Pictures

Pictures

Pictures draws images inside a chat. The writer drafts a prompt from the scene, you change what you like, and the picture lands as a message. It draws on your own ComfyUI, or on your own OpenAI or Google AI Studio key if you have no GPU.

For characters standing beside the story, with a face for each feeling, see Sprites.

Drawing in the chat

Opening the sheet

You can open the Picture sheet four ways:

• On a message, open Message Actions and choose the paintbrush. The sheet opens on This moment for that message.
• In the wand menu by the message box, choose Draw a picture.
• Write /imagine. It also answers to /sd, /img and /image.
• On a picture you already drew, choose Edit prompt (see The picture menu).

/imagine takes an optional first word that picks what to draw: moment, char, me, place or free. /imagine place a rainy street opens the sheet on The place with that text filled in. Text with no first word starts on Free. Add quiet=true to skip the sheet and draw straight away; the command returns the picture's address.

The sheet

The sheet works in two steps, shown under its title: Set up, then Write and draw. Nothing is written until you ask.

Step 1, Set up

1. Pick what to draw. This moment draws the message you came from, or the last one. The second chip carries the character's name and draws them on their own. Me draws your persona. The place draws the surroundings with nobody in them. Free is a prompt you write yourself.
2. Pick how to draw it: the model, pipeline and preset (see Picking a model in the sheet), any LoRAs, the Size, and on ComfyUI the Seed: Random or Pinned. A pinned seed shows its number, and the dice button rolls a new one.
3. Choose Write scene. The writer drafts the prompt once, with your connected chat model, in the style the chosen pipeline reads (tags or plain description).

Step 2, Write and draw

1. Read the Prompt and edit it freely. Choose Write it again to redraft.
2. Open Negative prompt to see or change what the picture should avoid. A preset that ignores negatives says so and hides the box.
3. The model, size and seed are still on the sheet if you want to change them.
4. Choose Draw. The sheet closes and the picture starts.

If you pick another subject after the prompt is written, the sheet does not rewrite it on its own. It says what the prompt was written for and offers Write it for the new subject, or Keep. With an empty prompt box, a new subject goes back to Set up. If you switch to a model that reads the other style, the sheet offers Rewrite as tags or Rewrite as description, or Keep.

Free, Edit prompt, and a sheet with no chat model to write with open straight on step 2, where the prompt box is yours.

If the writer cannot draft (no model connected, a sign-in that was refused, or a character with no description to draw from), the sheet says why and you can type the prompt yourself. Portraits (the character and Me) are drafted from the card or persona description alone, never from the chat.

While it draws

A tile appears at the end of the chat with a ring, a line such as "Drawing this moment", and a status: "Sending to the picture server…", "In the queue, 2 ahead", then "Step 12 of 30". Mikan-chan paints at her easel while you wait. She follows the step count, and when the picture arrives she paints the rest of the canvas before the tile leaves. On a lane that reports no step count she keeps to her palette until then.

To see the model's own preview frames instead, switch on Show the model's live preview while drawing in Settings → Pictures, under In the chat. If a theme hides the mascot, the tile shows the preview too. Kotatsu asks ComfyUI for those frames only when the tile will show them, so a stock ComfyUI sends them without any change on your side.

Choose Cancel to stop. If something goes wrong the tile keeps the message and offers Retry and Dismiss. If ComfyUI is not answering it also offers Set up ComfyUI, and Draw with your Gemini key (or OpenAI) when you have stored such a key. That last button switches the lane for good, the same as the switch in Settings.

You can leave the chat while a picture draws. If it finishes elsewhere, it is saved to the gallery and a notice says so.

The picture menu

Each picture has a small button over it, Picture options:

• Again: draws the same prompt with a new seed. The new picture joins the message, and the gallery arrows walk between them.
• Same seed, bigger: draws the same seed through the model's bigger version. It only appears when there is one.
• Edit prompt: opens the sheet titled Edit and draw again, filled with that picture's prompt, seed, model and LoRAs. The result joins the same message.
• Use as reference: the next picture you draw uses this one as its reference.
• Use as background: makes this picture the chat's background behind the characters in the Sprite look. A picture drawn as The place suits it best. The notice says where it shows when the Sprite look is not on.
• Delete this picture: asks first, and Also delete the file is ticked by default. The message keeps its other pictures.

Set as portrait is in the menu but switched off in this release.

Captions

A picture message carries a caption line, [Name sends a picture that contains: …], so the chat model knows what the picture shows on later turns. It is hidden by default. Switch on Show the caption line under pictures in Settings → Pictures to see it.

The character and yourself

The character chip and Me are portraits. They start from the card's description (or your persona's) with no scene attached. The picture message is kept out of the chat model's context, so a portrait does not become part of the story.

Reference pictures

If the model or workflow has a reference slot, switch on Use the card's art as reference to draw from the card's picture. After Use as reference, the sheet shows "Reference: a picture from this chat" with Forget this reference. OpenAI and Gemini always offer a reference.

Local ComfyUI

ComfyUI draws on your own computer, and the prompt stays on your machines. Pictures start on this lane.

Pointing Kotatsu at it

1. Open Settings, then Pictures. Under Where pictures are drawn, choose ComfyUI.
2. Under ComfyUI, check the Address. It starts as http://127.0.0.1:8188, where ComfyUI usually answers.
3. If your ComfyUI asks for a login, fill in Sign-in. Leave it empty otherwise.
4. Choose Check. Kotatsu reports the version and what the server has.

