niklaslenzHyperFlow H3 Flash in ComfyUI: 5-second clips with sound on an RTX 5090 and on a 16 GB RTX 5060 Ti
HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax-H3. In ComfyUI it makes a 5-second clip with sound in 86.8 s on one RTX 5090. The GGUF Q4_K_M file also runs on a 16 GB RTX 5060 Ti, at 562.2 s per clip. Below are the real clips of both cards with prompt and time.
Not ranked. Measured on the devices shown here only. No rating and no rank: this result has no rating by me yet.
RTX 5090 32 GB · INT8 · 6 steps
one RTX 5090 32 GB (Ryzen 9 9950X, 96 GB RAM), ComfyUI 0.44.0, INT8 single file for ComfyUI, server started with --enable-triton-backend
RTX 5090 32 GB · GGUF Q4_K_M · 6 steps
one RTX 5090 32 GB (Ryzen 9 9950X, 96 GB RAM), ComfyUI 0.44.0 with ComfyUI-GGUF, GGUF Q4_K_M file, default server start
RTX 5060 Ti 16 GB · GGUF Q4_K_M · 6 steps
one RTX 5060 Ti 16 GB (Core i7-12700K, 32 GB RAM), ComfyUI 0.44.0 with ComfyUI-GGUF, GGUF Q4_K_M file, server started with --fast-disk
The clips
Text to video, prompt 1: ramen counter (seed 42)
Prompt
A chef slides a steaming bowl of ramen across a wooden counter; rain taps on the window.
Prompt
A chef slides a steaming bowl of ramen across a wooden counter; rain taps on the window.
Prompt
A chef slides a steaming bowl of ramen across a wooden counter; rain taps on the window.
Text to video, prompt 2: street musician (seed 42)
Prompt
A street musician plays a weathered acoustic guitar on a cobblestone corner at golden hour. The camera slowly pushes in from a wide shot to a medium close-up as passers-by drift through the warm backlight. Soft strumming, distant chatter and a passing tram bell fill the soundtrack.
Prompt
A street musician plays a weathered acoustic guitar on a cobblestone corner at golden hour. The camera slowly pushes in from a wide shot to a medium close-up as passers-by drift through the warm backlight. Soft strumming, distant chatter and a passing tram bell fill the soundtrack.
Prompt
A street musician plays a weathered acoustic guitar on a cobblestone corner at golden hour. The camera slowly pushes in from a wide shot to a medium close-up as passers-by drift through the warm backlight. Soft strumming, distant chatter and a passing tram bell fill the soundtrack.
Image to video: autumn forest path (seed 42, start image below)
Prompt
The scene comes alive: leaves drift down through the light, the camera glides forward along the path, and birdsong and a gentle breeze fill the air.
Prompt
The scene comes alive: leaves drift down through the light, the camera glides forward along the path, and birdsong and a gentle breeze fill the air.
Prompt
The scene comes alive: leaves drift down through the light, the camera glides forward along the path, and birdsong and a gentle breeze fill the air.
Start image for the image-to-video clips
Time per 5-second clip at 1344×768 and 6 steps. My medians over warm clips next to the numbers the author publishes. The machines differ, so this is a cross-check and not a rank list.
| Setup | Text to video, to finished MP4 | Text to video, without saving | Author, without saving | Image to video, to finished MP4 | GPU memory peak | Author, peak VRAM |
|---|---|---|---|---|---|---|
| RTX 5090 32 GB, INT8 | 86.8 s | 85.6 s | 83 s | 96.4 s | 30.8 GiB | 31 GiB |
| RTX 5090 32 GB, GGUF Q4_K_M | 107.1 s | 105.7 s | 96 s and 99 s | 116.1 s | 29.4 GiB | 28 GiB |
| RTX 5060 Ti 16 GB, GGUF Q4_K_M, --fast-disk | 562.2 s | 560.2 s | not measured by the author | 608.1 s | 14.9 GiB | not measured by the author |
How it was run
- Two boxes with one GPU each: an RTX 5090 32 GB (Ryzen 9 9950X, 96 GB RAM) and an RTX 5060 Ti 16 GB (Core i7-12700K, 32 GB RAM). Measured on 9 Nov 2026.
- Software on both boxes: ComfyUI 0.44.0, PyTorch 2.14.1 with CUDA 13.0, the author's node ComfyUI-HyperFlow-Flash 1.1.0 (commit c807a89). The GGUF file is loaded by the author's HyperFlow Loader node. ComfyUI-GGUF by city96 (commit 6ea2651) is installed for it, as the author's guide asks for GGUF files.
- The author's API workflows, changed only in prompt, seed, output name and start image name: 1344×768, 124 frames at 24 fps (5.17 s), 6 steps, sampler euler, no CFG, stereo sound at 32 kHz.
