| name | hatch-pet-realistic |
| description | Use when a user wants a Codex desktop pet based on a realistic human, photo-style person, real-person reference image, portrait companion, or full-body human character that must stay visually consistent across pet animations. |
Hatch Pet (Realistic Human)
Create a Codex-compatible animated pet rendered as a photo-realistic human. The key risks are identity drift, small unreadable people, stale script arguments, opaque backgrounds, and accidentally composing un-matted gray-background frames.
Hard Rules
- All visual subject pixels come from
$imagegen. Local scripts may only copy, record, slice, matte, inspect, assemble, validate, and package.
- Record every selected
$imagegen output from $CODEX_HOME/generated_images/.../ig_*.png with record_imagegen_result.py; do not edit imagegen-jobs.json by hand.
- Running-left must be generated independently. Do not mirror running-right for realistic humans.
- The final atlas must be 1536x1872, 8 columns x 9 rows, 192x208 per cell, RGBA/transparent, with frame counts from
references/animation-rows.md.
- Never compose from
<run-dir>/frames_raw. Compose only from the clean rembg output directory <run-dir>/frames.
- Realistic humans must be large in the pet cells: base fills
92-96% image height; row-strip people fill each 192x208 slot with no extra empty margin.
Prerequisites
- Use
uv run --with ... for script dependencies. If dependencies are already cached, prefer uv run --offline --with ....
- First use may need network access for Python packages and the rembg
u2net_human_seg model. Warm the model with:
uv run --with rembg[cpu] --with pillow --with numpy hatch-pet-realistic/scripts/warmup_models.py
Visible Progress
Maintain these four visible milestones:
- Getting
<Pet> ready: name, description, optional reference image, run directory.
- Imagining
<Pet>'s main look: generate a tight full-body neutral base, inspect it, and lock identity.
- Picturing
<Pet>'s poses: generate tight idle + running-right row strips first, then the remaining rows.
- Hatching
<Pet>: slice, matte, scale-normalize, inspect, assemble, validate, QA, and package.
Workflow
1. Prepare the run
uv run --with pillow hatch-pet-realistic/scripts/prepare_pet_run.py \
--pet-name "<slug>" \
--display-name "<Display Name>" \
--description "<detailed adult human description>" \
--reference <optional-reference-path> \
--output-dir <run-dir> \
--force
This creates pet_request.json, prompt files, layout guides, and imagegen-jobs.json.
2. Generate and record the base
Use $imagegen with prompts/base-pet.md. Treat uploaded photos as character references only when policy and backend support allow it; otherwise use a text description derived from the target person.
The base must be one complete full-body person, tightly framed, subject height 92-96%, head or hair within 4-8 px of the top safe area, shoes within 4-8 px of the bottom safe area, and no extra empty margin. Regenerate if the person appears distant or small.
Record the selected built-in image output:
uv run --offline --with pillow hatch-pet-realistic/scripts/record_imagegen_result.py \
--run-dir <run-dir> \
--job-id base \
--source <CODEX_HOME>/generated_images/<run>/ig_*.png \
--force
Write pet_request.json#character_lock with a 30-50 word identity description: adult gender presentation, age range, build, skin tone, eyes, hair, outfit, shoes, and any stable visual details. Replace {{character_lock}} in prompts/rows/*.md.
3. Generate and record row strips
Generate idle and running-right first. Inspect identity, outfit, frame count, full-body visibility, row action, and spacing. Every invisible 192x208 slot should contain one complete person as large as possible while keeping the head, hands, and shoes visible. If identity drifts or the subject is too small, regenerate before continuing. Use references/animation-rows.md for frame counts and row action semantics.
For each row, record the selected built-in image:
uv run --offline --with pillow hatch-pet-realistic/scripts/record_imagegen_result.py \
--run-dir <run-dir> \
--job-id <state> \
--source <CODEX_HOME>/generated_images/<run>/ig_*.png \
--force
Rows are idle, running-right, running-left, waving, jumping, failed, waiting, running, and review. Subagents may generate images, but only the parent records jobs and writes manifests.
4. Hatch with clean rembg frames
The safest full pipeline is:
uv run --with pillow --with numpy --with rembg[cpu] hatch-pet-realistic/scripts/finalize_pet_run.py \
--run-dir <run-dir> \
--allow-slot-extraction
finalize_pet_run.py intentionally uses this sequence:
uv run --offline --with pillow hatch-pet-realistic/scripts/extract_strip_frames.py \
--decoded-dir <run-dir>/decoded \
--output-dir <run-dir>/frames_raw \
--states all \
--method slots \
--no-matte
uv run --with pillow --with numpy --with rembg[cpu] hatch-pet-realistic/scripts/matte_frames.py \
--frames-dir <run-dir>/frames_raw \
--output-dir <run-dir>/frames \
--qa-dir <run-dir>/qa \
--target-height-ratio 0.94 \
--running-target-height-ratio 0.88 \
--edge-padding 2
uv run --offline --with pillow hatch-pet-realistic/scripts/compose_atlas.py \
--frames-root <run-dir>/frames \
--output <run-dir>/final/spritesheet.png \
--webp-output <run-dir>/final/spritesheet.webp
uv run --offline --with pillow hatch-pet-realistic/scripts/validate_atlas.py \
<run-dir>/final/spritesheet.png \
--json-out <run-dir>/final/validation.json
matte_frames.py removes the background and normalizes the transparent foreground size. Use --no-normalize-scale only for diagnosis. Do not point compose_atlas.py --frames-root at frames_raw; those frames still contain the generated background.
5. QA and package
Create a contact sheet if finalize did not already do so:
uv run --offline --with pillow hatch-pet-realistic/scripts/make_contact_sheet.py \
<run-dir>/final/spritesheet.png \
--output <run-dir>/qa/contact-sheet.png
Package manually only when needed:
uv run --offline --with pillow hatch-pet-realistic/scripts/package_custom_pet.py \
--pet-name "<slug>" \
--display-name "<Display Name>" \
--description "<short description>" \
--spritesheet <run-dir>/final/spritesheet.webp \
--force
Writing to ${CODEX_HOME:-$HOME/.codex}/pets/<pet-name>/ may require filesystem approval outside the workspace.
Common Failures
| Symptom | Fix |
|---|
record_imagegen_result.py rejects the source | Use the original $CODEX_HOME/generated_images/.../ig_*.png, not a copied or locally edited image. |
| Atlas validates as nearly opaque | You composed from frames_raw or skipped rembg. Re-run matte_frames.py --output-dir <run-dir>/frames and compose from frames. |
uv tries to reach PyPI in a restricted environment | Retry with uv run --offline --with ... if dependencies are cached; otherwise request network approval. |
| Row inspection flags slot extraction | For realistic row strips, slot extraction is acceptable after visual inspection; pass --allow-slot-extraction to finalize. |
| Identity drifts across rows | Regenerate the row using the locked identity paragraph and base image as visual truth; do not locally edit faces or outfits. |
| Subject too small or unreadable in the pet UI | Regenerate the base or row with tight 192x208 slot framing and no extra empty margin, then finalize so matte_frames.py can scale-normalize transparent frames. |
References
references/animation-rows.md: row names, frame counts, durations, and motion semantics.references/qa-rubric.md: visual QA and repair guidance.references/realistic-style-guide.md: realistic prompt anchors and negatives.references/codex-pet-contract.md: Codex atlas loading contract.
