# Four 5-second text-to-video clips with sound, one RTX 5090 32 GB, 4-5 Nov 2026. 5 prompts (P1-P5), seed 42, one run per prompt, one warm-up run per model (seed 1, not in the median).
# Time = wall clock from request to finished MP4 (ComfyUI: submit to saved file; Kandinsky: wall_s of the driver). Median of 5. Clips shown here: prompt P2 (rally car).
# Per-job files: jobs/<model>_P<n>.json (wall_seconds / wall_s, full ComfyUI graph for the three ComfyUI models). Summary: timings.json.

# MiniMax-H3 base, ComfyUI 0.44.0 (master 7ad7adb3), template video_minimax_h3_t2v:
#   UNet minimax_h3_fl2va_pruned_int8_convrot, text encoder qwen3vl_32b_minimax_h3_nvfp4_awq (type minimax),
#   video VAE minimax_h3_video_vae_int8_convrot, audio VAE minimax_h3_audio_vae_fp32
#   MiniMaxH3ImageToVideo 1344x768, length 124 frames, 24 fps | RandomNoise seed 42
#   KSamplerSelect res_multistep | BasicScheduler simple, 20 steps | BasicGuider (no CFG) | SamplerCustomAdvanced
#   prompt: prompts/P2-action-rallyauto.h3.txt (three fields: integrated_multimodal_description, overall_soundscape, non_diegetic_music)

# MiniMax-H3 + Pika-H3-Flex (6 steps), same graph as above plus:
#   LoraLoaderModelOnly pika_h3_flex_fl2va_v1_bf16, strength 1.0
#   KSamplerSelect lcm | ManualSigmas "1.0, 0.972973, 0.923077, 0.8, 0.571429, 0.333333, 0.0" (6 steps) | BasicGuider (no CFG)

# LTX-2.5 22B distilled, ComfyUI 0.44.0, our own earlier two-stage ComfyUI workflow (not an author template):
#   UNet ltx-2.5-22b-distilled-transformer-comfy-int8-convrot, text encoder gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot (type ltxv),
#   video VAE ltx-2.5-video-vae-bf16, audio VAE ltx-2.5-audio-vae-bf16, latent upscaler ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0
#   stage 1: 672x384, 121 frames, 24 fps, euler_ancestral, ManualSigmas "1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0" (8 steps), LTXVDualCFGGuider video/audio CFG 1.0
#   latent x2 upsample to 1344x768, stage 2: ManualSigmas "0.85, 0.75, 0.62, 0.48, 0.32, 0.0" (5 steps), seed 42 both stages
#   negative prompt: blurry, low detail, warped geometry, melting shapes, extra limbs, deformed hands, flicker, jitter, frozen frame, still image, text, watermark, logo
#   prompt: prompts/P2-action-rallyauto.ltx.txt

# Kandinsky 6.0 Pro-distill 5s, official inference repo with its rtx-5090 device preset (not ComfyUI):
python3 bench_k6.py --config kandinsky/configs/devices/rtx-5090.yaml --tag pro --prompts <prompts-k6.json> --outdir <clips-dir> --only P2-action-rallyauto
#   864x480, 121 frames, 24 fps, 10 denoising steps, seed 42, block offload, SageAttention, prompt beautifier on (per the preset)
#   prompt: prompts/P2-action-rallyauto.kandinsky.txt (sound marked with <AUDCAP> ... <ENDAUDCAP>); the built-in prompt beautifier rewrote it before generation: prompts/P2-action-rallyauto.kandinsky.beautified-by-model.txt; step count: kandinsky-run-excerpt.txt