Your character sheet is ready and you need 40 scenes of her. Clicking Generate 40 times is the slow way. Hand your PixAI API key to Muse, tell it what to draw, and the agent submits each task, waits for it, and saves the images for you.
Muse is Meta's AI agent. Since it launched in Japan in October 2026, "Muse × PixAI" setups have been all over Japanese creator communities: people batch-producing illustration sets, running multi-page manga pipelines, and even letting Muse handle LoRA training from dataset to finished model.
This guide covers getting a PixAI API key, wiring it into Muse, prompts you can paste straight into Muse, and the workflows Japanese creators are actually using.
New to PixAI? Spend a few minutes in the web generator first. Once you know which model and style you like, the settings you hand to Muse are easy to pin down.
01 · WHAT IS MUSE
What Muse is, and why it pairs well with PixAI
Muse is a personal AI agent from Meta. You give it instructions in chat, and it plans the steps and carries them out. The big difference from a regular chatbot: Muse has its own computer in the cloud.
CLOUD
Runs in the cloud
Tasks keep going after you close the app. Your own computer doesn't need to stay on.
WORKSPACE
Handles code and files
Its workspace can hold scripts and config files, and it can call outside APIs.
ANYWHERE
Web and mobile
Use it on the web or in the iOS and Android apps. There's a free tier.
CONFIRM
Asks before it acts
It checks with you before sending, buying, or spending credits.
PixAI brings the anime models, such as Tsubaki.3 and Haruka V2, plus community LoRAs. Muse plans and runs the workflow; PixAI draws the pictures. For more on Tsubaki.3 itself, see the Tsubaki.3 launch post.
02 · TWO WAYS
Let Muse click through the website, or let it call the API?
If you want consistent results in bulk, use the API. If you only need a handful of images now and then, having Muse drive the website works too.
Muse drives the website
BROWSER
How it works
Muse opens pixai.art in its cloud browser and clicks buttons the way a person would
What you need
Just a PixAI account, signed in inside Muse's browser
Speed
Slow. Every step means reading the screen and finding the next button
Reliability
Layout changes, pop-ups, and sign-in checks can stall it
Best for
A dozen images now and then, without applying for a key
Muse calls the API (this guide)
RECOMMENDED
How it works
Muse writes a script that sends requests straight to the PixAI API
What you need
A PixAI API key
Speed
Fast. Submit a task, then wait for the result
Reliability
Fixed parameters, so results are easy to reproduce
Best for
Batch runs, manga panels, a series with the same character
One Japanese user who tested the browser route got a dozen-plus images out of free accounts, and was also frank that doing it by hand would have been much faster.
03 · BEFORE YOU START
Step 0: get an API key and a model ID
You need two things: an API key and the version ID (modelVersionId) of the model you want.
Get an API key
Members can create one instantly
With a PixAI membership, click your avatar in the top right → Settings → API Platform in the left menu → Create API key. The key is ready on the spot, so this is the quickest way to start.
① Settings → ② API Platform → ③ Create API key (Japanese UI shown)
The key is shown only once, when you create it. Copy it before you close the dialog and store it somewhere safe, such as a password manager.
This is the only time you'll see the key. Copy it, then click Done
Or apply by email or form
Email api@withpixai.art or fill in the beta test form. Manual review takes up to 5 business days.
Both kinds of key work exactly the same. A key looks like sk-xxxxxx, and it's as sensitive as your password, so never post it anywhere public. The PixAI API docs walk through the application in detail.
Find the model version ID
Open a model's page on PixAI and pick the version you want. The number at the end of the URL is the modelVersionId. Tsubaki.3's, for example, is 2024383379556065549. LoRA version IDs work the same way.
① Pick a version → ② the last number in the URL is the modelVersionId
TIP
Generate a few images by hand on the PixAI website first and settle on a recipe: model, prompt, negative prompt, and LoRA. If Muse starts from scratch, it can churn out a whole batch you don't want and burn through credits. The LoRA weight settings guide helps you find the right strength before you hand things off.
04 · 6 STEPS
Six steps to get Muse generating through the PixAI API
The idea is simple. Put your key in a .env file in Muse's workspace, then have Muse write a small script that calls the PixAI API. These steps use Muse on the web.
Start a chat and a project folder
Open a new chat in Muse on the web and ask: "Create a pixai-batch folder in the workspace with a Sample.env.txt file in it."
Muse creates pixai-batch and Sample.env.txt
Add your key
Muse writes a PIXAI_API_KEY= line into the file. Open Library → System files on the left, find the file, paste your key after the equals sign, and save. Because you never type the key into chat, it won't show up in your conversation history.
Paste your key after PIXAI_API_KEY= and save
Rename it to .env
Ask Muse to rename the file to .env. It may disappear from the web view afterward. That's normal; the file is still there.
Just ask Muse to rename the file to .env
Prepare a prompt list
Create prompts.csv in the pixai-batch folder, with one prompt per row. Write it yourself, or have Muse draft it and review it before use. For Tsubaki.3, descriptive sentences beat piles of tags.
