FEATURE REQUEST: Make character chat render its character consistently

SUMMARY

Character chat has a "visualize" button but no visual controls, so every render

has to be manually prompt-engineered from the chat. The character's appearance

lives in a prose description that gets summarised before it reaches the image

model, and the avatar is only a weak hint. Result: the character drifts, extra

people get drawn, and clothing/state is inconsistent.

WHAT ACTUALLY HAPPENS (reproduced)

1. Summariser drops constraints. The character description contains hard rules

   ("one woman only", "no other person", "full body"). Before reaching the image

   model the description is compressed, and the constraints are dropped. Renders

   then show a second person, a cropped shot, or the wrong state.

2. The avatar does not lock identity. The avatar is treated as a soft hint. A

   longer description outvotes it, so the character renders as a similar-looking

   stranger - different face, different body, different tattoo placement. With a

   short description she matches better; with a long one she does not match at all.

3. The image prompt is generated as prose, and prose casts a cast. Example

   generated prompt: "...Below, on the sprawling lawn, a figure stands looking up

   at her, creating a tense, introductory moment between neighbours." Because the

   prose describes a second person, a second person is drawn. Nothing in the

   prompt says the character is alone.

4. Naming the other character in the prompt causes a body to be drawn. If the

   prompt contains a proper name ("Tommy") or a person-noun ("a figure", "a man"),

   the model renders that person. "You" and "the user" behave better; a proper

   name reliably produces an unwanted figure.

5. Nude vs clothed has no toggle. The reference/avatar being nude makes every

   render nude, regardless of text saying she is clothed. There is no reliable way

   to say "clothed this time".

6. Cropped avatars leak their crop. A circular-cropped avatar on a white canvas

   causes the white corners / circular shape to be reproduced inside renders.

7. No wardrobe state. Clothing continuity has to be carried by the character

   narrating what she is wearing, so a summariser picks it up. If she does not

   narrate it, the outfit is lost.

PROPOSED SOLUTION

A. Visual bible (highest priority)

   A short, separate, never-summarised field holding the character's exact look,

   tattoos, and default wardrobe. Re-injected verbatim on every render, bypassing

   the chat summariser entirely.

B. Avatar as an identity lock

   Feed the avatar to image generation as image-to-image at high strength by

   default, with a user-adjustable strength slider. The avatar should define

   identity; text should only define scene and state.

C. Template-based prompt assembly

   Replace LLM prose prompts with slot filling:

     [character] + [wardrobe state] + [pose] + [camera] + [setting]

   Prose is what invents extra people. A template cannot.

D. Explicit render toggles

   - Solo / with-user (with a "user visible: none | hand | partial | full" option)

   - Clothed / nude / custom outfit

   - Camera