Results · Watch the clips · INT8 explained · Downloads and U.S. license request
What can the computer you already own actually do?
Can a consumer GPU with only 12 GB of video memory, paired with 32 GB of system RAM, do useful local AI video work? On the HP OMEN desktop tested here, the answer was yes: it completed five-second clips at 480p, 720p and 1080p. The waiting time changed dramatically with the settings.
This is a case study of working within a constrained computer’s limits. It shows what this particular machine accomplished before any hardware upgrade. It is not a claim that an RTX 5070 is the minimum supported GPU, or that every 12 GB card will perform the same way.
The practical opportunity is to start making and judging short takes on available hardware. The practical limit is that generating a video file and getting the performance a story needs are separate achievements.
Five seconds of video in about one minute
A five second image to video shot completed in 61.65 seconds on an HP OMEN desktop with an NVIDIA RTX 5070 with 12 GB VRAM and 32 GB of system RAM. This result used MiniMax H3 INT8 with the four step Turbo adapter at 480p. It demonstrates a useful starting point for rapid shot auditions on this particular computer. It does not establish a universal render time or final production quality.
The practical question for an independent filmmaker is how quickly an idea can become footage that can be judged in an edit. Waiting roughly a minute for a candidate take changes that interaction compared with waiting nearly twelve minutes for the tested 1080p Turbo take. The faster workflow makes it plausible to test expressions, framing and story rhythm before spending more time on finishing.
The test computer
The machine used its existing RTX 5070 with 12 GB VRAM, 32 GB DDR5 system RAM, and AMD Ryzen 7 8700F processor. It ran Windows 11 Home and ComfyUI 0.39.2 with Python 3.13.12 and PyTorch 2.12.1+cu130. No proposed GPU or RAM upgrade was installed. System RAM and GPU VRAM are separate resources; this test should not be described as a 44 GB GPU configuration.
Why 12 GB VRAM and 32 GB RAM are different limits
VRAM is the GPU’s dedicated working memory. System RAM serves the rest of the computer and can also hold model data while a workflow moves what it needs to the GPU. That movement is called offloading. It lets some workflows run beyond the GPU’s resident memory capacity, but transfers can add waiting time. The two memory pools do not combine into one equally fast 44 GB GPU.
For scale, ComfyUI’s documentation lists the pruned H3 INT8 diffusion checkpoint at about 21 GB, before the other components and working buffers. That file is already larger than this GPU’s 12 GB VRAM. A successful render therefore does not mean the whole pipeline fitted in VRAM at once. This benchmark records completion and elapsed time; it does not measure peak memory use or isolate how much time offloading added. ComfyUI checkpoint guidance
The useful evidence for 32 GB of RAM is concrete: the six short jobs below finished on the existing machine. It does not establish the spare memory available for other apps, longer clips, or a different workflow.
INT8 in plain language
A model stores learned numerical values called weights. Quantization represents numbers with fewer bits, reducing the storage needed for those values. INT8 means eight-bit integers. For the same number of values, eight-bit storage uses half the raw space of a 16-bit format and a quarter of a 32-bit format, before scales, metadata and other components are counted.
Think of it as keeping numbers with less precision so the model is easier to store and move. The tradeoff is numerical approximation, which can affect results. It does not promise a fixed speed increase or identical output. Hugging Face’s quantization explanation
Here, “INT8” describes the main H3 diffusion checkpoint. The pipeline also used an NVFP4/AWQ text encoder, an INT8 video decoder and an FP32 audio decoder. The whole system was not uniformly eight-bit. Nor does INT8 mean eight-bit video color or a particular output resolution.
Turbo is a separate choice. The tested Turbo adapter changes the generation workflow so it can use four sampling steps. INT8 concerns numerical representation; Turbo concerns how the generation proceeds. Every run in the table used the same INT8 base checkpoint, including the runs without Turbo. This test did not compare INT8 against an unquantized model.
