Add peft to requirements (needed by LCM-LoRA load_lora_weights)
Matys BIECHE+1: Claude Opus 4.8 (1M context)
faec225
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add 20 WarioWare-style microgame templates with modular canvas engine
Matys BIECHE+1: Claude Opus 4.8 (1M context)
5214e3b
Replace the 3 hardcoded templates with a 20-game registry (grab/catch/whack/
find/swat/shoot/protect/dodge/chase/drag/balance/mash/pump/hold/rhythm/stop/
quickdraw/sort + timing/avoid). All share one asset contract (bg + target +
2 decoys). LLM prompt updated to pick a varied template and keep instruction
coherent with the target object. Engine rewritten as plain string (no f-string
brace escaping) with per-game setup/frame/click/move/key/timeout hooks.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Better game quality: instruction<->image coherence, retries, white cutout
Matys BIECHE+1: Claude Opus 4.8 (1M context)
a313479
- Enforce in system prompt that instruction's object == target_prompt object
- Retry spec generation 3x (decreasing temperature) before example fallback
- Cut out white background of sprites (border floodfill + alpha) so the named
object actually stands out on the canvas instead of blending into the card
- Clearer object prompts (single, centered, fully visible, large)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Speed up SDXL: LCM-LoRA (5 steps) + batched image generation
Matys BIECHE+1: Claude Opus 4.8 (1M context)
c04d9b2
Replace 18-step CFG=7 SDXL with LCM-LoRA scheduler (~5 steps, guidance 1.2)
on the same base model. Batch target+2 decoys into a single GPU call instead
of 4 sequential calls. ~4-6x faster asset generation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add .gitignore for pycache
Matys BIECHE
f67d493
Add ZeroGPU-ready Gradio app for HF Space
Matys BIECHE+1: Claude Opus 4.8 (1M context)
6b66235
Adapt copie_de_poc.py into app.py: remove Colab pip-install/share/debug,
add spaces lib with @spaces.GPU decorator, load models on cuda at startup,
queue() the demo. Add requirements.txt.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
