video_minimax_h3_r2v.json explained

R2V uses an image, video or audio as a semantic and motion reference for a newly generated clip. It does not guarantee frame-by-frame alignment, and it does not provide the denoise-strength control found in traditional diffusion editing.

The template may reconstruct the scene, crop the reference, or reorder actions and shots. If the job requires frame-by-frame preservation, an exact lip-sync timeline or a pixel-level mask, evaluate a dedicated video-editing tool for that requirement.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── minimax_h3_ref2va_pruned_int8_convrot.safetensors
    ├── text_encoders/
    │   └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    └── vae/
        ├── minimax_h3_video_vae_fp16.safetensors
        └── minimax_h3_audio_vae_fp32.safetensors

The R2V template loads minimax_h3_ref2va_pruned_int8_convrot.safetensors, approximately 20.97GB. Ref2VA and the FL2VA file loaded by T2V and I2V are separate diffusion-model downloads.

Sources: ComfyUI MiniMax H3 tutorial and Comfy-Org MiniMax-H3 repository. Checked 2026-08-13.

Conditions for every figure in this test card: physical GPU 0 was an RTX 3060 12GB; Proxmox VE VM 100 had 16 vCPU, 47.05 GiB RAM and no swap; ComfyUI 0.31.0 used PyTorch 2.13.0+cu130; the official R2V template loaded Ref2VA pruned INT8 at 1344×768 for 124 frames with a fixed seed; audio, Turbo LoRA and SageAttention were not enabled.

The two output SHA-256 values were identical. In this fixed environment with a fixed seed, the result was deterministically reproducible.

Conditions for this same-load incident telemetry: physical GPU 1 was used for the failed attempt and physical GPU 0 for the comparison; Proxmox VE VM 100 had 16 vCPU, 47.05 GiB RAM and no swap; ComfyUI 0.31.0 used PyTorch 2.13.0+cu130; the official R2V template loaded Ref2VA pruned INT8 at 1344×768 for 124 frames with a fixed seed; audio, Turbo LoRA and SageAttention were not enabled.

One attempt on physical GPU 1 froze the entire host VM and required a manual cold start. The root cause was abnormal cooling at that card's physical location: under the same load, GPU 1 reached 79 °C while GPU 0 reached 60 °C, a 19 °C difference. After moving the work to GPU 0, both formal runs completed without the fault.

The root cause was not memory exhaustion. Sampling during the freeze showed MemAvailable on an approximately 4.9GB plateau, without the pattern of progressive depletion.