Same Seed, Different Take: Why Step Count Changes the Performance
Deterministic at fixed seed and step count; altering step count changes actor performance and blocking. Always record both parameters.
OMNI LabsUpdated
Key takeaways
MiniMax H3 is fully deterministic at fixed seed and steps; re-rendering four days later produces byte-identical files (md5 verified).
Changing the seed from 79111 to 51477 with identical prompt yields a completely different performance with PSNR 15.1.
Prompt revisions with unchanged seeds produced zero changed pixels across 124 frames, confirming seeds dictate spatial blocking.
Altering step counts alters choreography: at 4.8 s, an 8-step take has open eyes toward lens while 4-step take has eyes closed.
Turbo-SLA 4-step LoRA at 704x1280 completes in 5m15s versus 31m47s baseline (31m), with 4-step preferred in 3 of 3 blind reviews.
Full determinism: verifying byte-identical reproducibility
MiniMax H3 is fully deterministic at a fixed seed and fixed step count. A re-render four days later came out byte-identical, matched by md5.
In commercial advertising pipelines, the ability to regenerate approved footage deterministically underpins quality control, compliance archiving, and post-production re-assembly. When executing inference with MiniMax H3, holding both the pseudorandom noise seed and total inference step count constant guarantees complete mathematical reproducibility. In our verification benchmarks, re-rendering a complex video sequence four days after initial generation produced a byte-identical bitstream confirmed by cryptographic md5 hashing.
Same prompt, different seed is a different film. Changing only the seed (79111 to 51477) with a byte-identical prompt gave a PSNR of 15.1: different blocking, different ending pose.
Conversely, modifying only the seed parameter transforms the entire dramatic staging. When testing a locked prompt across seeds—transitioning from 79111 to 51477—the engine synthesized a completely distinct performance yielding a Peak Signal-to-Noise Ratio (PSNR) of 15.1. The talent adopted altered physical blocking, altered pacing, and concluded the shot in an entirely different terminal posture.
Prompt inertness versus seed authority in camera moves
Changing only the prompt can change nothing. Reversing the walk direction, adding a framing lock and changing the camera move produced a file with zero changed pixels across 124 frames.
Creative directors frequently assume that refining prompt prose directly directs actor motion and camera movement. However, in models operating without strong classifier-free guidance, prompt text exerts minimal geometric force once noise trajectories are initialized. In a controlled diagnostic run, our engineers drastically revised prompt instructions—inverting the talent's walking direction, enforcing a framing lock, and swapping a pull-out move for a lateral camera track. The resulting video output matched the prior render exactly: across all 124 frames, zero pixels changed value.
Step count changes the take, not just the speed. On the same shot and seed, at 4.8 seconds the 8-step take has the subject drinking with eyes open toward the lens, while the 4-step take has her eyes fully closed.
Step count functions as an artistic staging variable rather than a simple compute budget knob. Because denoising schedulers compute discrete numerical integration steps across latent trajectories, altering the step count alters the latent path. In a benchmark tracking a subject drinking at timestamp 4.8 seconds, the 8-step take rendered the talent with eyes wide open looking directly toward the camera lens, whereas the 4-step take captured her with eyes fully closed in relaxed focus.
Turbo-SLA acceleration and the mandatory step recording rule
Turbo-SLA 4-step LoRA on the same shot and seed at 704x1280 and 124 frames: 5 minutes 15 seconds at 4 steps, 7 minutes 21 seconds at 8 steps and 31 minutes 47 seconds at the 20-step baseline. A Codex vision review preferred the 4-step frames in 3 of 3 pairs.
Integrating accelerated distilled adapters drastically compresses turnaround times while maintaining aesthetic quality. Evaluating a Turbo-SLA LoRA on identical seeds at 704x1280 resolution across 124 frames demonstrated dramatic compute differentials: the 4-step schedule concluded in 5 minutes 15 seconds (5m15s), the 8-step schedule required 7 minutes 21 seconds, and the unaccelerated 20-step baseline demanded 31 minutes 47 seconds (31m). In automated Codex vision review, the evaluator preferred the 4-step frames over baseline outputs across 3 of 3 evaluation pairs.
