Methods, perspectives &behind the scenes

MiniMax H3 vs Seedance 2.5: We Fed One Previz to Both, 274 Times

After the heist film, we had 41 previz units and one question: what actually changes when the same previz goes to Seedance 2.5 or open-source MiniMax H3? One listens to the video, one to the text. When the heist film was done, we had 41 Blender previz units, 70% of a Seedance 2.5 resource pack left, and cheap compute. Why not play with open-source MiniMax H3? Up front: nothing in this series is sponsored. We just wanted to spend some money on compute, and it's cheaper than Seedance 2.5 anyway. So we fed the previz through again, this time to three community quantized builds of MiniMax H3: QuantFunc, Kijai and lilcheaty. Every unit got two prompts, one written by MiniMax's paid prompt optimizer, one written by following the official skill. 41 × 3 × 2 = 246 clips. Seedance 2.5 had 28 takes, 32 of the units overlapping with H3. That's 274 clips, and we took every one of them apart frame by frame. Fair warning: this isn't a lab-grade model benchmark. The prompts differ between the two models, so do the input lengths. We're comparing two pipelines we actually make films with. But at this volume, some things are impossible to miss. Sorry, "one-shot video" is clickbait 1. Seedance watches the video. H3 reads the text. If you want to be a director, I honestly recommend spending your time on each model's prompts. The previz cuts at 1.5 seconds. One H3 prompt said 1.1, and H3 cut at 1.1. The other said 1.75, and H3 cut at 1.75. The whole previz was sitting right there as a reference, and it just... didn't look??? Seedance went the other way: 31 of 35 previz cuts, zero frames off. So with Seedance, your edit lives in the previz. With H3, it lives in the text. The details are in the cut-timing article. 2. The harder H3 follows the previz, the more likely it hands the previz right back. We forgot to give one warning lamp a reference image, so in the previz it was just a red ball. All three H3 builds dutifully delivered... a glowing red ball. Seedance drew an industrial warning beacon, with a housing and a pipe along the wall. Same ball, two readings: see the previz-leak article. 3. Quantized builds: for speed, it's QuantFunc. On a 15-second clip, the 8-step QuantFunc build needs only 742 seconds (about 12.4 minutes), about 2.2 times faster than Kijai and lilcheaty, and it keeps every crane, pan and split screen. The price: on 13 of 39 units it handed the previz's walls straight back, and it followed written cuts less often. Kijai has the cleanest surfaces, but now and then a whole clip turns into color noise, and no prompt rewrite brings it back. Open-source folks, start with the quantization article and the speed article. 4. Swapping builds is a costume change. Swapping prompts is a recast. Same prompt on different builds, and the clips still recognize each other. Same build with a different prompt, and the picture moves as far as if you'd changed everything. When a shot breaks, don't blame the weights first. See the prompt article. 5. Some things neither model does. Editors, please keep showing up. A three-second freeze? People keep moving. A dolly zoom? You get an ordinary shot with someone walking in. A speed ramp? The cart just stops. Stop burning generations on these and hand them to the edit. Who gets which shot is in the shot-split article. Pick your path You are Start with Running open-source H3 on your own GPU Quantization, speed, prompt vs build Choosing between Seedance and H3 Cut timing, previz leak, shot split Here to copy a prompt H3 template, Seedance template, porting guide How we ran it (said once, here) Previz: Blender previz from the eight-episode heist film, cut into 41 units of 4–15 seconds, uploaded at 854×480. How we build previz: Stop Rewriting Prompts. MiniMax H3: QuantFunc (W4A4, 8 steps), Kijai (W4A8, 20 steps), lilcheaty (NVFP4, 20 steps), all on an RTX 5090. Two prompts per unit: V1 from MiniMax's paid optimizer, V2 written by an agent following the official skill. Random seed for every job. Seedance 2.5: full previz (up to 22 seconds) with Chinese structured prompts, 28 takes, overlapping H3 on 32 units. Measuring: every frame decoded, comparisons paired on the same unit and prompt. Previz leakage, cuts, motion sync and layout similarity are automatic measurements; the strongest results were checked against side-by-side frames. Camera-technique retention comes from 438 manual judgments. Making video with AI really is cheap. Cheap enough that we fed the same previz to the models 274 times, and the compute cost us under $60. That still doesn't mean AI can replace directors, editors, writers, lighting crews, VFX artists, illustrators and the other crafts of the film and TV industry.