# age_shift_cli.py

End-to-end CLI for ComfyUI-AgeShift. Talks to a running ComfyUI server,
uploads a face image, runs detection, and (optionally) drives PuLID-Flux to
generate age-shifted portraits.

## Quick start

```bash
# Detect age only
uv run python age_shift_cli.py face.jpg --detect-only --ip 10.8.0.12

# Single target age
uv run python age_shift_cli.py face.jpg --target-age 65 --ip 10.8.0.12

# Sequence
uv run python age_shift_cli.py face.jpg \
    --target-ages 5,15,25,45,65 \
    --gender male --out ./out --ip 10.8.0.12

# Debug — save generated workflow JSON(s) for inspection in UI
uv run python age_shift_cli.py face.jpg --target-age 65 \
    --dump-workflow /tmp/agewf.json --ip 10.8.0.12
```

## Argparse reference

| Group | Flag | Default | Notes |
|---|---|---|---|
| **server** | `--ip` | `127.0.0.1` | ComfyUI host |
| | `--port` | `8188` | ComfyUI port |
| | `--server` | — | overrides `--ip:--port` |
| **mode** | `--detect-only` | False | skip generation |
| | `--target-age` | — | single target |
| | `--target-ages` | — | csv of targets, e.g. `5,15,25,45,65` |
| **subject** | `--subject` | `person` | injected into prompt |
| | `--gender` | `unspecified` | `male` / `female` / `non-binary` |
| | `--style` | `photorealistic ...` | base style prompt |
| | `--extra-positive` | `same identity, ...` | identity-preserving anchor |
| **gen** | `--width / --height` | 1024 | Flux native is 1024 |
| | `--steps` | 20 | Flux dev: 20–28 is fine |
| | `--cfg` | 1.0 | Flux uses CFG=1 + `--guidance` |
| | `--guidance` | 3.5 | Flux distillation guidance |
| | `--sampler / --scheduler` | `euler / simple` | |
| | `--seed` | 0 (random per age) | |
| **pulid** | `--pulid-weight` | 0.9 | identity strength |
| | `--pulid-start / --pulid-end` | 0.0 / 1.0 | active range |
| **models** | `--flux-unet` | `flux1-dev.safetensors` | filename on server |
| | `--flux-clip-l` | `clip_l.safetensors` | |
| | `--flux-t5` | `t5xxl_fp8_e4m3fn.safetensors` | |
| | `--flux-vae` | `ae.safetensors` | |
| | `--pulid-weights` | `pulid_flux_v0.9.1.safetensors` | |
| | `--insightface-provider` | `CUDA` | `CPU` / `ROCM` also valid |
| **detector** | `--detector-model` | `nateraw/vit-age-classifier` | HF ID or local path |
| | `--detector-device` | `auto` | |
| | `--detector-dtype` | `fp32` | |
| **debug** | `--out` | `./out` | output dir |
| | `--dump-workflow` | — | writes `*.detect.json` / `*.shift_ageN.json` |
| | `--timeout` | 600 | per-workflow seconds |

## Troubleshooting

**`/prompt 400 - node 'AgeDetectorLoader' not found`**
The node package isn't loaded on the server. Check
`http://server:8188/object_info | grep AgeDetect`.

**`/prompt 400 - node 'PulidFluxModelLoader' not found`**
Install `ComfyUI-PuLID-Flux` on the server.

**`Estimated age: 0`**
Detector hit a corrupt or all-black input. Confirm the source image renders
correctly: `curl http://server:8188/view?filename=face.jpg`.

**Hangs after `executing node sampler_node`**
PuLID weight too aggressive, or InsightFace download in progress. First run
of PuLID downloads ~600 MB of weights — watch the server log.

**Output looks like the source age, not the target**
- Bump `--guidance 4.5` (Flux follows prompt harder).
- Drop `--pulid-weight 0.75` (less identity dominance).
- Steps to 24–28.
- Make sure your style prompt isn't pulling toward "young model".