Muse wraps prompts that contain commas in double quotes so each stays on one row
Have Muse write the script
Send Muse Template 1 from the next section. The script submits a task, waits for it, downloads the images, and saves them with numbered filenames.
Before running, Muse shows the task count and fixed parameters for you to check
Confirm, then run
Muse lists the prompts and image count and waits for your OK. Run a single test image first, then the full batch.
Test passed: Muse reports where the file went and what it contains
What the script does behind the scenes
The PixAI API is an asynchronous REST API. Every request goes to https://api.pixai.art with an Authorization: Bearer <key> header.
Submit a task
POST /v2/image/create
Send the model, prompt, and settings. You get a task ID back, with status 201 on success.
Check progress
GET /v1/task/{taskId}
Poll no more than once every 1.5 seconds, or use a webhook instead.
Download images
GET /v1/media/{mediaId}/image
You can also download straight from the finished task's mediaUrls.
The parameters you'll reach for most (full list in the image creation API reference):
modelVersionId
Which model
Taken from the model page URL
prompt / negativePrompt
Prompts
Up to 10,000 characters on Tsubaki.2 and Tsubaki.3
aspectRatio
Aspect ratio
1:1, 2:3, 3:4, 9:16, 16:9, and more
size
Resolution tier
1k or 1.5k
mode
Quality (Tsubaki only)
lite, standard, pro, ultra
batchSize
Images per task
1 or 4
loras
Stack LoRAs
Up to 3 without a membership, 15 max
contextImages
References (Tsubaki.3 only)
Up to 3 images, for character consistency
seed
Fixed seed
When you want to reproduce a result
05 · COPY & PASTE
Copy-paste prompts for Muse
Swap the parts in [brackets] for your own details and send the whole block to Muse.
Template 1: build the batch script
In the pixai-batch folder, write a Python script that batch-generates images with the PixAI API:
1. Read PIXAI_API_KEY from .env. Never print the key.
2. Submit tasks with POST https://api.pixai.art/v2/image/create, sending the header Authorization: Bearer <key>. Use a 120-second timeout, and treat both 200 and 201 as success (a created task returns 201).
3. Fixed parameters: modelVersionId=[model version ID], aspectRatio=[2:3], batchSize=[1], negativePrompt=[negative prompt].
4. When I want a LoRA, add a loras parameter, for example "loras": [{"modelVersionId": "[LoRA version ID]", "weight": [0.7]}]. Leave it out entirely when I don't. Match the field names to the PixAI API docs.
5. Read prompts.csv one row at a time and submit one task per row. As soon as a task is accepted, write "row number + task ID" to tasks.log.
6. If a submission times out, do NOT resubmit automatically (the server may already have created the task, and a retry would charge credits twice). Log the row to failed_tasks.log and tell me.
7. Poll GET https://api.pixai.art/v1/task/{taskId} every 3 seconds. If a task fails, log the reason and move on.
8. Support a --task-id option: given an existing task ID, skip creating a task and just poll and download it.
9. When a task finishes, save the images from outputs.mediaUrls to output/. Pick the file extension from the actual format (WebP, for example) and name files "row_index.ext".
Before generating anything, tell me how many tasks you will submit and wait for my OK. Start by testing row 1 only.Template 2: plan the scenes, then generate
I want a set of [number] illustrations about [story or scene]. Character sheet: [hairstyle, eye color, outfit, and other traits that must stay fixed]. First, write a short scene description and a prompt for each image, in the format that suits the model I'm using, and save them to prompts.csv for me to review. Once I approve, generate them with the script in pixai-batch and send me all the images when you're done.
Template 3: let Muse build a character LoRA
Help me train one SDXL character LoRA of my character on PixAI. Don't start right away. First, check two things with me: 1. My PixAI credit balance, whether I have any free training slots, and the most I'm willing to spend this time: [credit limit]. 2. Using the character sheet I send you, which visual traits must stay fixed and which (outfit, background) can change. Then work through these in order: - Make 30 training images: a mix of close-ups, upper body, and full body, spread across front, three-quarter, profile, and back views. Show me a few samples first. - In PixAI's basic training, choose Character and SDXL, set a trigger word, and write it down. - Before training starts, tell me the credit cost shown on screen. Stop and ask if it's over my limit. Don't buy credits or a membership. - Keep the LoRA private. Don't publish it. - After training, show me the base model, prompt, LoRA strength, and credit cost you plan to use for a test, and generate one test image only after I agree. Finish with a summary: LoRA name, trigger word, matching base model, where it's saved, and total credits used.
You can also hand the prompt-writing in Template 2 to another chat AI and give Muse only the finished CSV. A Japanese user who tried it found that split much faster than leaving everything to Muse.
06 · USE CASES
How Japanese creators are using it
Three patterns stand out: batch-producing scene illustrations, generating multi-page manga, and drawing the same character again and again with their own LoRA.
1. Generate manga automatically
A Japanese creator released an open-source manga kit called MangaAgentsStudio-A. Have Muse clone it into the workspace, set up .env, and say "draw a manga with PixAI." It goes from story and panel layout all the way to finished images. By default it generates one image per panel; switch to one per page to save credits. To build manga inside PixAI alone, see this Tsubaki.3 manga zine workflow.
