Image Studio

Front Porch does not draw pictures. You run a picture program (or a cloud image API). Front Porch sends it a prompt.

Your chat model is not the painter.

Phone: there is no full Studio. Phone Models can still generate one picture and insert it. /image works in web chat.

What you need

1. Settings → Voice & Media → Image Generation → on. Until that is on, the ✨ button is hidden.
2. A picture program running, with an image model loaded — or a Remote API key (the same key as Settings → Backend).
3. Open a chat → ✨ in the input bar.
4. Pick the matching chip inside Studio. Wait until the card says connected (it lists models / LoRAs). Then Generate.

URL and model lists live inside Studio, not on the Voice & Media row.

Front Porch will not start Draw Things, ComfyUI, or Automatic1111 for you.

Backends

Pick a chip in Studio. The tab auto-tests on open and when you switch chips. Retry if it failed.

Connected means the picture program answered and listed checkpoints. If the hint says it is not running — start it, load a model, Retry.

Create vs Edit

Two tabs at the top.

Create — new picture from a prompt. Model, LoRA, size, style, negative prompt, Advanced samplers.

Subject

• Freeform — you write it. Empty + Write it for me pictures the current scene (your chat LLM drafts the prompt; you can edit it).
• Character — close-up from appearance + current expression. Personality text is not stuffed in. In a group, pick one member. A group shot is allowed; diffusion is bad at several specific faces, and the UI says so.
• Your persona — from persona appearance.

Prompt style: natural language (FLUX / SD3) or Danbooru tags (SD 1.5 / anime). Match the checkpoint.

LoRAs: compatible ones show. Confirmed mismatches hide behind Show N incompatible. Weight slider is there.

Reference image (img2img): local backends only (hidden on Remote). Attach a picture + denoise. Not the same as Edit.

Edit — change a picture you already have. Needs an edit image model (Qwen-Image-Edit, Flux Kontext, …). A normal txt2img checkpoint will not do this. The app tells you why; it does not silently fake txt2img.

• Draw Things Edit: Use recommended on the recipe strip. Weird CFG on Qwen-Image-Edit often returns a blank image.
• Comfy Edit: built-in recipe or upload a workflow.

From a result: Variations, edit prompt & regenerate, send to chat, save to gallery, save to disk.

/image in chat

/image — current scene.
/image me · /image char · /image raw <prompt> · /image <description>

If Review AI prompts before generating is on (Studio generation settings), /image pauses so you can edit the crafted prompt.

Expression packs

From one base portrait, so the pack stays the same person. Studio or the character editor portrait panel.

Keep faces you like; only generate missing ones.

QC (the app looking at each face) needs a vision-capable chat model: local GGUF + mmproj in Model Settings → Vision, or a vision API. “Couldn't check vision” usually means the server is still loading.

No base portrait → it stops. No edit-capable image model → pack-from-portrait cannot run.

When it is grey / broken

More Q&A: FAQ → Pictures.