Setting up local pictures

Until everything is ready, Settings shows Setting up local pictures with three steps that tick themselves:

1. Install ComfyUI and start it: links to ComfyUI Desktop and its install guide for Windows, macOS and Linux. Then press Check.
2. Get a model family's files: lists the files each family still needs, the ComfyUI folder each goes in, and a Download link. Put the files in place and choose Check again.
3. Set up your models below: confirm what Kotatsu suggests.

Models

Under Models, every model file ComfyUI has is listed with a chip: Set up, Suggested or Not set up. Kotatsu suggests a family from the file's header or name; it never guesses when nothing says. Pick a family and a preset for each model, or choose Confirm (or Confirm N suggestions for all at once). The sheet only offers models you have confirmed. Tick Only models not set up to see what is left.

The families are Anima, KREA 2, and SDXL / Illustrious / Pony. Each family has presets, such as Anima's Base and Turbo, that set steps, CFG and sampler to what the model's makers recommend.

Picking a model in the sheet

Once a model is set up, the sheet shows Draw with, grouped by family, then Pipeline (Plain, or Bigger (hires) where there is one) and Preset. With no model set up, it shows a Recipe menu instead.

The sheet opens on the last model you drew with on this server, with its pipeline and preset. If you last drew with one of Your own workflows, it opens on that. Edit prompt is the exception: it opens on the picture's own model and pipeline. /imagine quiet=true follows the same rule.

A model only draws with a pipeline made for its own family. If your install has none for that family, the sheet says so in place of the menu, for example "No KREA 2 pipeline installed". Write scene and Draw stay off, and Open Settings takes you to Settings → Pictures. Pick another model, or add a recipe for that family.

If ComfyUI is not answering when the sheet opens, the sheet says so with Set it up, and Draw with your Gemini key (or OpenAI) when you have stored such a key.

Recipes

A recipe is a ComfyUI workflow plus what Kotatsu knows about where its prompt, seed and size go. Kotatsu ships recipes for its model families, and they appear in the sheet as pipelines.

To use your own, export a workflow from ComfyUI in API format and put the file in the user\workflows folder of your Kotatsu data (Documents\Kotatsu\data). It shows under Recipes in Settings → Pictures and under Your own workflows in the sheet. Your workflow draws as it is built, with its own settings. If it needs a node your ComfyUI lacks, the recipe says Needs: and names it. Under Default recipe for each kind of picture you can set one per purpose, or leave Ask each time.

LoRAs

In the sheet, LoRAs lists the ones on the stack, each with a strength. Use Add a LoRA… to add more, up to eight. The list starts with LoRAs made for the model's family; Show all reveals the rest, and a LoRA made for another family is marked. A LoRA's trigger words go into the prompt on their own, and come out when you remove it.

A card can carry its own LoRAs. Choose Edit card in the library and switch Card Studio to Advanced. In the looks chapter, named after the character (such as How Hana looks), find How new pictures are drawn. In the Simple view the chapter shows a one-line summary and an Edit the look button, which switches to Advanced for you.

• Always include: tags put in front of every picture of this character.
• Never include: tags for the negative prompt.
• Written as: Whatever the recipe wants, Tags or Prose.
• Reference with Use card art.
• LoRAs: pick them from your ComfyUI's list, set a strength and edit the Trigger words.

The card's LoRAs start on This moment and the character's own picture. Me, The place and Free start without them. The look is saved in the card file, so it travels when you share the card.

Your own OpenAI or Gemini key

If you have no GPU, Pictures can draw with a key you already added under Connection: an OpenAI key or a Google AI Studio key.

1. Open Settings, then Pictures. Under Where pictures are drawn, choose OpenAI key or Gemini key.
2. Pick the Model, Quality and Size (OpenAI), or the Model, Size and Shape (Gemini).
3. Open the sheet as usual. It says Drawn with and names the lane and model.

What each lane can and cannot do:

• No seed, no negative prompt, no LoRAs and no recipe. Those parts of the sheet are gone.
• The writer drafts in plain description.
• You choose a Size (OpenAI, including Let the model choose) or a Shape (Gemini).
• A reference picture is always on offer.
• Again repeats the same request. There is no Same seed, bigger.
• Gemini marks every picture with Google's SynthID watermark.
• A key draws pictures, but not a sprite set. Sprite sets need ComfyUI.

The prompt, which may quote the chat, and any reference picture are sent to OpenAI or Google. Each picture is billed to your key at the provider's rates. Settings shows the provider's own listed price for your choice, with the date it was read and a link to their pricing page.

Good to know

• Pictures are saved in your data folder, in user\images, in a folder named after the character. File names start with the date and time.
• A picture is an ordinary chat message, so chats still open in stock SillyTavern and the pictures show in the gallery.
• Portraits are kept out of the chat model's context. Other pictures are not: their caption line is how the chat model knows what was drawn.
• In a scene, a picture comes from the narrator when one is seated.
• Your ComfyUI also draws whole sprite sets for a card, in Card Studio's looks chapter. A card's look (How new pictures are drawn) is where that starts too.
• Pictures are part of Kotatsu's own layout. Under the classic layout, /imagine says so.
• The old Image Generation extension is gone. Its ComfyUI address carries over once, and a card with no look yet gets its old character prompt as one.
• The writer uses your connected chat model, so each draft is a small extra request on that model.