Six measured configurations
Every completed output contained 120 frames at 24 fps, lasting five seconds. All six used the same first frame, prompt and seed for a close-up of Nina from HER #1 Driver.
| Setting | Finished five second clip |
|---|---|
| 1080p Base · 10 steps | 28m 09s |
| 1080p Turbo · 4 steps | 11m 53s |
| 720p Base · 20 steps | 13m 04s |
| 720p Turbo · 4 steps | 3m 03s |
| 480p Base · 20 steps | 3m 34s |
| 480p Turbo · 4 steps | 1m 02s |
Download the six measured results as CSV.
These are elapsed times to a finished delivery file, including submission, model loading required by the current cache state, generation, decoding, saving, completion polling and delivery encoding. They exclude setup, downloads, troubleshooting, shot discovery and creative review. There was one completed run per setting.
The 1080p base result used ten steps. The 720p and 480p base results used twenty steps. Turbo also changes the adapter and sampler configuration, so these comparisons measure complete workflows rather than the isolated effect of reducing step count.
For this shot, 480p Turbo offered a short feedback loop; 720p Turbo took roughly three minutes; 1080p Turbo took nearly twelve. Higher resolution was possible on the existing hardware, but it changed how long each creative attempt took. These are observations from individual runs, not promised production averages.
Fast generation and usable footage
The tested shot asked for a silent listening reaction with closed lips. Six sampled frames from each completed clip showed mouth opening despite that instruction. The 1080p Turbo samples also showed additional background people and a larger gaze turn. These are creative problems that a successful file export does not resolve.
All six videos passed technical checks for frame count, dimensions and duration. That does not make them approved episode footage. The sampled review was not full motion playback, a blind quality comparison or a study of the proportion of takes an editor would accept.
For production planning, time per accepted shot matters more than time per generated shot. A fast configuration that needs several attempts may still be useful, but its practical cost includes those attempts, review and finishing.
Watch the six completed takes
These silent, AI-generated clips show Nina from HER #1 Driver. Each is five seconds at 24 fps. They are benchmark outputs, not approved episode footage. All use the same first frame, prompt and seed; higher-resolution runs are different generated takes, not upscaled copies. Use fullscreen to inspect detail. The page may scale a player to your screen, but the delivered file dimensions remain those listed below.
480p Turbo · 4 steps · 854 × 480
Elapsed time to delivery: 1m 02s. Silent AI-generated test clip.
720p Turbo · 4 steps · 1280 × 720
Elapsed time to delivery: 3m 03s. Silent AI-generated test clip.
1080p Turbo · 4 steps · 1920 × 1080
Elapsed time to delivery: 11m 53s. Silent AI-generated test clip.
480p Base · 20 steps · 854 × 480
Elapsed time to delivery: 3m 34s. Silent AI-generated test clip.
720p Base · 20 steps · 1280 × 720
Elapsed time to delivery: 13m 04s. Silent AI-generated test clip.
1080p Base · 10 steps · 1920 × 1080
Elapsed time to delivery: 28m 09s. Silent AI-generated test clip.
Editing at 480p and finishing at 1080p
The delivered 480p image is 854 by 480 pixels. It can be resized to a 1920 by 1080 delivery frame, approximately 2.25 times larger in each dimension. That produces a 1080p file, but file dimensions alone do not establish the detail quality of direct 1080p generation.
Ordinary scaling enlarges the existing frames. AI enhancement estimates additional detail and needs inspection for altered facial features, unstable texture and flicker. DaVinci Resolve Studio provides AI Super Scale as one available enhancement tool. Neither enlargement nor enhancement should be expected to correct an unwanted facial performance, mouth movement or added character.
Upscaling an accepted take allows the editor to retain its action and timing, provided frame rate, duration and frame order are preserved. Rerunning image to video generation at higher resolution is a new take. The same prompt and seed do not promise identical movement across resolutions. A high resolution replacement may therefore require another editing pass.
A practical workflow is to generate quick 480p auditions, build a rough cut, and upscale only the selected takes or selected ranges with enough extra frames for transitions. Inspect those results in motion at the intended viewing size. Regenerate individual shots at 720p or 1080p when the enlarged image or original performance is not acceptable.
The measured 720p Turbo result, three minutes and three seconds, is another candidate for a production compromise. It needs only a 1.5 times enlargement to reach 1080p. Whether that is worth the extra generation time depends on the actual finished images. Upscaling speed and quality have not yet been benchmarked on this computer.
