Vibe Tavern 1.4.1 — image generation right in the chat, plus lorebooks and macros that work the way ST's do
Hi again! I'm the developer of Vibe Tavern, an open-source AI roleplay client. Versions 1.4 and 1.4.1 are out, and this post covers both.
Previous posts:
https://www.reddit.com/u/NoiNeri/s/SSR4F6zg6D
https://www.reddit.com/r/SillyTavernAI/s/0DGbon3rzb
For anyone new: VT is built around ease of use without giving up deep controls, especially for people who create and work on character cards with AI. It runs on Windows, Linux, Docker and Android (a single APK, no Termux).
A few weeks ago I asked here what people actually generate during RP. Thank you to everyone who answered — the image generation below is the result.
Image generation
Image generation is now built into the chat. No extensions to install, and no switching to another window to generate and carry the picture over by hand.
You pick a mode, VT writes the image prompt from the chat for you, and the picture lands in the message. Don't like it? Generate another variant and swipe between them, the same way you swipe replies. Swiping works even while the model is busy writing.
There are eight modes: the five you know from SillyTavern (scene background, scene illustration, full-body character, character portrait, your persona), plus the two people asked for — a selfie and the character's reaction to the current moment — and a free mode where your prompt goes in as is.
The prompt assist knows which model family you're generating for. Pony, Illustrious and NoobAI want tags, Qwen or a realism SDXL wants plain sentences, and VT detects the family from your checkpoint and picks the matching template. You can pin it by hand if it guesses wrong. It also only sees what the mode needs: a portrait gets the character's appearance, not the whole chat. And if the assistant refuses instead of writing a prompt, VT stops with a clear error instead of feeding the refusal to the image model.
Where the images come from
- Your own GPU. ComfyUI works through a normal settings form: no node graphs. Pick a checkpoint, sampler, scheduler and LoRAs, watch the step progress, interrupt if you need to. A1111-compatible servers (Forge) work too.
- The cloud. 28 image services with ready-made presets: OpenAI, OpenRouter, Google Gemini, Black Forest Labs (FLUX), Stability AI, fal.ai, Replicate, Together AI, NanoGPT, NVIDIA, Hugging Face and more.
- Free. No GPU and no budget? Pollinations and AI Horde work out of the box.
The fine-tuning chip
Everything you'd want to tweak before a generation sits in one chip in the chat: target and resolution presets, LoRA strength, hires fix, an AI-written prompt you can edit, and a Generate button. On a wide screen it opens as two panes, so the prompt and the sampler settings are side by side.
The image prompts themselves live in the Prompt Manager's new Images tab: edit the template for every mode, save them as profiles, and keep separate variants per model family.
Lorebooks: SillyTavern 1.18 parity
If you've been keeping your world in ST lorebooks, they now behave the same way in VT. Activation follows ST 1.18: delay, recursion, min activations, whole-word matching, per-book budgets with an overflow alert, `{{outlet::…}}`. Dry-run previews no longer burn your sticky and cooldown timers.
Import and export round-trip without losing anything: roles, character filters, groups, probabilities, timed effects, match sources. When you import a card with an embedded lorebook, VT offers to turn it into a lorebook bound to that character.
1.4.1 adds a few things on top:
- Finding the right book. A «Current» tab shows everything that takes part in this chat, «Bound» filters by character or persona, and every tab has search by name. Same for scripts.
- More places to insert and match. Entries can go before or after the persona description, and can trigger on alternate greetings, the chat's dynamic prompt and the chat summary.
- No hidden owners. A bound book belongs exactly to the characters and personas you link it to, and deleting a character no longer deletes the books it had.
SillyTavern macros
Cards and presets from ST rely on macros, and 1.4.1 closes the gaps that actually show up in them:
- Random macros stay put. `{{random}}`, `{{pick}}`, `{{roll}}` and the time macros in greetings and in your messages are fixed when the message is written, like in ST, instead of re-rolling every turn. Name and persona macros stay live, so switching personas still works.
- Full `pick` / `roll`, `trim`, variable shorthand (`.x`, `$x`) and globals within one prompt, plus more ST macro names that used to leak into the prompt as plain text.
- When you import a card or preset that uses a macro VT doesn't support, it tells you instead of failing silently.
And a few more things
- Co-Author got its own settings. The card-writing assistant now has its own model, samplers and limits per connection, separate from your RP settings, an optional separate model for lore, and it can edit every lorebook setting.
- Impersonation can now improve your draft. Type a rough line, switch on "Improve draft", and the model rewrites it in your persona's voice instead of writing from scratch.
- AI editor: save instructions you keep retyping («shorten», «tone down the drama») as templates, and merge replies from as few as two swipes.
- Cut-off replies are kept. If the provider drops the connection mid-reply, the part that already arrived stays in the chat, as if you had pressed Stop.
- Android. The 10–12 crash right after launch is fixed, and Play Protect shouldn't flag the APK as harmful anymore (the release notes link its VirusTotal scan if you want to check). On phones, the token counter is now a small context ring in the input row.
- Faster binding picker on big libraries, lighter avatars, search in the character pickers, and search and drag-and-drop in every provider list.
Repository: https://github.com/Noineri/vibe_tavern
Release: https://github.com/Noineri/vibe_tavern/releases/tag/v1.4.1
Full changelog: https://github.com/Noineri/vibe_tavern/compare/v1.3.0...v1.4.1
Vibe Tavern is free and open source under AGPL-3.0. I still build it mostly on my own, so feedback is genuinely useful — especially from people who already generate images during RP: which modes you actually use, and what's missing.