storytold/cloud-worker
2
0
Fork 0
mirror of https://github.com/storytold/cloud-worker.git synced 2026-10-15 00:04:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

bench: v2 reference set — JP stills, Ashitaka/Bebop, forests, volcanos (up to 23k px)

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Hanashi
2026-08-06 02:45:34 -04:00
parent 94a94a268f
commit 1debde7673
2 changed files with 104 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
bench/minimax_bench.py
+12 -9
View File
@@ -41,12 +41,14 @@ VIDEO_VAE = "minimax_h3_video_vae_fp16.safetensors"
AUDIO_VAE = "minimax_h3_audio_vae_fp32.safetensors"
FL2VA_DEFAULT = "minimax_h3_fl2va_pruned_int8_convrot.safetensors"
REF2VA_DEFAULT = "minimax_h3_ref2va_pruned_int8_convrot.safetensors"
# Real images fetched by scripts/pod/fetch-bench-images.py (anime characters,
# film stills, giant panoramas) — realistic detail load for the ref encoder.
BENCH_IMAGE = "bench_movie_01.jpg"
BENCH_REF_SET = ["bench_anime_01.jpg", "bench_anime_02.jpg", "bench_anime_03.jpg",
"bench_movie_01.jpg", "bench_movie_02.jpg", "bench_movie_03.jpg",
"bench_big_01.jpg", "bench_big_02.jpg"]
# Real images fetched by scripts/pod/fetch-bench-images-v2.py: Jurassic Park
# stills, anime characters (Ashitaka, Spike), forests, volcanos — large files
# (up to 23040x3840) for realistic reference-encoder load.
BENCH_IMAGE = "bench2_jp_01.jpg"
BENCH_REF_SET = ["bench2_jp_01.jpg", "bench2_jp_gate_01.png",
"bench2_ashitaka_01.jpg", "bench2_bebop_01.jpg",
"bench2_forest_01.jpg", "bench2_forest_02.jpg",
"bench2_volcano_01.jpg", "bench2_volcano_02.png"]
T2V_PROMPT = (
"Cinematic aerial shot slowly orbiting a coastal lighthouse at golden hour, "
@@ -68,9 +70,10 @@ def ref2v_prompt(ref_count):
return REF2V_PROMPT
tags = ", ".join(f"<Picture {i}>" for i in range(1, ref_count + 1))
return (
f"A sweeping cinematic montage that travels through the {ref_count} city scenes "
f"from {tags} in order, crossfading between their times of day and color moods, "
"camera drifting forward the whole time, ambient city sounds evolving with each scene."
f"A sweeping cinematic montage inspired by {tags}: dinosaurs stalking through "
"ancient forests, an anime hero surveying a volcanic ridge, crossfading between "
"the scenes and moods of each reference in order, camera drifting forward the "
"whole time, an adventurous orchestral score building throughout."
)