GameCult/Ghostlight
Watch
9
0
Fork
You've already forked Ghostlight
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Split BFL base and edit prompts

Browse source
This commit is contained in:
Meta Vens 2026-05-03 13:00:41 +01:00
commit 664b8a2ced
3 changed files with 133 additions and 37 deletions
Download patch file Download diff file Expand all files Collapse all files

21
scripts/README.md
View file

@ -6,6 +6,18 @@
API, polls the returned `polling_url`, downloads the signed `result.sample`
image immediately, and writes a metadata receipt beside the image.
For Ghostlight visual-plan modifiers, the script uses the durable continuity
workflow:
1. Generate the scene base image from the assembled base prompt.
2. Save the base image locally.
3. Send that base image back as `input_image`.
4. Send only the selected modifier prompt as the edit instruction.
Dry-runs write `.base.prompt.txt` and, when modifiers are selected,
`.edit.prompt.txt` separately so the two prompts can be inspected before
credits get thrown into the machine.
The script reads the API key from `BFL_API_KEY` first. If that is not set, it
falls back to `E:\Projects\gamecult-ops\bfl-api.txt`. It never prints the key.
@ -21,12 +33,21 @@ Dry-run one Ghostlight visual scene prompt:
npm run images:bfl -- --visual-plan examples\visual\pallas-species-strikes.branch-and-fold.v0.visual.json --scene-id pallas_yard_wake --dry-run
```
Dry-run a base scene plus one edit modifier:
```powershell
npm run images:bfl -- --visual-plan examples\visual\pallas-species-strikes.branch-and-fold.v0.visual.json --scene-id pallas_ilya_arrival --global-modifier-id formal_stoppage_visible --dry-run
```
Submit one paid render:
```powershell
npm run images:bfl -- --visual-plan examples\visual\pallas-species-strikes.branch-and-fold.v0.visual.json --scene-id pallas_yard_wake --width 1536 --height 864 --output-format png --yes
```
If modifier ids are supplied, one visual scene costs two BFL jobs: one base
generation and one edit pass from that base image.
Render every base scene from a visual plan:
```powershell

