Two-Image Blend · مهارة Open Skill عامة

Two-Image Blend

Two-Image Blend هي Open Skill عامة على Digen. افتحها لتعرف ماذا ينشئ هذا التدفق.

@wwws

Product

2 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/njhf72ua.jpg

# Two-Image Blend

Fuses two user-supplied images into one coherent portrait frame: the identity of the first image is kept, the traits of the second image are grafted onto the same anatomy, and the two halves are divided by a glowing energy seam. Inputs: reference images + optional text. Output: one image.

There is no structured project state. Everything the model needs for the next call lives in the conversation, in the analyzed image descriptions, and in the canvas assets.

## When to use

- The user attaches two images and asks to "mix", "merge", "blend", "fuse", "combine" or "turn into one" picture.
- The user wants a human-and-creature hybrid portrait (or two-person / human-animal / material fusion) in a single frame.
- The user wants a short title text rendered on top of such a fused portrait.
- Follow-up requests such as "make the scales cover more of the face", "warmer light", "change the title" — only redo the affected part, do not restart the whole conversation flow.

## Reference files (read on demand, never all at once)

| File | What it holds | When to read |
|---|---|---|
| `references/style-and-prompt.md` | Style, subject/content, layout, must-keep, must-avoid, master prompt template and reusable prompt fragments | Always, right before composing the generation prompt |
| `references/prompt-templates.md` | Filled-in template variants (creature hybrid, animal hybrid, two-person blend, material/style fusion) and weighting rules | When the blend type is not the default creature hybrid, or when the first result is unbalanced |
| `references/step-details.md` | Detailed sub-steps: input validation, image analysis checklist, call checklist, review criteria, retry ladder, edge cases, refusal rules | When gathering inputs, reviewing the result, or a generation needs a retry |

## Steps

1. **Collect the two inputs.** Take the images from the user's attachments, canvas assets or the asset library. Designate image A as the identity/base (body, face shape, pose, hair, framing) and image B as the trait donor (creature, animal, material, second person). Note any overlay text the user gave verbatim.
   - If fewer than two images are available, ask for the missing one; if the user supplied one image plus a written description of the second subject, proceed using that description as B.
   - If the user asks for text only with no image, do not run this flow.

2. **Analyze both images.** Call `recognize_image` once per image with a prompt asking for: subject and framing, facial structure and features, hair, skin/material texture, dominant colors, lighting direction, background, and any visible text. Keep the returned descriptions — they feed both the prompt and the must-keep list. Follow the checklist in `references/step-details.md`.

3. **Load the visual rules.** Call `read_skill_file(name="two-image-blend", path="references/style-and-prompt.md")`. Use it as the single source of truth for palette, materials, lighting, composition and wording of the prompt. Do not paste the file content back to the user.

4. **Choose the blend variant.** Default is the fantasy-creature hybrid described in the style file. If the user's images imply another blend (two people, human plus animal, human plus material/object), also read `references/prompt-templates.md` and use the matching variant.

5. **Write the generation prompt (English).** Fill the master template from `references/style-and-prompt.md` with the descriptors gathered in step 2: what each side keeps from A, what each side inherits from B, the seam, the palette split, the framing, the overlay text (exact characters, if any). State weights explicitly (e.g. "left half 50% human, right half 50% scaled creature") and never invent subjects, clothing or backgrounds that are absent from the inputs.

6. **Generate the image.** Call:

   `generate_image(prompt=<composed English prompt>, input_urls=[image A, image B], model="i2i.hd.lite", orientation="portrait", resolution="2K", name=<short desc