Denoising step schedules, execution runtimes, and performance variation · measured 21 Oct 2026
Architecture / Model Rail
Denoising Step Schedule
Measured Runtime (124 frames @ 704x1280)
Observed Performance & Denoising Output
MiniMax H3 (Turbo-SLA)
4 steps (fast take)
5 minutes 15 seconds (5m15s)
Subject drinking with eyes fully closed at 4.8 s
MiniMax H3 (Turbo-SLA)
8 steps (intermediate)
7 minutes 21 seconds
Subject drinking with eyes open toward lens at 4.8 s
MiniMax H3 (Standard Baseline)
20 steps (full schedule)
31 minutes 47 seconds (31m)
Slower render; Vision QA preferred 4-step frames in 3 of 3 pairs
LTX 2.5 (Distilled cfg 1.0)
8 steps (seed 4242, 97 frames)
Fast inference baseline
One-line prompt gave woman (two runs of two); seven sentences gave apple
Text-Critical Diffusion Rails
25 steps vs 50 steps / Turbo
Standard sweep schedule
25 steps broke headline text twice on identical seed
Measured on 2026-10-21; seed and step count must both be recorded to reproduce takes.
The rule we now follow: record the step count alongside the seed. Never re-render an approved clip at a different step count, because you will not get the same take.
Production logs must treat step count with the same rigor as random seeds. In studio manifests, every generated take records the precise step count alongside seed numbers. Re-rendering an approved take with an adjusted step count fails to reproduce the approved performance.
LTX 2.5 prompt collapse and multi-seed testing discipline
On LTX 2.5 we compared prompt lengths at a fixed seed of 4242, cfg 1.0, 8 steps, 704x1280 and 97 frames. A one-line red apple produced a woman in two runs out of two; the same apple described in seven sentences produced the apple.
Distilled architectures require comprehensive prompt density to prevent visual collapse. Testing LTX 2.5 at fixed seed 4242, cfg 1.0, 8 steps, 704x1280, and 97 frames revealed that sparse single-line prompts describing a red apple yielded an unrelated portrait of a woman in two runs out of two. Expanding the description into seven sentences detailing surface reflection, texture, and table geometry successfully materialized the apple.
A torch-kernel upgrade made one seed look broken on an infographic. Seeds 1 to 3 rendered it fine, so the seed was bad, not the kernel. We do not judge a kernel or model change on one seed.
Engineers must avoid misattributing noise anomalies to system infrastructure. Following an environment PyTorch kernel upgrade, a benchmark seed generated corrupt infographic layouts. However, evaluating seeds 1 to 3 produced flawless layouts, isolating the defect to an unfavorable seed rather than driver corruption. Pipeline upgrades must never be judged on a single seed.
Typographic step schedules and non-reproducible video-to-video rails
For text-critical frames, the 25-step default broke headline spelling twice on the same seed where the 50-step and turbo settings spelled it right. We sweep seeds on turbo instead of trusting the full schedule.
Typographic accuracy exhibits complex non-linear responses to step intervals. During headline rendering benchmarks, standard 25-step schedules introduced letterform errors twice on a seed that rendered correctly under 50-step and turbo configurations. Studio workflows sweep multiple seeds on turbo schedules rather than trusting default full-step allocations.
Genjutsu video-to-video outputs have no seed parameter, so there is no reproducible re-render. We keep the take we like instead of expecting to regenerate it.
Finally, platform API capabilities dictate asset retention policies. In Higgsfield Genjutsu video-to-video transformations, the API exposes no seed parameter, eliminating mathematical re-render reproducibility. When a satisfactory take is generated, production teams archive the output directly.
