic "keep the appearance" instruction.
The master beside the pose and background variants
Character stayed recognizable across both tests. One drift point, the sleeve markings, fixed with a single named correction. A reference doesn't guarantee every detail holds. I checked each result against the master before moving on.
For consistency across a full image set, PixAI has a longer guide on it.
Step 3: Generating the Expression Set
A stream needs more than one face. Generated four expressions from the master: neutral, happy, surprised, sleepy. Each used the master as reference (@image1), started with a short instruction in Tsubaki.3. Sleepy was my own call. Hoodie and headphones read as a late-night streamer.
Prompt for Neutral
@image1. neutral.
Prompt for Happy
@image1. happy.
Prompt for Surprised
@image1. surprised.
Prompt for Sleepy, first attempt
@image1. sleepy.
The four expressions generated from the master
Neutral, happy, surprised worked first try. Happy: open smile, curved eyes, blush. Surprised: wide eyes, raised brows, small open mouth, mint-green eyes read clearly. All three matched the master on hairstyle, clip, headphones, hoodie, background.
Sleepy needed a second pass. First result had fully closed eyes. Read as sleeping, not sleepy, and the eye color was unverifiable.
Added one clause:
@image1. sleepy, with heavy half-open eyelids
Sleepy expression before and after adding the eyelid clarification
Retry gave a drowsy look, lowered lids, a sliver of pale green visible. Short instructions worked for three of four expressions; the fourth needed more description.
This covers four still images, not a full turnaround sheet. That's a different job, covered in PixAI's Character Asset Kit guide.
Step 4: Stream Screen Output via the AI VTuber Generator
Opened the AI VTuber generator, uploaded the master avatar. Single text box, example line instead of separate name and language fields. Test input:
Character's name is Avatar Test, response language is English
PixAI markets this as a free AI VTuber generator running on daily credits. Generate button showed a cost of 8,000 credits; balance dropped from 712,400 to 704,400 after the run.
The AI VTuber generator with the master avatar uploaded and the test inputs entered
Ran it once. Output: a single still image in a phone-style live-stream layout. Header reading "Avatar Test" with a LIVE badge and viewer count, bedroom backdrop with string lights, floating hearts, chat panel with comment bar at the bottom.
The generated stream screen
Character matched the master. Hairstyle, mint-green eyes, crescent-moon clip, headphones, hoodie all carried over, hands still in pockets. Hood lining and coral sleeve details read slightly darker and cooler under the room lighting. Small header avatar is a headshot of the same character.
Frame keeps the character clear, though the chat panel covers the lower hoodie. At full size, "Avatar Test" and the chat messages were legible in the original file. PixAI's FAQ says small generated text is usually not legible, so this run outperformed that expectation. Still a generated image, not a working chat.
Chat panel cropped from the original output
No defect requiring correction on this one. Tool's FAQ says it outputs still images, not a rigged model. That matches what I got.
Step 5: Kit Assembly
Everything up to here was generated. This step is assembly: profile icon, channel banner, combined layout built by hand in Canva. Manual edits of existing generated images, not PixAI exports or editing controls.
The complete VTuber Starter Kit, built from one persona
Kit has five parts: master avatar, four expressions, stream screen, profile icon, channel banner. Icon: square crop of the master, keeps face, crescent-moon clip, and top of headphones in frame. Banner: master on the right of a wide canvas, plain pale-gray background, left side open.
Persona holds across all five parts. The bob, mint-green eyes, clip, headphones, navy hoodie with coral markings stay consistent throughout. Sleepy expression is the outlier, since lowered lids hide most of the eye color, but the rest of the character holds.
Expression set came out strongest. Four faces easy to distinguish at a glance, hairstyle, hoodie, headphones and background consistent across the set. Banner is the simplest asset: pale background leaves room for channel branding, master avatar anchors the right side.
Total corrections across the workflow: four. Two on Direction B's outfit and framing, one on the sleeve markings in the background test, one on the sleepy eyelids. No new persona needed, no rebuilt prompt. Stream screen took a single run.
Handoff Notes for a Rigger
Handing this to a rigger, send:
The master avatar at the highest resolution you have
The expression set
A front-facing full-body image, prepared later
A note on non-negotiable details: the chin-length silver-lavender bob with blunt bangs, the mint-green eyes, the crescent-moon clip, the white headphones, the navy hoodie with coral sleeve markings
These are design references for an artist, not rigging-ready files.
Summary
The route: one persona, a Tsubaki.3 avatar, reference-based variants, an expression set, a stream screen from the AI VTuber generator, a Starter Kit.
Reusing one master as the reference kept the character consistent across the still images. Not flawless. Sleeve markings softened once, sleepy needed a second pass, but both fixed with a small, specific clarification. Banner stayed plain, chat panel covers part of the hoodie in the stream screen.
None of this replaces a rigged model. It gets you a face to test and post while you decide whether and when to commission a rig. If you want to see what your own character looks like on a stream screen, try the AI VTuber generator with a character you already have.
Top comments (0)