Local, uncensored AI roleplay — with story-consistent image generation.
The same character. The right scene. Turn after turn.
One character, four framings, same face — the hard part of AI roleplay, solved. An open-world, uncensored RPG with unbound AI narration and a storybook you keep on your own disk. Runs on your card. No subscription. No ID upload.
Bring your whole roster. Fervyn imports your TavernCard V2/V3 and chub.ai cards — a PNG, a JSON, or a chub.ai / JanitorAI / HuggingFace link — carrying the name, persona, scenario, greeting, example messages, system prompt, tags, and your lorebook (up to 100 entries). The card's own portrait — when it's a PNG — becomes its avatar; every new scene can render from a synthetic face you pick and pin — a pinned face stays consistent scene to scene. No ID upload. No face scan. Unlike where you came from.
Persistent character faces
A storybook you keep
Verifiable privacy
Free · Windows · NVIDIA 8 GB minimum — your card sets its config (detection settles after first launch) · installer ~165 MB, then a one-time ~28 GB model download at the 8 GB floor, more on bigger cards (resumes if interrupted) · nothing else to install · unsigned — how to verify → · SHA-256
Don't take our word for it. Look at the panels.
Most AI roleplay apps bolt an image generator onto a chat box. Fervyn was built the other way around — so the same character looks like the same character from the first panel to the last, scene after scene.
Same character. Panel 1 and panel 4. Same face — curated from multiple takes, the way you would pick panels in-app.
The render matches what the story just said.
A scene extractor reads the prose and renders what was actually written — the setting, who's in the room, what's happening. The picture follows the story, not a static prompt.
"The archery grounds lie fallow in winter — the target-butts wrapped in canvas, the bow-rack dusty, frost forming on the stone steps leading down to the practice area."
"By nightfall, the Upper Lake has been transformed — a thousand floating lanterns drift on its surface… The high balcony overlooks all of it."
"Lihua crawls across the lacquered floor with the robe slung low over one shoulder, eyes front…"
Each render above carries the burned-in AI-generated badge the pipeline applies to every image it makes.
Uncensored by design.
Fervyn is an open-world harem RPG with unbound, AI-driven narration — it ships uncensored image and language models because adult roleplay is the point, not a bolt-on. No filter walls, no "content policy" refusals, no railroaded story. The engine runs on your machine, so the only rules are yours — and you own every panel it makes.
18+. One hard line, non-negotiable: Fervyn permanently blocks any sexual content involving minors — enforced by a prompt-side filter and a post-generation image check that run by default and cannot be turned off. Everything else is yours to direct.
Your models. Your settings. Your machine.
Fervyn is an open engine, not a black box. It runs the local tools you already have, and lets you swap in your own.
Chat: auto-detects KoboldCPP, Ollama, or LM Studio running on your machine — or bring your own GGUF and drop it into any of them; Fervyn picks up whatever you have installed. Sampling is tuned automatically per scenario and model, with per-character sampler presets in the Character editor; response length and context size are editable under Settings. The uncensored chat path makes zero outbound calls.
The shipped chat model scales to your card — TheDrummer's Rocinante-X 12B on 8-to-16 GB cards (IQ3 quant at 8 GB, Q4 above) and Cydonia-24B from 24 GB up (co-resident with the image model at 32 GB), plus Skyfall-31B as an optional max-quality pick on 32 GB. Older GTX-class cards fall back to a 3B. Every one is swappable — see writing quality by GPU tier.
Image:ComfyUI with SDXL under the hood. Drop your own checkpoints and LoRAs into the standard ComfyUI model folders (Civitai is a first-class source), or open and edit the workflow — it's a plain JSON file on your disk.
Consistent faces: character identity is held by an IP-Adapter face conditioner on a fully permissive licence — Apache-2.0 weights, an MIT image encoder, and no biometric face-scan at render. That's the honest engineering reason the same face holds render after render.
Every scene becomes a panel you keep.
Your panels live on your disk — every render, plus a curated picture-book view — and there's no platform that can delete them. And you can export the whole thing: the prose, every image, and a verifiable provenance sidecar on each panel.
Exports as a portable bundle — the prose plus every image, each carrying a per-panel C2PA Content Credential (a tamper-evident provenance signature, on an open industry standard).
Every panel you export is signed — open one in any Content Credentials verifier. The signature is written and checked entirely on your machine: it records that the panel was made in Fervyn, and never phones home. A panel that cannot be signed is labelled unsigned in the export — reported honestly, never faked.
Each frame is a single woman, kept consistent across her world's setting and framing. The hard part — the same face, frame after frame — is the part that's solved.
