Current findingVerified

Local best verified configuration

Frame
1024×576
Length
243 frames
Sampling
20 steps
Duration
10.125 sec
Generation
46m 03s

A complete, decodable result produced locally on an RTX 4060 Laptop 8GB — not the absolute hardware limit.

A/B synchronized comparison workbench

Align by progress or clock time to compare narrative, clarity, and stability directly.

Choose other runs
A

Final C / 20 steps

1024×576 / 243 frames / 10.125 sec

View 1:1
B

C tier / 4 steps

1024×576 / 243 frames / 10.125 sec

View 1:1
A/B synchronized playback

Reading video metadata…

0% / A 00:00 / B 00:00
Synchronization mode
Audio source

Resolution

A 1344×768A has more pixelsB 1024×576

Frames

A 124 framesB has more framesB 243 frames

Duration

A 5.167 secB is longerB 10.125 sec

Sampling

A 4 stepsEqual sampling stepsB 4 steps

Generation time

A 537.49 secA generated fasterB 650.96 sec

Result

A SuccessBoth runs verifiedB Success
Published runsAssign any run to A or B
5 videos · 2 contact sheets
  • 01
    Final C / 201024×576 · 243 frames2762.89s · Success
  • 02
    C tier / 41024×576 · 243 frames650.96s · Success
  • 03
    B-short / 41344×768 · 124 frames537.49s · Success
  • 04
    C preflight / 11024×576 · 243 framesTiming not recorded · Success
  • 05
    B-short preflight1344×768 · 124 framesTiming not recorded · Success

Choose the shot objective before the parameters

This laptop cannot maximize resolution, duration, and sampling steps at the same time. Each practical target below has a result you can inspect.

Best verified

C / 20 steps

The strongest verified balance of ten-second narrative, motion, and final detail.

Resolution first

B-short / 4 steps

Use the higher 1344×768 frame when a five-second shot is enough.

Length first

C / 4 steps

Use the full ten seconds to validate shot rhythm before spending on sampling.

Fast preflight

1 step

Check composition, motion direction, and workflow connectivity first.

All results in one player

All five videos and two contact sheets retain their original files. The viewer fits media to the window by default and can switch to native 1:1 pixels.

Seven runs reveal one real boundary

With the same model and seed, spatial pixels and temporal frames decide whether a run can start; sampling steps mainly change generation time and detail.

MiniMax H3 local configuration results
ExperimentResolutionFrames & durationStepsGeneration timeResultVRAM evidenceRAM / private memoryMP4 sizeMeasurement basis
A upper bound1344×768362 / 15.08s1Failed before outputOOM≈4.32 GiB allocated + 4.38 GiB requestPeak not recordedRuntime error log
B-long1344×768243 / 10.13s1StalledStalled≈7.9 / 8.2 GB≈27.8 GiB privateProcess / GPU observation
B-short preflight1344×768124 / 5.17s1Not recordedSuccessCompleted within 8188 MiB; peak not recordedPeak not recorded6.38 MiBOutput file measurement
B-short1344×768124 / 5.17s4537.49sSuccessCompleted within 8188 MiB; peak not recordedPeak not recorded4.83 MiBRuntime + output file
C preflight1024×576243 / 10.13s1Not recordedSuccessPeak not recordedPeak not recorded16.94 MiBOutput file measurement
C balanced1024×576243 / 10.13s4650.96sSuccessPeak not recordedPeak not recorded8.42 MiBRuntime + output file
Final result1024×576243 / 10.13s202762.89sRecommended≈6.46 GB observed peakPeak not recorded6.82 MiBGPU observation + output file

CPU utilization and power were not continuously sampled; successful-run RAM peaks and several VRAM peaks were not recorded.

What the failed runs reveal

More steps increase runtime but barely change peak VRAM. The real threshold is the activation scale created by spatial pixels and temporal frames.

4.38GiB

VRAM hard wall

At 1344×768 and 362 frames, one attention QKV allocation exceeds the remaining VRAM.