- Version 1, INT8: minimax_h3_fl2va_hyperflow_flash_14b-int8_convrot.safetensors, 16.01 GB, sha256 d38f0c0c4682. Run on the RTX 5090 only.
- Version 2, GGUF Q4_K_M: minimax_h3_fl2va_hyperflow_flash_14b-Q4_K_M.gguf, 9.05 GB, sha256 ee251661dffd. Run on both cards.
- Both from VideoRebirthAI/hyperflow-h3-flash, revision 61f1311b. With them: text encoder qwen3vl_32b_minimax_h3_nvfp4_awq (15.69 GB, sha256 35a88d510442), video VAE fp16 (5.21 GB, sha256 7c1f131492e7) and audio VAE fp32 (0.61 GB, sha256 8e505d95dd15) from Comfy-Org/MiniMax-H3. Every sha256 matches Hugging Face.
- Time is the wall clock from request to finished MP4 with sound. Cold is the first clip after a fresh server start, with model load. Warm is every later clip on the same server. Medians are over warm clips only.
- Inputs: two text prompts from the author (ramen counter, street musician) and the author's image-to-video prompt. The start image is my own, made with Qwen-Image-2.1, seed 42.
- RTX 5090: a fresh server and 10 clips per version, all with seed 42. The INT8 server was started with --enable-triton-backend, the GGUF server with no extra flag.
- RTX 5060 Ti 16 GB: GGUF Q4_K_M only, one server for 9 clips, started with --fast-disk. That is the author's start line for PCs with 16 or 32 GB of RAM. Five clips use seed 42, four use seeds 43 and 44.
What to keep in mind
Measured on these two cards only, not ranked. There is no quality score on this page. It shows the clips and the times.
The author has not measured a 16 GB GPU. The model card says: "16 GB GPUs are not measured yet." The 16 GB numbers here are my own run with the author's start line for small PCs.
On the 16 GB card one clip takes about 5 times as long as on the RTX 5090 with the same file (562.2 s against 107.1 s).
Where the time goes on the 16 GB card (medians of the warm text-to-video clips): sampling 392.0 s for 6 steps, prompt encoding with loading the text encoder 88.2 s, video decode 76.6 s. The transformer does not fit in full. The server log shows 6,792 MB loaded on the GPU and 2,145 MB offloaded. The text encoder is loaded again for every clip.
Forced deviations on the 16 GB box: the model files were on a SATA hard disk, the only disk with enough free space. The author recommends an NVMe SSD with --fast-disk. The box also runs other services. Before the first clip 19.6 of 31.2 GiB of host RAM were free, and swap use peaked at 4,096 MiB. A PC with an NVMe SSD and free RAM may be faster. I did not test that.
Compared with the author on the RTX 5090, both without saving the MP4: INT8 85.6 s here and 83 s at the author. GGUF Q4_K_M 105.7 s here, 96 s and 99 s at the author. That is 3 % and 7 to 10 % slower. I did not look for the cause. Known differences: PyTorch 2.14.1 here against 2.13 at the author, and a different machine.
GPU memory peak on the RTX 5090: 30.8 GiB with INT8 and 29.4 GiB with Q4_K_M. ComfyUI fills free memory on purpose, so this is not the minimum the model needs. The author states 31 GiB and 28 GiB.
On the RTX 5090 the server ran without --fast-disk. The log still prints "Model storage policy: fast_disk=True", and for the GGUF file "transformer weights in disk (weights_in=auto)". I did not check what that means for speed.
Seeds: every clip on the RTX 5090 uses seed 42, so the warm clips repeat the same three inputs. On the 16 GB card four of the six warm text-to-video clips use seeds 43 and 44. Repeats of the same prompt and seed on one card gave the same sound each time. The picture was identical in some repeats and slightly different in others. The MP4 files of the repeats are not bit-identical.
Image to video on the 16 GB card has only 2 warm clips. Read the 608.1 s as a first value.
Shown are 9 of the 29 timed clips: for each setup the first warm clip of each input with seed 42. The other 20 clips are in the timings file. I did not pick clips by look.
Not measured: 15-second clips, reference to video, INT8 on the 16 GB card, and any score for picture or sound.
A quality hint from the author for the two files, not checked by me: "against the `BF16` GGUF (one H200, same seed, `t2v` 5.2 s), mean frame PSNR is 18.0 dB for INT8, 17.5 dB for `Q8_0`, and 14.8 dB for `Q4_K_M`."
License: the weights come under the MiniMax H3 Community License Agreement (see the model card). The code is Apache-2.0, the ComfyUI node GPL-3.0.
Exact command · All timings (every clip and the medians) · Per-clip job files (times, memory peaks, full ComfyUI graph) · Server logs, run logs and system stats · Prompts · Model files with sha256 · The author's published numbers, word for word · ffprobe of the original clips · sha256 of the original clips · All raw files