2. Train a LoRA, then keep drawing the same character
This one has spread widely among Japanese users: Muse turns your character into a LoRA, and PixAI draws the outfits, poses, and action scenes other tools struggled with. Your job is to answer Muse's questions and check the images.
Prepare training images
Give Muse your character sheet and have it make about 30 training images (PixAI can train on as few as 10). Mix close-ups, upper body, and full body, plus front, three-quarter, profile, and back views. Review a few samples first, and drop anything broken or off-model.
Prompt you can send to Muse as-is:
Using the character sheet I uploaded, make LoRA training images of this character with your own image generation: 1. Treat the sheet as the source of truth and keep these traits fixed: [hairstyle, hair color, eye color, body type, etc.]. Outfits and backgrounds can vary; don't use the same one every time. 2. 30 images total. Each is a separate single image, at least 1024px on the long side, one person only, no text or watermarks. 3. Framing: about 10 each of close-ups, upper body, and full body. Angles should include front, three-quarter, profile, and back. 4. Make 4 samples first and wait for my OK before making the rest. 5. When all are done, show them as a contact sheet and flag any that look like a different person or are broken (warped faces, bad hands, and so on). Remove those and fill back up to 30. 6. Save the final 30 to a lora-dataset folder in the workspace, numbered 01–30.
Train it on PixAI
Muse uploads the images, picks the category and base model, sets a trigger word, and starts training. Check the credit cost on screen before it starts, and keep the LoRA private.
Prompt you can send to Muse as-is:
Use the images in the lora-dataset folder to train a LoRA on the PixAI website: 1. Open PixAI's LoRA training page. If it asks me to sign in, stop and tell me; I'll sign in myself. 2. Upload every image in lora-dataset. Set the category to [Character] and the base model to [base model name]. 3. Set the trigger word to [trigger word] and record it exactly. We'll use it as-is when generating later. 4. Keep the LoRA private. Don't publish it. 5. Before you click start, tell me the credit cost shown on screen and my current balance, then wait for my OK. Stop and ask if it's over [credit limit]. Don't buy credits or a membership. 6. When training finishes, tell me the LoRA name, its version ID (the last number in the URL), the trigger word, and the matching base model, and confirm it's still private.
Test one image
Use the same base model you trained on, with the trigger word and a simple outfit, pose, and background. Check whether the face and hair come through. If they don't, adjust the LoRA strength, or compare with and without the LoRA while keeping every other setting the same.
Change outfits and scenes
Once the character holds up, ask Muse for the outfits and scenes you had in mind. Over the API, just put the LoRA version ID in the loras parameter.
You don't need a membership: credits from daily login bonuses and in-app ads can cover training (see every way to earn free credits). Keep some aside for test images. Training starts on the LoRA training page, and the credit cost guide and credit earning guide explain the numbers. For training settings in depth, read Train LoRA on PixAI.
On Tsubaki.3 you can also pass up to 3 reference images through contextImages, which keeps a character consistent without training a LoRA at all. More options are in 3 beginner methods for character consistency.
3. Batch-produce a full illustration set
Lock in your style recipe first, then have Muse generate a dozen to several dozen images in story order. It suits character reference sheets, scene sequences, and expression sets for stickers. Set batchSize to 4 for batch runs; it takes fewer steps than submitting images one by one.
07 · FAQ
Things to know before you run a batch
Waiting tasks per account
10
Polling interval
≥ 1.5 s
Images per task
1 or 4
Are API-generated images public?
No. Anything generated through the API is private and doesn't appear in your account's public generation history.
Are the images stored forever?
No, generated media isn't kept indefinitely. Make sure your script downloads every image to your computer or Muse's workspace.
How many tasks can I submit at once?
Each API account can have up to 10 tasks waiting at a time; tasks that are already running don't count. Keep polling to once every 1.5 seconds or slower. Details are on the limits page.
Which features does the API support?
Text-to-image, inpainting and outpainting, LoRAs, and Tsubaki.3 reference images. Reference Pro, PixAI Edit, and Edit Lite aren't available through the API yet.
Does it use credits?
Yes. Credits are deducted based on the model and quality settings, so run one test image before a batch to see the cost. For the bigger picture, see the PixAI pricing guide.
What if a submission times out?
Don't resubmit right away. Check your credit balance and generation history first: the server may have created the task anyway, and a retry would charge you twice. If you have the task ID, pick it up again with --task-id.
How do I keep my key safe?
Keep it in the .env file and never paste it into chat. Don't upload files containing the key to GitHub or any other public repository.
Anything else?
Follow the terms of use for both Muse and PixAI. If you plan to use the images commercially, check each model's and LoRA's permission settings first.
START WITH ONE ROW
Your first batch starts with one row in prompts.csv
Dial in one recipe you like on PixAI, then hand the repetitive part to Muse. Your time goes to picking the best images and thinking up the next idea.
Questions about the API? Ask in the #api-beta-test channel on the PixAI Discord.