146
scripts/generate_bfl_images.py
View file

@ -1,8 +1,10 @@
"""Generate Ghostlight images through the Black Forest Labs API.
The script can submit a plain prompt or assemble prompts from a Ghostlight
`.visual.json` plan. It writes each downloaded image beside a metadata receipt
that preserves the prompt, request id, polling result, and source scene.
`.visual.json` plan. For modifier renders, it generates the base image first,
then sends that image back to BFL with only the modifier prompt as the edit
instruction. It writes downloaded images beside metadata receipts that preserve
the prompt, request id, polling result, and source scene.
"""
from __future__ import annotations
@ -38,6 +40,7 @@ class RenderJob:
source_ref: str
scene_id: str | None = None
modifier_ids: tuple[str, ...] = ()
edit_prompt: str | None = None
def load_json(path: Path) -> dict[str, Any]:
@ -146,14 +149,11 @@ def selected_global_modifiers(
return modifiers
def assemble_visual_prompt(
def assemble_visual_base_prompt(
data: dict[str, Any],
scene: dict[str, Any],
scene_modifier_ids: set[str],
global_modifier_ids: set[str],
) -> tuple[str, tuple[str, ...]]:
) -> str:
plan = data.get("visual_scene_plan", {})
scene_id = scene["visual_scene_id"]
sections: list[str] = []
style = plan.get("global_style_cue", "").strip()
@ -173,19 +173,25 @@ def assemble_visual_prompt(
sections.append("Visible character references:\n" + "\n".join(character_lines))
sections.append("Scene base image:\n" + scene.get("base_image_prompt", "").strip())
return "\n\n".join(section for section in sections if section.strip())
def assemble_visual_edit_prompt(
data: dict[str, Any],
scene: dict[str, Any],
scene_modifier_ids: set[str],
global_modifier_ids: set[str],
) -> tuple[str | None, tuple[str, ...]]:
plan = data.get("visual_scene_plan", {})
scene_id = scene["visual_scene_id"]
applied: list[tuple[str, str]] = []
applied.extend(selected_scene_modifiers(scene, scene_modifier_ids))
applied.extend(selected_global_modifiers(scene_id, plan, global_modifier_ids))
if applied:
sections.append(
"Apply these additive visible state changes:\n"
+ "\n".join(f"- {modifier_id}: {prompt}" for modifier_id, prompt in applied if prompt)
)
if not applied:
return None, ()
return "\n\n".join(section for section in sections if section.strip()), tuple(
modifier_id for modifier_id, _ in applied
)
prompt = "\n".join(prompt for _modifier_id, prompt in applied if prompt).strip()
return prompt, tuple(modifier_id for modifier_id, _ in applied)
def build_visual_jobs(args: argparse.Namespace) -> list[RenderJob]:
@ -212,7 +218,13 @@ def build_visual_jobs(args: argparse.Namespace) -> list[RenderJob]:
jobs: list[RenderJob] = []
for scene in selected:
scene_id = scene["visual_scene_id"]
prompt, applied = assemble_visual_prompt(data, scene, scene_modifier_ids, global_modifier_ids)
prompt = assemble_visual_base_prompt(data, scene)
edit_prompt, applied = assemble_visual_edit_prompt(
data,
scene,
scene_modifier_ids,
global_modifier_ids,
)
suffix = "-".join(slugify(item) for item in applied)
render_id = slugify(f"{visual_plan_id}-{scene_id}" + (f"-{suffix}" if suffix else ""))
jobs.append(
@ -222,6 +234,7 @@ def build_visual_jobs(args: argparse.Namespace) -> list[RenderJob]:
source_ref=f"{args.visual_plan}#{scene_id}",
scene_id=scene_id,
modifier_ids=applied,
edit_prompt=edit_prompt,
)
)
return jobs
@ -260,7 +273,7 @@ def download_file(url: str, output_path: Path) -> None:
output_path.write_bytes(response.read())
def build_payload(args: argparse.Namespace, prompt: str) -> dict[str, Any]:
def build_payload(args: argparse.Namespace, prompt: str, input_images: list[str] | None = None) -> dict[str, Any]:
payload: dict[str, Any] = {
"prompt": prompt,
"output_format": args.output_format,
@ -278,12 +291,24 @@ def build_payload(args: argparse.Namespace, prompt: str) -> dict[str, Any]:
payload["webhook_url"] = args.webhook_url
if args.webhook_secret:
payload["webhook_secret"] = args.webhook_secret
for index, value in enumerate(args.input_image or [], start=1):
for index, value in enumerate(input_images if input_images is not None else (args.input_image or []), start=1):
key = "input_image" if index == 1 else f"input_image_{index}"
payload[key] = encode_input_image(value)
return payload
def receipt_payload(payload: dict[str, Any]) -> dict[str, Any]:
receipt = dict(payload)
for key, value in list(receipt.items()):
if key.startswith("input_image") and isinstance(value, str):
parsed = urllib.parse.urlparse(value)
if parsed.scheme in {"http", "https"}:
receipt[key] = value
else:
receipt[key] = f"<base64 image omitted: {len(value)} chars>"
return receipt
def poll_until_ready(
polling_url: str,
api_key: str,
@ -316,11 +341,34 @@ def extension_for(args: argparse.Namespace, sample_url: str | None) -> str:
return guessed.lstrip(".")
def submit_and_download(
args: argparse.Namespace,
api_key: str,
payload: dict[str, Any],
output_base: Path,
label: str,
) -> tuple[Path, dict[str, Any], dict[str, Any]]:
print(f"submitting {label}")
submit = request_json(args.endpoint, api_key, method="POST", payload=payload)
polling_url = submit.get("polling_url")
if not polling_url:
raise RuntimeError(f"BFL response did not include polling_url: {json.dumps(submit)}")
result = poll_until_ready(polling_url, api_key, args.poll_interval, args.timeout)
sample_url = (result.get("result") or {}).get("sample")
if not sample_url:
raise RuntimeError(f"BFL ready result did not include result.sample: {json.dumps(result)}")
image_path = output_base.with_suffix("." + extension_for(args, sample_url))
download_file(sample_url, image_path)
return image_path, submit, result
def run_job(args: argparse.Namespace, api_key: str | None, job: RenderJob, index: int, total: int) -> None:
output_dir = Path(args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
render_base = output_dir / f"{index:03d}-{job.render_id}"
payload = build_payload(args, job.prompt)
base_payload = build_payload(args, job.prompt)
metadata: dict[str, Any] = {
"render_id": job.render_id,
@ -328,39 +376,65 @@ def run_job(args: argparse.Namespace, api_key: str | None, job: RenderJob, index
"scene_id": job.scene_id,
"modifier_ids": list(job.modifier_ids),
"endpoint": args.endpoint,
"request_payload": payload,
"base_request_payload": receipt_payload(base_payload),
}
prompt_path = render_base.with_suffix(".prompt.txt")
prompt_path = render_base.with_suffix(".base.prompt.txt")
prompt_path.write_text(job.prompt + "\n", encoding="utf-8")
if job.edit_prompt:
edit_prompt_path = render_base.with_suffix(".edit.prompt.txt")
edit_prompt_path.write_text(job.edit_prompt + "\n", encoding="utf-8")
metadata["edit_prompt_path"] = display_path(edit_prompt_path)
if args.dry_run:
metadata["dry_run"] = True
if job.edit_prompt:
edit_payload = build_payload(args, job.edit_prompt, input_images=[])
edit_payload["input_image"] = "<base image generated by this job>"
metadata["edit_request_payload"] = edit_payload
write_json(render_base.with_suffix(".metadata.json"), metadata)
print(f"[{index}/{total}] dry-run wrote {display_path(prompt_path)}")
if job.edit_prompt:
print(f"[{index}/{total}] dry-run wrote {display_path(render_base.with_suffix('.edit.prompt.txt'))}")
return
if api_key is None:
raise SystemExit("Internal error: api_key missing outside dry-run mode.")
print(f"[{index}/{total}] submitting {job.render_id}")
submit = request_json(args.endpoint, api_key, method="POST", payload=payload)
metadata["submit_response"] = submit
polling_url = submit.get("polling_url")
if not polling_url:
raise RuntimeError(f"BFL response did not include polling_url: {json.dumps(submit)}")
print(f"[{index}/{total}] generating base {job.render_id}")
base_image_path, base_submit, base_result = submit_and_download(
args,
api_key,
base_payload,
render_base.with_name(render_base.name + ".base"),
f"{job.render_id} base",
)
metadata["base_submit_response"] = base_submit
metadata["base_poll_result"] = base_result
metadata["base_image"] = display_path(base_image_path)
result = poll_until_ready(polling_url, api_key, args.poll_interval, args.timeout)
metadata["poll_result"] = result
sample_url = (result.get("result") or {}).get("sample")
if not sample_url:
raise RuntimeError(f"BFL ready result did not include result.sample: {json.dumps(result)}")
if job.edit_prompt:
print(f"[{index}/{total}] editing from base {job.render_id}")
edit_payload = build_payload(args, job.edit_prompt, input_images=[str(base_image_path)])
metadata["edit_request_payload"] = receipt_payload(edit_payload)
edit_image_path, edit_submit, edit_result = submit_and_download(
args,
api_key,
edit_payload,
render_base.with_name(render_base.name + ".edit"),
f"{job.render_id} edit",
)
metadata["edit_submit_response"] = edit_submit
metadata["edit_poll_result"] = edit_result
metadata["edited_image"] = display_path(edit_image_path)
metadata["downloaded_image"] = display_path(edit_image_path)
write_json(render_base.with_suffix(".metadata.json"), metadata)
print(f"[{index}/{total}] wrote {display_path(edit_image_path)}")
return
image_path = render_base.with_suffix("." + extension_for(args, sample_url))
download_file(sample_url, image_path)
metadata["downloaded_image"] = display_path(image_path)
metadata["downloaded_image"] = display_path(base_image_path)
write_json(render_base.with_suffix(".metadata.json"), metadata)
print(f"[{index}/{total}] wrote {display_path(image_path)}")
print(f"[{index}/{total}] wrote {display_path(base_image_path)}")
def parse_args() -> argparse.Namespace:

1
state/evidence.jsonl
View file

@ -59,3 +59,4 @@
{"ts": "2026-05-03T10:55:19+00:00", "type": "review", "status": "ok", "note": "Ran narrative and visual reviewer subagents on the Pallas fixture, patched required revisions, scoped global visual modifiers by scene, added the red-alert worker-judgment visual beat, and resolved the final Orrin/Nara route-memory staging mismatch."}
{"ts": "2026-05-03T10:58:29+00:00", "type": "doctrine", "status": "ok", "note": "Corrected visual reviewer scope: it reviews playable Ink plus .visual.json for illustrated replay and website value, ignores clean-run prose receipts, and treats visual artifacts as auxiliary corpus presentation rather than core social-model training signal."}
{"ts": "2026-05-03T11:53:40+00:00", "type": "implementation", "status": "ok", "note": "Added BFL image generation script for direct prompts and Ghostlight .visual.json plans. It reads BFL_API_KEY or gamecult-ops/bfl-api.txt, supports dry-run prompt receipts, submits FLUX.2 pro preview jobs, polls returned polling_url, downloads signed result.sample images immediately, and keeps generated outputs under ignored experiments/images/."}
{"ts": "2026-05-03T12:00:19+00:00", "type": "implementation", "status": "ok", "note": "Updated BFL visual-plan rendering so modifiers use the intended two-pass continuity workflow: generate the base image from the base prompt, then edit that saved image with only the selected modifier prompt. Dry-run receipts now emit separate .base.prompt.txt and .edit.prompt.txt files."}