| {{ t('chatId') }} | {{ t('role') }} | {{ t('model') }} | {{ t('updatedAt') }} | {{ t('actions') }} |
|---|---|---|---|---|
| {{ s.chat_id }}{{ s._label }} | {{ s.role_name }} | {{ s.model || '-' }} | {{ s.updated_at }} |
{{ t('noData') }}
| {{ t('id') }} | {{ t('chat') }} | {{ t('role') }} | Role | {{ t('content') }} | {{ t('time') }} | {{ t('actions') }} | |
|---|---|---|---|---|---|---|---|
| {{ m.id }} | {{ m.chat_id }} | {{ m.role_name }} | {{ m.role }} | {{ m.content?.substring(0, 120) }} | {{ m.created_at }} |
{{ t('noData') }}
| {{ t('chatId') }} | {{ t('role') }} | {{ t('msgCount') }} | {{ t('summaryPreview') }} | {{ t('updatedAt') }} | {{ t('actions') }} |
|---|---|---|---|---|---|
| {{ s.chat_id }} | {{ s.role_name }} | {{ s.message_count }} | {{ s.summary?.substring(0, 100) }} | {{ s.updated_at }} |
{{ t('noData') }}
Image/video models are chosen by the JailBreaker pools above — there is no separate default model.
{{ typeGroup.label }} {{ typeGroup.count }} models
{{ runtimeGroup.label }} {{ runtimeGroup.count }}
{{ platformGroup.label }} {{ platformGroup.count }}
{{ m.id }}
{{ field.description }}
No Media Models
{{ p.name }}
No LLM Providers
{{ llmModal.isNew ? 'Add Provider' : 'Edit ' + llmModal.data.name }}
{{ t('renameHint') }}
{{ roles.length }} roles
{{ r.rolename }}
{{ r.persona?.trim() }}
No roles
{{ roleEdit.rolename }}
Default Role · Unsaved{{ vnEffective.text }}
Identity
card blurb + cast summaryHer look
appearance — image promptsYour role
the interlocutorVoice
TTS + live avatar{{ voiceOptions.length }} online voices Not loaded —
{{ minimaxVoiceOptions.length }} voices bundled Not loaded —
System prompt
persona base · anti-censorship CTA + reply formatScenes
story cards · sent to the role LLM on start{{ scenesError }}
Brainstorm a character (the assistant hands you a card — click "Fill form"), or ask it to read / change this role's config and scenes: it reads the real fields, proposes a diff card, and only writes after you click Apply.
{{ p.summary }}
⚠ {{ w }}
new value
Role card incomplete (missing {{ assistMissingIn(assistCardOf(m)).join(', ') }}), ask the assistant to fill it in
Assistant is typing…
Couldn't read “{{ er.name }}”: {{ er.reason }}
Edit the assistant's personality / anti-censorship rules (saved to role_assist_persona.md); applies on the next conversation. Keep @@fields output-format notes.
icon = avatar (1:1); ref = reference sheet (16:9, four face crops + full body, anchors the face to icon); ref_user = reference sheet from user_appearance; voice_ref = manual upload only. About 1-2 minutes per image.
{{ heroPrompt }}
{{ livePrompt }}
{{ a.note }} · {{ assetMetaText(a.kind) }}
{{ genStatus[a.kind].error }}
Generation complete
System Prompt JailBreaker
Global uncensored policy prepended to the system message of EVERY LLM request, for all models & providers. Cached as a constant prefix (only the first call pays the tokens).
Media JailBreaker
Order = priority. All of these failing falls through to the Full Uncensored pool.
Order = priority. Celebrity scenes use this pool only (Ban Celebrity models would reject them).
Order = priority (falls back on failure). Used whenever an image is generated without an explicit model — video first frames, scene plates and role assets keep their own models. Candidates = enabled models whose prototype operation is image_edit.
