diff --git a/tools/sprite-generation/experiments/2026-03-27_model_comparison.yaml b/tools/sprite-generation/experiments/2026-03-27_model_comparison.yaml new file mode 100644 index 00000000..76078c2e --- /dev/null +++ b/tools/sprite-generation/experiments/2026-03-27_model_comparison.yaml @@ -0,0 +1,64 @@ +# Model Comparison Experiments — 2026-03-27 +# Testing which SDXL model produces game-ready unit sprites +# Target: overhead isometric, green chroma key, painted fantasy style, southwest-facing + +experiment: model_comparison +date: 2026-03-27 +sprite: spearmen_dwarves_f +prompt_version: "YAML-driven v1 (unit_classes.yaml + composition.yaml)" + +models: + juggernaut-xl-v9: + full_name: "Juggernaut-XL_v9_RunDiffusionPhoto_v2" + file_size_gb: 6.9 + vram_mb: ~8500 + sdxl_base: true + results: + total_generated: 18 + gates_passed: 0 + common_failures: + - "facing_southwest: always faces camera or forward" + - "no_shadow: always renders cast shadows" + - "no_base_or_ground: always adds ground plane" + - "not_photorealistic: renders 3D photorealistic (it's a PHOTO model)" + - "green_background: produces blue/purple/teal backgrounds" + strengths: + - "correct_subject_type: SPEAR + SHIELD always present (unit_classes.yaml works)" + - "anti-green-bleed: copper/brown/silver armor (NON-GREEN directive works)" + verdict: "REJECT — fundamentally a photorealistic model, can't produce painted game art" + note: "Name literally says 'RunDiffusionPhoto' — wrong model for stylized game art" + + juggernaut-xi-v11: + full_name: "Juggernaut-XI-byRunDiffusion" + file_size_gb: 7.1 + vram_mb: ~9000 + sdxl_base: true + results: + total_generated: "~800 (across many prompt iterations)" + common_failures: + - "facing_southwest: ~50% hit rate (SDXL limitation)" + - "no_shadow: sometimes renders shadows (improved with negative)" + - "green armor bleed: characters pick up green from chroma background" + strengths: + - "Produces painted/stylized art when prompted correctly" + - "Responds to 'close-up isometric view from above'" + - "Green chroma key background works ~70% of time" + - "Better at non-photorealistic styles than v9" + verdict: "CURRENT BEST — not perfect but produces usable painted art" + note: "Earlier loading crash was VRAM contention, not model issue" + + # Models NOT YET TESTED (available in model-boss manifest): + epicrealism-xl: + status: untested + notes: "Realism-focused, likely similar issues to v9" + + illustrious-xl-v2: + status: untested + notes: "Anime/illustration focused — CLAUDE.md says NEVER use anime models" + +# Key findings: +# 1. SDXL models with "Photo" in name can't produce painted game art +# 2. SDXL can't reliably control facing direction (~50% southwest hit rate) +# 3. Green chroma key causes color bleed onto subject — anti-bleed negatives help +# 4. Unit class weapon descriptions (SPEAR, LONGBOW) work well when placed early in prompt +# 5. v11 is the right model but needs model-boss VRAM stability to load reliably