Mirror Floor Photo · সর্বজনীন Open Skill
Mirror Floor Photo
Mirror Floor Photo Digen-এ একটি সর্বজনীন Open Skill। খুলে দেখুন এই ওয়ার্কফ্লো কী তৈরি করে।
Product
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/np2hevwd.jpg
# Mirror Floor Photo
Restyle an uploaded photo into a "mirror floor photo": the subject stays exactly who they are, while the ground under them becomes a polished, glass-smooth mirror surface that reflects them at their feet. This is a single-capability, image-to-image skill. There is no structured project state — the source photo comes from the conversation or the canvas, and the output is one generated image.
## When to use
- The user uploads a photo and asks to turn it into a mirror / reflective floor photo, a "reflection on the ground" shot, or a glossy floor restyle.
- The user asks to redo, re-generate, or tweak a previous mirror-floor result (different floor material, stronger or softer reflection, different crop).
- The user describes the effect in words ("make the ground a mirror with my reflection at my feet") and attaches or references a photo.
Do not use it for generic image generation from text only, for video, or for edits that have nothing to do with a reflective ground surface.
## Steps
1. **Identify the inputs.** Required: a source photo (an uploaded asset, a canvas asset, or an image URL the user provided) plus an optional text instruction (floor material, mood, crop, strength of reflection). If there is no photo, ask the user to upload one — this skill is image-to-image. Only if the user explicitly insists on a text-only version, fall back to `generate_image` with `model='t2i'` and tell them it is a concept image, not a restyle of their photo.
2. **Read the stage plan.** Call `read_skill_file` with `path='references/workflow-stages.md'` and follow the stage order it defines for this request (single pass vs. a refine pass on an existing result).
3. **Read the input/check rules.** Call `read_skill_file` with `path='references/step-details.md'` and run the input validation checks, the orientation mapping, the prompt-assembly checklist, and the retry rules described there. Use `recognize_image` on the source photo when you need to confirm what is in it (subject, pose, framing, visible ground) before writing the prompt.
4. **Read the visual spec.** Call `read_skill_file` with `path='references/style-and-prompt.md'` and build the final English prompt from the template and fragments in it: the subject description, the scene description, the mirror-floor treatment, the must-keep list, and the avoid list. Never skip this step and never generate from a generic one-liner.
5. **Generate the image.** Call `generate_image` with:
- `prompt`: the English prompt assembled in step 4 (subject + mirror floor + lighting/material + must-keep + avoid list).
- `input_urls`: the source photo (single image, first entry). This drives the restyle.
- `model='i2i.hd.lite'` (image-to-image channel; `t2i` only in the text-only fallback from step 1).
- `orientation`: match the source photo — `portrait` for vertical shots (the default for mirror-floor shots, since the floor reflection needs room below the subject), `landscape` for horizontal shots, `square` for 1:1 sources.
- `name`: a short canvas title, e.g. "Mirror Floor Photo".
- `entity_type='mirror_floor_photos'`.
- Leave `resolution` unset unless the user asks for a specific quality tier.
6. **Reply and guide the next step.** Summarize in natural language what was kept from the original and what the new floor looks like; do not paste long URLs. If the user may want variations, call `set_guidance` in the same turn (e.g. "Make the floor wet black marble", "Soften the reflection", "Keep my original framing").
## Notes
- Preserve identity above everything: face, hair, body proportions, outfit, pose, and framing must survive the restyle. If a result visibly changes who the person is, regenerate once with a stronger must-keep clause instead of shipping it.
- The mirror floor must obey perspective: the reflection sits directly beneath the subject, aligned with their feet, with foreshortening consistent with the camera angle. No floating, offset, or duplicate