Allocated
4.32GiB
Outcome
CUDA OOM
27.8GiB

System-memory unloading stall

At 1344×768 and 243 frames, sampling begins but model unloading cannot keep progressing.

GPU
7.9 / 8.2GB
Activity
0 CPU / 0 GPU
10.125s

Recoverable work zone

At 1024×576 and 243 frames, enough margin remains to finish both VAEs and muxing.

GPU peak
≈ 6.46GB
Final steps
20

Complete generation also means complete decoding

The final file was read frame by frame for video and audio, with queue, history, encoding, and content statistics checked independently.

Video243 / 243

H.264 / yuv420p / 24fps

Audio317 frames

AAC / 32kHz / Stereo

Motion15.979

Mean adjacent-frame change; not black or frozen

Run2762.89s

End-to-end local RTX 4060 Laptop run

Final MP4 SHA256BA978D87ECAD4DDF5A099FF842AA1E487FDF55D97F7E7C1051F4969ABF5856D5

Information required to reproduce the run

The video node graph keeps the same seed and sampling path while changing only resolution, frame count, and steps.

Hardware and runtime

System
Lenovo 21J8
CPU
Intel Core i9-13900H
CPU topology
14 cores / 20 threads
GPU
NVIDIA GeForce RTX 4060 Laptop GPU · 8188 MiB
GPU driver
581.08
RAM
2 × 16 GiB Samsung DDR5-5600
Configured memory speed
5200 MT/s
Storage
SAMSUNG MZVL21T0HCLR-00BL2 · 1 TB NVMeWD PC SN740 SDDPTQE-2T00 · 2 TB NVMe
Workspace
E: · WD 2 TB NVMe
OS
Windows 11
Python
3.13.9
ComfyUI
0.30.0
PyTorch
2.13.0 + CUDA 13.0
Launch
--fast-disk --reserve-vram 2

Measured local footprint

Model files
39.554 GiB
Python environment
4.041 GiB
Published site copy
46.58 MiB
Recommended free space
≥ 60 GiB

The 60 GiB figure is operating headroom, not measured file use. Keep models and temporary generation data on the 2 TB E: drive because the system drive was nearly full during this audit.

Model stack

  • Diffusionminimax_h3_fl2va_pruned_int8_convrot.safetensors
  • Textqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
  • Video VAEminimax_h3_video_vae_fp16.safetensors
  • Audio VAEminimax_h3_audio_vae_fp32.safetensors

Sampling parameters

Samplerres_multistep
Schedulersimple
Seed20260805
Denoise1.0
View the final English prompt
High-end Unreal Engine 5 gameplay footage of a 3D side-scrolling action platform game starring a heroic chubby black-and-white kung fu panda warrior. One continuous side-view tracking shot through a lush ancient Chinese bamboo mountain valley. Detailed fur simulation, physically based materials, global illumination, volumetric sunlight, mist, flying bamboo leaves, polished AAA game graphics, expressive animation, powerful readable combat impacts.

Timeline:
[0s-2s] The panda sprints left to right on a mossy stone path and leaps across a broken bridge over a misty ravine.
[2s-4s] He rolls under a swinging wooden trap, jumps over spikes and lands in a kung fu stance.
[4s-7s] Armored wolf soldiers attack; he blocks, punches, palm-strikes and kicks them away with dust and leaf impact effects.
[7s-10s] A large armored yak boss swings a hammer; the panda dodges, runs up a bamboo ramp and defeats it with a slow-motion spinning flying kick in golden sunset light.

Camera: strict 2.5D side view, smooth horizontal tracking at constant distance, one uninterrupted take, no cuts, no zoom, no orbit, no first-person view. No HUD, menu, subtitles, captions, logos, watermark, written text, live action, extra limbs or deformed anatomy.

Audio: synchronized footsteps, cloth movement, bamboo forest wind, bridge creaks, landing impacts, trap whooshes, crisp kung fu punches and kicks, armor hits, a deep boss hammer impact, restrained cinematic Chinese drums and low strings. No dialogue, narration or vocals.
MEDIAMedia viewer
1 / 7