Order = priority (falls back on failure). Used when a video scene has NO reference image at all (the character is not in frame) — every R2V model needs one, so those scenes must go to a pure text-to-video endpoint. Candidates = enabled models whose prototype operation is text_to_video.
LLM for scene classification (also reused by the NSFW asset classifier)
LLM that rewrites the director brief into the final video / image prompt (video rewriter + T2V rewriter). Pick a low-guardrail provider (e.g. Deepseek-Flash / Venice) — high-guardrail models refuse explicit briefs. On refusal the rewriter retries once with the fallback provider, then falls back to the raw brief.
[celebrity?] any cast role with is_celebrity = true
├─ yes → Full Uncensored pool (in order)
└─ no → Ban Celebrity pool, then Full Uncensored pool
[chain] first model that succeeds wins
[image] no explicit model → Uncensored Image Pool (in order)
Aspect Ratio Auto
LLM outputs portrait/landscape → auto-map to image size & video ratio
Image size follows each model's own size enum (ratio-matched). Chat photos use the chat resolution below; role assets (icon/ref) and scene plates keep the model default (2K). Video ratio auto-maps: portrait→9:16, landscape→16:9.
Scene Plate
{{ t('desc_scene_plate') }}
Lightweight LLM for location routing & env distillation
Scene state distillation LLM (character wardrobe/location continuity). Choose an uncensored provider for NSFW scenes.
LRU eviction — least recently used plates are removed beyond this limit
Diptych plate render model (gpt-image-2 recommended)
Diptych, width x height (21:9 recommended)
[route] same location → reuse existing plate
└─ new location → generate clean plate (one extra image gen)
[inject] plate as last reference image + anchor prompt
fixed: structure/layout/props/palette/lighting · variable: angle/framing/detail states
{{ t('sp_locations') }}
Scene States
Distilled per-role scene state injected every turn as continuity baseline. Delete to reset a role's baseline — it rebuilds on the next media generation.
Pricing (cents per use)
📖 Guide
What it is for:Set the unit price of the four billable actions, incents (100 cents = $1). Click "Save pricing"; it reaches the whole site within 60 seconds.
- Chat: one charge per message the user sends (including group replies). Reference cost: DeepSeek Flash about 0.1-1 cent each.
- Image: one charge per generated image. Reference cost: Seedream about 3-5 cents each.
- Video: one charge per generated video. Reference cost: Venice 480P from about 30 cents (the priciest — do not set it too low).
- Membership 30d: the "ticket" price for BYOK users (having your own key still requires membership; effective in P6b).
Billing rules:UserBalance > 0 the table above applies per action; once the balance hitsis 0 it falls back to the free daily quota (trial tier). Admins and BYOK users are not charged.
Refund on failure:Media generation charges up front; if it failsautomatic full refund (a refund entry appears in the ledger), so users never pay for nothing.
Suggested pricing:Set it at cost ×1.5 to leave room for platform margin and failed retries.
Top-up code
📖 Guide
What it is for:Generate codes to sell (top-up / membership); buyers redeem them in playground. Good for offline payment (WeChat/Xianyu/DM) then sending the code.
Relation to USDT auto top-up:If you set USDT_TRC20_ADDRESS; users pick a tier in Settings to auto-generate a "unique remainder amount" (e.g. 20.37 USDT); about a minute after transfer it credits automatically — no code needed. Codes are for buyers without a crypto channel, or for gifting/promos.
Two kinds:
- Points card: enter the face value incents (500 = $5). Redeeming adds $5; later generation is charged per the pricing table.
- Membership card: enter the face value inDays (30 = 30 days of membership). Redeeming grants membership directly (even BYOK users need membership).
Three parameters:
- Count: how many times it can be redeemed. 1 = single-use; 10 = the same code can be sold to 10 people (campaign code).
- Valid days: the code itselfvalidity (blank/0 = forever). Unused codes expire.
- Note: for your own reference, e.g. "$10 giveaway".
Full flow:① Generate → ② send the code to the buyer → ③ they open playground, click their avatar top-right → Settings → Account → paste the code → Redeem → ④ balance/membership lands instantly (page refreshes).