What these results support
This case study supports testing a fast draft and selective finishing workflow on the existing 12 GB GPU and 32 GB RAM computer. It does not establish that all shots take one minute, that 480p enhancement matches direct 1080p detail, or that a particular hardware purchase will deliver a predicted speed increase.
Keeping ComfyUI open between runs did not create a controlled cold versus warm comparison. Cache state, model loading and the order of configurations can affect elapsed time. Repeated tests across several shots are needed before assigning reliable production averages.
The next useful measurement is a finished-shot comparison: the same 480p Turbo take upscaled to 1080p, a 720p Turbo take upscaled to 1080p, and the existing direct 1080p Turbo generation. Measure enhancement time separately, preserve timing, and assess identity, motion, texture and editorial usefulness. This would test the proposed finishing workflow directly.
Get the model from Hugging Face—and check the license first
For readers in the United States: use MiniMax’s H3 license request form to request authorization before locally deploying the model. The published Community License excludes the U.S., EU, UK and South Korea from its standard territory; MiniMax links this application from its official model card. Submitting the form is a request, not an automatic license grant. Read the authorization MiniMax issues for your intended use. Official license · MiniMax licensing Q&A
The model files are available on Hugging Face. Download access and permission to use them are separate matters; a quantized copy or Turbo adapter does not remove the underlying license obligations.
- MiniMaxAI/MiniMax-H3: original model, documentation and official licensing links.
- Comfy-Org/MiniMax-H3: ComfyUI-packaged checkpoints, text encoders and decoders. This is the package family used for the test. Its model card lists the component folders and recommends the INT8 ConvRot diffusion build with a compatible PyTorch cu130 setup.
- lightx2v/Minimax-h3-Turbo: the source of the four-step adapter in this benchmark.
- ModelTC’s Turbo setup guide: adapter installation and inference settings.
Licensing links checked October 24, 2026. Consult the current terms and your authorization when planning deployment or commercial use.
Methods and reproducibility notes
The test ran on October 24, 2026, in the America/New_York time zone. The recorded UTC timestamps fall on October 25. The source shot was E03-RB13-NINA-EXPECTATION, with seed 202610240502. Only the first reference image was supplied. The source records verify the same prompt, reference and seed across the six completed runs.
The main checkpoint was minimax_h3_fl2va_pruned_int8_convrot.safetensors. Supporting components were qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, minimax_h3_video_vae_int8_convrot.safetensors and minimax_h3_audio_vae_fp32.safetensors.
Turbo used minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors at strength 1.0, four steps, Euler sampling, video shift 6 and audio shift 3. This is a 768p adapter tested at multiple resolutions. Direct 1080p generation here is an experimental configuration, not a claim about the model’s trained native resolution or a manufacturer quality guarantee.
Generation produced 124 frames. Delivery trimmed to 120 frames at 24 fps and center-cropped 1920 × 1088 to 1920 × 1080, 1280 × 736 to 1280 × 720, or 864 × 480 to 854 × 480. No upscaling was performed. The published delivery files are silent; generated audio was not evaluated.
There was one completed run per setting, one source shot and a nonrandomized run order. There was no controlled cold-versus-warm comparison, measured power draw, upscale benchmark or hardware-upgrade comparison. The Turbo adapter downloaded during part of the 1080p base test, which may have affected system load. A Turbo submission initially failed validation and was corrected; that troubleshooting is outside its successful-run timing. An earlier canceled 1080p twenty-step attempt used a different shot and is excluded from the completed-results table.
The author’s retained benchmark archive contains per-run timing records, workflows, prompts, raw outputs, delivery files, sampled-frame reviews and checksums. Published clips were matched to those delivery checksums; the downloadable results contain no personal filesystem paths. The observations about mouth movement and background changes come from the original six-frame-per-clip review, not a blind quality study or an acceptance-rate measurement.
The next test should measure the complete path to an accepted shot: generation, retries, review and finishing. These six runs establish that the existing computer can participate in that work—and show where its waiting times begin to shape the creative process.