The Inner Palace 內廷Topkapi — The Felicity WithinCrowns and CoronetsThe Convivium of CaesarThe Brothel of Seven ColoursYoshiwara — Floating World
…and there's a whole cast of worlds to explore — imperial China, an Ottoman seraglio, a European court, late-Republic Rome, a modern luxury house, Edo's floating world. Period-correct settings and attire, rendered to match.
Verify us. Don't trust us.
The uncensored chat path makes zero outbound network calls — and a continuous-integration (CI) test fails the build if any code path tries to phone home. Fair question for any adult installer from a small project — so here's what you can check today.
Scope, plainly: the zero-outbound guarantee covers the uncensored chat path. Optional, clearly-labelled features that are off by default may use the network — never on this path.
Honest signing note: the picture-book panels are C2PA-signed and the installer ships with a published hash you can verify. We don't claim the installer itself is code-signed — verify it by the published SHA-256 instead. Because it isn't code-signed, Windows SmartScreen will likely show an "unknown publisher" warning — that's what unsigned looks like, not a malware verdict. Verify the SHA-256 first (or upload the file to VirusTotal, which runs 70+ scanners), then choose More info → Run anyway.
Built in the open. Shaped by you.
Fervyn is built by one person — and steered by the people who actually run it. Tell us what to build next, get help installing, and report what broke — the anonymous feedback form takes all three, no account needed.
#showcase — what people are generating (work-safe + adult rooms).
#help — install + GPU questions, answered fast.
#feature-requests — vote on what ships next.
#bug-reports — tell us what broke; we read every one (include your GPU + the installer build).
No account? A couple of ways in without one: the anonymous feedback form (no sign-in, your IP is never stored), or the public roadmap — Now / Next / Later.
Runs on your card. Scales with it.
Fervyn detects your VRAM — the memory on your graphics card — and picks an image workflow for your hardware automatically. The free installer runs on an 8 GB card and scales up automatically.
Consistent characters and a storybook you keep on disk. The picture-book keeps the story's cover panel at 8 GB; it grows to six panels per story at 12 GB and up.
12 GB (RTX 3060 12GB / 4070 / 5070)
The full-quality fp16 image engine plus a higher-quality build of the chat model — the biggest single step up in the ladder.
16 GB (RTX 4080 / 5080 / 5060 Ti 16GB)
The same full-quality image engine with headroom to spare — on the Speed preset, chat and pictures stay loaded together (no waiting).
24 GB+ (RTX 3090/4090/5090)
The biggest chat model in the ladder driving the same image engine; at 32 GB, chat and image stay loaded together, with no wait between a reply and its picture.
The installer is a free download. There's nothing to buy to use what you install.
Does it run locally or in the cloud?
On your machine. First launch downloads the models for your hardware — ~28 GB for the 8 GB config, ~32 GB at 12–16 GB, ~38 GB at 24 GB and up (runtime stack + models) — and the download resumes if your connection drops. After that, the game runs locally. Budget ~34 GB free at 8 GB, ~38 GB at 12–16 GB, ~45 GB at 24 GB+. Full flow: installation guide.
How do updates work?
New builds are announced on the changelog (RSS available). Updating means running the new installer — your characters, chats, and picture-books live in your user data folder and survive reinstall. Nothing updates itself behind your back.
Do I need an account or to upload ID?
No account, and no ID upload, to use what you install.
Can I import my existing characters?
Yes. Drop in a TavernCard V2/V3 or chub.ai card — as a PNG, a JSON, or a chub.ai / JanitorAI / HuggingFace link. Fervyn carries the name, persona, scenario, greeting, example messages, system prompt, and tags, plus your lorebook (up to 100 entries). A PNG card's own portrait becomes its avatar; every new scene can render from a synthetic face you pick and pin — a pinned face stays consistent scene to scene. Alternate greetings, author's-note / post-history instructions, and custom card extensions are not imported.
Can I tune it for speed or quality?
Above the 8 GB floor, Settings (switch to Advanced) lets you pick a Speed preset — chat and image models stay loaded together, so there's little or no wait between a reply and its picture — or a Quality preset (a larger chat model + a higher-fidelity image engine, swapped in as needed). A 32 GB card can run both at full quality at once.
Can I keep and move my creations?
Yes. Export any saved chat as a portable bundle — prose, images, and a C2PA provenance signature on each panel, written and verified locally. Nothing phones home.
Is the source open?
Not yet fully — the companion repository opens at launch so the privacy guarantee is auditable. Until then, verify the installer by its published SHA-256 and this page's own Content-Security-Policy. Image models and some assets are distributed separately.
Is it available everywhere?
Not yet everywhere — some regions are excluded at download while legal review completes. We don't publish a compliance claim for any specific jurisdiction. 18+ only.