Revoke:Clicking "Revoke" voids the code immediately; unredeemed ones stop working (already redeemed ones are not clawed back).
Pricing reference:Common tiers $5 / $10 / $20; for USDT collection convert at the day's rate and enter cents.
USDT auto top-up (TRC20)
📖 Guide (including how to get a deposit address)
How it works:the user picks a tier in playground → Settings and the system generates aAmount with a unique remainder (e.g. 5.37 USDT — the remainder is the order id); the user sends it to your TRC20 address and the backend scans the chain every 60s, crediting the balance on a match.
How to get a deposit address (pick one):
- Self-custody wallet (recommended): install TronLink / TokenPocket / OKX Web3 on your phone → create a wallet → Hand-write the seed phrase (never screenshot or cloud-save it) → choose the TRON network → copy the T-address → put it in .env.
- exchange deposit address: register and pass KYC on Binance/OKX → deposit USDT via TRON (TRC20) → copy the address. Downside: KYC and risk controls on frequent small deposits.
Three hard rules:① the user picks an TRC20/TRON network (ERC20 transfers will not arrive); ② test small before scaling; ③ never leak the seed phrase or private key (this system only uses the address).
How to test without USDT:no real money needed — have the test account create an intent in Settings, then click its "Mark as received", the system credits the intended amount; the whole chain (intent → credit → balance → charge) is auditable.
"Mark as received" is also an operational must:For wrong amounts, transfer-before-order, or missed auto-matches: credit the order manually.
Optional:Apply for a free TRONGRID_API_KEY at console.trongrid.io and put it in .env to raise the on-chain query rate limit.
Invite code
📖 Guide
What it is for:registration gate. This siteRegistration requires an invite code— used to control volume, block scrapers and gate entry on "18 or older".
Three parameters:
- Count: how many people can register with it. 1 = single invite; 20 = promo code, the first 20 users can register.
- Valid days: code validity (blank/0 = forever). Unused parts expire.
- Note: for your own reference, e.g. "Twitter campaign batch 1".
Full flow:① Generate → ② send the code to the invitee → ③ they open playground (site home) → at the login wall click "No account? Sign up" → fill email/password/invite code and tick 18+/ToS → ④ registered; sign in and chat.
Revoke:Clicking "Revoke" voids the code immediately; unused seats stop working (already registered accounts are unaffected).
Difference from top-up codes:invite codes handle "Can they get in", top-up codes handle "Can they afford to play". New users start on the trial tier (daily free quota); top up or subscribe for unlimited use.
User
📖 Guide
What it is for:All registered users; adjust balance, quota and ban status.
Per-line fields:
- Email: login name (used to sign in to playground).
- Balance $X: while >$0 the user is charged per the pricing table; at $0 it falls back to the free daily quota.
- Member until: how long the BYOK ticket lasts (BYOK users still need membership; effective in P6b).
- Today's usage: today's usage count for each of the three actions.Free-quota tier only; paid charges are not counted.
- chat_id: this account's memory isolation key (assigned automatically; never shared across users). The tg suffix is the bound Telegram id (P4).
Quota (three boxes on the left):
- blank = use the global default (currently Chat {{ defaultQuotas.chat }} / Image {{ defaultQuotas.image }} / Video {{ defaultQuotas.video }} per day).
- 0 = this user's absoluteUnlimited (gifting / your own alt).
- positive = per-user daily caps (e.g. relax a heavy user to 500/50/10).
- Click "Save" to apply.
Adjust balance (right side):Enter cents, positive = credit, negative = debit. e.g. 500 = credit $5, -200 = debit $2. Every adjustment is logged (kind=admin_adjust, visible in the ledger). Use it for manual compensation, refunds and gifts.
Ban / unban:A banned user loses access immediately (their token expires within 60s), but all data and history are kept; unbanning restores access. Admin accounts cannot be banned.