Krea2 VAE Image Generation · सार्वजनिक Open Skill

Krea2 VAE Image Generation

Krea2 VAE Image Generation Digen पर एक सार्वजनिक Open Skill है। खोलें और जानें कि यह वर्कफ़्लो क्या बनाता है।

@eggs

Roleplay

1 0 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-19/xqmmzyzh.jpg

# Krea2 VAE Image Generation

Turn a text description into one finished, high-fidelity still image in a single consistent visual language (the "Krea2 VAE" look: painterly semi-realistic rendering, warm earthy palette, sculpted cinematic light, crisp micro-texture). There is no structured project state — everything lives in the conversation and on the canvas.

## When to use

- The user gives a text description and wants a picture made from it.
- The user wants a character portrait, concept art, key art, or prop/scene image in this specific painterly-realistic style.
- The user wants an existing image redone in this style or made sharper/cleaner (image-to-image branch).
- Not for animation, storyboards, or clips — this Skill produces stills only.

## Steps

1. **Extract the brief.** From the user message pull out: main subject, identity-defining attributes (age, build, hair, wardrobe, props), setting, mood, color accent, framing, and any explicit must-keep detail. If a critical attribute is missing and the request is ambiguous, ask at most one short question; otherwise choose sensible defaults and state them in one line.

2. **Track multi-image work.** If the user wants more than one image (variants, a set), call `write_todos` first with one item per image, and keep exactly one item `in_progress`.

3. **Load the visual conventions — mandatory before any generation.** Call
   `read_skill_file(name='krea2-vae-image-generation', path='references/style-and-prompt.md')`.
   Follow it for Style, Subject/content, Layout, Must keep, Must avoid, and the copy-ready English prompt template. Do not generate before reading it.

4. **Load the operational details when the call is not trivial.** Call
   `read_skill_file(name='krea2-vae-image-generation', path='references/step-details.md')`
   when you need the orientation/resolution matrix, the negative-prompt assembly rules, the image-to-image branch, or the quality-check and retry rules.

5. **Compose the English prompt.** Fill the template from `references/style-and-prompt.md` with the user's subject and the style lock. Every prompt must contain the three style anchors (painterly semi-realistic rendering, warm earthy cinematic palette, crisp high-frequency texture) plus the short "Negative prompt" line from the reference file. Keep the user's own wording for names, labels, and identity details; everything else stays in English.

6. **Call `generate_image` (text-to-image).**
   - `prompt`: the assembled English prompt from step 5.
   - `model='t2i'` (always the channel string, never a physical model name).
   - `orientation`: `portrait` for characters/people and vertical key art, `landscape` for scenery and wide scenes, `square` for objects, icons, and centered product-like shots.
   - `resolution='1080P'` by default (this Skill is an HD-oriented renderer); raise to `2K`/`4K` only when the user asks for a large print file, and keep `720P` for fast drafts or variants.
   - `name`: a short 2–5 word title of the image.
   - `entity_type='images'`; when the subject is clearly a reusable character, location, or prop, use `characters` / `locations` / `props` instead.

7. **Image-to-image branch (only when the user supplies a reference picture or asks to restyle/refresh an existing result).**
   - `generate_image(prompt=<assembled prompt + "preserve the reference subject's identity, pose and silhouette">, input_urls=[reference image asset id or uri], model='i2i.hd.lite', orientation=..., resolution='1080P', name=..., entity_type='images')`.
   - Keep the reference image's identity, structure, and composition; change only what the user asked to change.

8. **Quality check.** When the returned image is available and the user cares about clarity (or when you suspect blur/low resolution from the prompt), call `recognize_image(image_url=<generated image>, prompt='Check for blur, low resolution, plastic oversmoothed skin, broken anatomy, garbled text, and missing requested details.')`. If a defect is rep