MiniMax H3 image-to-video, 15s clips on one RTX 5090 32GB: 0.8MP worked, 0.9MP OOM'd
retained text
Made a 15s dancing clip from one reference photo (face is from a person who agreed to it, clip is AI generated). Wanted to see how far one 32GB card goes with MiniMax H3 image-to-video, so I kept raising resolution at 15s until it OOM'd. [the result workflow with 0.8 pixels](https://reddit.com/link/1x1ee9q/video/sve8m7e30euh1/player) Setup i use to training.. https://preview.redd.it/hw7n9bq90euh1.png?width=1906&format=png&auto=webp&s=975fe3b1f919a3ea5e431e5024ccc0c84835445f GPU RTX 5090, 32607 MiB RAM 90 GB (cgroup limit) ComfyUI commit d91ed5f5, native "MiniMax H3: Image to Video" template torch 2.11.0+cu128 unet minimax\_h3\_fl2va\_pruned\_int8\_convrot (\~20 GB) text enc qwen3vl\_32b\_minimax\_h3\_nvfp4\_awq (\~15 GB) vae video int8\_convrot (2.7 GB) + audio fp32 lora lightx2v minimax\_h3\_fl2v\_turbo\_8step\_v1.0, strength 1.0, 8 steps settings 15.0 s, aspect 3:4, same prompt, same seed 757358688076805 every run simple workflow deafult image to video. https://preview.redd.it/l9rtiuji0euh1.png?width=1905&format=png&auto=webp&s=19c662fdbc894ce403ca66055fa11b0200ba70b3 Results (15 s, turbo on, 8 steps) MP s/it total 0.6 51.09 476 s 0.7 65.64 590 s 0.8 83.42 746 s (00:12:26) 0.9 - OOM at step 2/8 Gpu Ram used for 0.6 0.7 0.8 almost the same. https://preview.redd.it/zzodub921euh1.png?width=963&format=png&auto=webp&s=ff4b2c664191f83d486204b829e5c4ed70d75073 Second prompt at 0.8 MP (different choreography): 83.45 s/it, 00:12:24. Prompt content didn't change the time, only res and duration did. nvidia-smi during 0.6 MP sampling: 31758 / 32607 MiB, 100% util, 574W, 78C. VRAM sits near the top whatever you pick, so s/it tells you more. then i try to run change pixels to 0.9 and the result was OOM, even though turbo mode was enabled. OOM at 0.9 MP Allocation on device 0 would exceed allowed memory. Currently allocated : 20.39 GiB Requested : 5.38 GiB Device limit : 31.36 GiB Free (according to CUDA): 391.69 MiB Died inside int8\_linear in the MLP at step 2 of 8. Looks like an activation allocation, dynamic VRAM loading already moves weights to RAM and it still didn't fit. "Prompt executed in 113.53 seconds" is just time until the crash. What changed between runs Same seed, same prompt, only megapixels changed, and the motion changed every time. 0.6 had the best motion and music but looked soft. 0.7 was sharper but motion and music got worse. 0.8 was best overall to my eye so I kept it. Latent size changes the noise, so every resolution is basically a new roll. Audio comes out of the same pass so the music changed too. [the result 0.8 pixels with difrent prompt](https://reddit.com/link/1x1ee9q/video/9n1orl6b1euh1/player) And. last result on 0.6 pixels. [the video result 0,6 pixels](https://reddit.com/link/1x1ee9q/video/8t1sypga2euh1/player) Prompt is written as timed phases (0-4s, 4-8s, 8-12s, 12-15s) with one small repeating move each. Can paste the full one in a comment. Not tested: turbo off at 0.8, --reserve-vram, other seeds, motion reference video. Haven't tried --reserve-vram yet. If someone got 0.9MP at 15s on 32GB, what did you change?
document fetched 2026-10-10 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1x1ee9q/minimax_h3_imagetovideo_15s_clips_on_one_rtx_5090