README.md

metadata

title: ShotCraft
emoji: 🎬
colorFrom: purple
colorTo: pink
sdk: gradio
app_file: app.py
license: apache-2.0
short_description: AI art director for e-commerce product shots
tags:
  - backyard-ai
  - modal

🎬 ShotCraft — AI Shot Director

Upload a product photo → an 8B vision model designs 5 marketing shot concepts grounded in your actual product → a 12B image model renders them as production-ready stills.

Build Small Hackathon 2026 | Models: MiniCPM-V-2_6 (8B) + FLUX.1-schnell (12B) — both ≤32B ✅

Pipeline

1. Upload product photo + brand description + category + style preset
2. MiniCPM-V-2_6 analyzes the photo → 5 shot concepts (scene, camera, lighting, palette, props, marketing angle, optimized prompt)
3. Review & edit concepts
4. FLUX.1-schnell renders 5 stills → pick hero frames → export ZIP

Team: Pawel (Stage 1) · Rafal (Stage 2) · Stefan (integration)

Architecture — powered by Modal ⚡

This Space is the Gradio interface; all model inference runs on Modal (modal_backend/shotcraft_inference.py in the repo):

Stage	Model	Modal GPU
1 — Shot Director	MiniCPM-V-2_6 (8B, bf16)	A10G
2 — Frame Generator	FLUX.1-schnell (12B, bf16, 4-step)	L40S

• One HTTP call per stage; the whole 5-frame reel renders in a single batched call, per-frame regen is a 1-frame call (seeded, FR-2.3/2.4).
• Weights cached on a Modal Volume; containers stay warm 5 min between calls.
• The Space needs SHOTCRAFT_API_URL → the Modal endpoint. The deploy workflow can set it automatically from SHOTCRAFT_MODAL_WORKSPACE, SHOTCRAFT_MODAL_APP, and SHOTCRAFT_MODAL_FUNCTION.

Dev notes

• Backend deploy: modal deploy modal_backend/shotcraft_inference.py (or modal serve for a dev loop). Health check: GET /health.
• Tests: python test_smoke.py (pure logic, no backend) and python test_app_e2e.py (real end-to-end against the deployed backend — spends GPU time).
• Plan: see docs/shotcraft-IMPLEMENTATION_PLAN.md for the slice-by-slice build order.

Deployment configuration

Keep personal tokens in GitHub Secrets, not in code. This shared repo deploys one shared Hugging Face Space, but can point that Space at Rafal's or Pawel's Modal backend. Manual workflow runs ask for deployment_profile (rafal or pawel). Pushes to main still deploy the default rafal profile.

Rafal uses the existing GitHub Secrets:

• HF_TOKEN — Hugging Face write token for the target Space.
• MODAL_TOKEN_ID / MODAL_TOKEN_SECRET — Modal token for the target workspace.

Pawel uses prefixed GitHub Secrets for Modal only:

• PAWEL_MODAL_TOKEN_ID
• PAWEL_MODAL_TOKEN_SECRET

Rafal/default GitHub Variables:

• HF_USERNAME — username used for HF git auth. Defaults to rafalbog.
• HF_SPACE_REPO — Space repo id, for example build-small-hackathon/ai-video-generation.
• SHOTCRAFT_MODAL_WORKSPACE — Modal workspace slug. Defaults to rafalbogusdxc.
• SHOTCRAFT_MODAL_APP — Modal app name. Defaults to shotcraft-inference.
• SHOTCRAFT_MODAL_FUNCTION — Modal ASGI function name. Defaults to api.
• SHOTCRAFT_API_URL — full backend URL override. If unset, the workflow builds https://<workspace>--<app>-<function>.modal.run.
• SHOTCRAFT_HF_SECRET_NAME — Modal secret that contains the HF token for model downloads. Defaults to huggingface-secret.
• SHOTCRAFT_HF_CACHE_VOLUME — Modal volume for cached weights. Defaults to shotcraft-hf-cache.

Pawel-specific GitHub Variables use the same names with a PAWEL_ prefix:

• PAWEL_SHOTCRAFT_MODAL_WORKSPACE
• optional: PAWEL_SHOTCRAFT_API_URL
• optional: PAWEL_SHOTCRAFT_MODAL_APP
• optional: PAWEL_SHOTCRAFT_MODAL_FUNCTION
• optional: PAWEL_SHOTCRAFT_HF_SECRET_NAME
• optional: PAWEL_SHOTCRAFT_HF_CACHE_VOLUME

For Rafal, the existing repository-level secrets/variables can stay as-is. For Pawel, add the prefixed Modal secrets/variables under the same repository Actions settings page, then run either deploy workflow manually and choose deployment_profile=pawel. The frontend workflow still uses the existing HF_TOKEN to update the shared Space at build-small-hackathon/ai-video-generation.