CharactersAPEX workspace
Tools
/
-
U
No character selected
Pick one from the list, or start a new one.
+
Character
New Character
● Unsaved
More
Identity
Paste from lacpanel. Read-only after creation.
Realistic
Anime
Furry
Mystic
Selects the image pipeline/model family for this character.
Affects grammatical gender in prompts. Defaults to female if unset.
Personality
AI Enhance
Sharpest natural-feel lever. The agent and humanizer both read this separately to keep her voice consistent — e.g. Wednesday is "dry, deadpan, clipped, no Gen-Z slang", Scarlett is "casual high-energy, lots of lol/omg, contractions, fragments". Specific vocabulary > vague adjectives.
Fed to the agent as concrete voice/style samples. Useful when communication_style is hard to pin down in words — drop 3-5 real-feeling exchanges and let the LLM imitate.
Free-form. Appended to the personality description in prompts.
Used only when adult content is enabled. If empty, runtime falls back to the SFW communication style.
Used only for NSFW dialogue. Keep it explicit about what this persona allows, wants, refuses, and how fast intimacy should escalate.
Settings
Who can chat with this character on the site.
Allowed values depend on Sex.
Tags
Click chips to toggle. The site shows characters under matching filter pills. To add or rename a tag, use Tools → Manage tags.
Required by the Coming soon tag. 1–336 hours (14 days max). The site starts the countdown when the character goes live there and opens the chat to everyone once it runs out — nothing to switch off by hand.
Visibility
Where this character is shown. All envs require cover, video, voice, reference image and personality. Prod additionally requires the vision template.
Reference Image
No reference image
Info
Vision Template
-
Created
-
Updated
-
Delete character
Permanently removes this character from every environment.
Cover & Video
Cover image and a short video for the card on the site. The cover must match the first video frame — otherwise the card will "jump" on hover.
SFW
No cover
No video
NSFW
No cover
No video
Limit: 50 MB per video. Formats: png/jpg/webp for covers, mp4/webm/mov for video.
Description
6 variants: mainstream/adult × card-mobile / card-web / profile. English is the source; other languages can be auto-translated. Limits: card-web ≤ 77, card-mobile ≤ 46, profile — no limit (counter shown for awareness).
Mainstream / Adult
The selector applies to starter messages, user suggestions, and starter media. Each kind has its own pool; an empty pool falls through to LLM (no cross-kind blending). Starter request binds the chat to adult/mainstream once; later runtime reads that chat-bound mode.
Starter messages
One is picked at random when a chat starts. Fetish/pose tabs override Default. If nothing is filled in for the language — fallback to LLM.
User suggestions (first 3 user replies)
3 reply suggestions for the user, surfaced before they send their first message. The first 3 lines of the selected language/fetish go to the site; fewer than 3 falls back to LLM. After the user's first reply we always switch back to LLM.
Starter media
Photos and videos share one pool — one item is picked at random when a chat starts, and the backend can ask for a specific fetish or pose. Fetish/pose tabs override Default. If nothing is set and the chat has no fetish — fallback to the legacy cache lookup.
png, jpg, webp · mp4, webm, mov
Send a message to start testing this character's personality
Training history
Reference & generation settings Upload a reference and describe the character.
One waist-up photo → ~100 turnaround frames. Eyes, hair color, hair length and breasts are required; everything else can stay — and is omitted from captions.
Waist-up photo
Generated dataset
Generation, quality review and culling stay together here.
—
—
Similarity is scored by a VLM that fails open — a scoring outage shows as a run of suspiciously perfect 100s.
Click a tile to open it full-size; ← → to move, D to drop, R to retry, Esc to close.
Training identity & captions Trigger, appearance tags and caption preview.
These are baked into every caption ahead of the pose/framing tags, and protected from shuffling during training. Editing them here rewrites the captions at submit time — the stored captions are never modified, so you can retrain with a different trigger.
Unique token the LoRA binds to. Mangle it (
b1@nk4, 4k1k0, !a1ann@) so it doesn't collide with a real CLIP token.Changing this after generation only affects how captions are re-headed, not the stored text.
Comma-separated. Any number of tags — the trainer derives its keep-tokens count from this list.
Pick a run with generated images to preview.
Training settings Choose the model and review the run before starting.
Hands the reviewed dataset to avatar-lora-service, which trains on RunPod Serverless and uploads the LoRA to HuggingFace.
—
Goes into the artifact filename on HuggingFace.
Optional class token paired with the trigger. Leave “—” to train on the trigger only.
—
Advanced training parameters
Rendered from the selected preset, and validated against the same schema on the service side — a field here can never be something training would reject. Only values you change are sent; everything else uses the preset default.
Start training above, or choose an earlier run from History.
—
—
Rendered by the trainer as it goes, so you can watch the likeness converge and stop a bad run early.
◇
No trained artifact yet
Finish training first. Model-pack promotion and validation will appear here without moving training progress out of its page.
Artifact ready
The next step promotes this LoRA to the model packs, then enables repeated one-image tests with different prompts.
HuggingFace artifact
1 · Prepare model packs Not started
Copy the trained LoRA to the model packs used by the realistic or anime Vast worker.
2 · Generate test images
Diagnostic log Show
3 · Use in Studio
Generate a test image, check the likeness, then enable the avatar in Apex Studio.
—
Chat packs
Bundles of pregenerated media delivered in chats (teaser message + pack). Only published packs are ever sent; delivery walks packs in order and never repeats one in the same chat. Паки для галереи — отдельный раздел ниже.
New pack
Пусто = дефолт consumer-стороны
Gallery packs
Платные паки в галерее персонажа — отдельно от чатовых паков выше. Собираются здесь и уезжают на backend по кнопке; правки сохраняются локально, пока не нажать «Обновить». Порядок (↑↓), «Опубликовать/Скрыть», «Media Gallery» и «Private Content» уходят сразу — они не переотправляют состав. Купившие пак остаются с тем содержимым, которое купили: правки на них не влияют.
Куда попадает пак — два независимых флажка, по умолчанию оба выключены: Media Gallery — галерея самого персонажа, Private Content — общая витрина платных паков по всем персонажам. Можно выбрать одно, другое или оба; без единого флажка опубликованный пак не показывается нигде. Позиция на витрине не связана с порядком в галерее; пусто = в конец ленты. Показываются только опубликованные и отправленные паки.
Куда попадает пак — два независимых флажка, по умолчанию оба выключены: Media Gallery — галерея самого персонажа, Private Content — общая витрина платных паков по всем персонажам. Можно выбрать одно, другое или оба; без единого флажка опубликованный пак не показывается нигде. Позиция на витрине не связана с порядком в галерее; пусто = в конец ленты. Показываются только опубликованные и отправленные паки.
New gallery pack
Уезжает как
name.en. Пустое имя backend не примет.В коинах, целое ≥ 0
Library
Curated pregenerated media for this character. «Import from chats» browses liked / CV-approved chat media straight from the DB (no re-upload). «Upload files» accepts files from the export zip too — they resolve to their existing R2 objects by filename, so only genuinely new files get uploaded.
No assets yet. Import from chats or upload files.