The “random” seed is not random in the way the name suggests.

The seed is an integer that selects the specific pattern of random noise the model starts from. Same seed, same prompt, same settings produces the same final image, bit for bit. A different seed means different starting static, and so a different output even with an identical prompt.

Practitioners use seeds two ways: as a reproducibility tool (regenerate exactly this image later) and, on SDXL, as a hidden compositional control (“seed 47 gives a tight portrait; seed 109 gives a wider shot”). The randomness lives at the step where the seed is chosen, not in what the seed then does.

Because the number has been hiding a control surface that was never labeled as one. Every tutorial since 2022 described the seed as “a number to fix for reproducibility — random otherwise.” We measured that on SDXL the seed is doing about half the work of placing the subject in the frame (NoobAI centroid_y seed fraction 54.4%, 95% CI [42.8–66.0]) — comparable in influence to the prompt itself.

That makes the seed a covert compositional knob practitioners have been intuitively exploiting (“seed banks”, “lucky seeds”) for years without anyone formally measuring it. On Flux the same knob has moved — from the seed to the prompt — which has real consequences for any workflow that depended on it.

Not yet. It is currently a preprint — a public manuscript released before formal peer review. The predictions are pre-registered (outcome probabilities committed before the data lands), and the headline inversion is defensible at workshop scope today: p ≤ 0.0002 on the permutation test, replicated across two independent feature extractors, with non-overlapping 95% confidence intervals on the two architectures.

The plan is a workshop submission soon, then a main-track submission after the disentanglement experiments complete. It has not yet been submitted. The preprint format lets the result circulate and get critiqued in the open; everything reported is backed by public data, so anyone can reproduce or contest the numbers.

Yes, on SDXL-family models. Fix the seed to lock composition — seed banks are exactly the right practice on SDXL, Illustrious, NoobAI, Animagine, and Pony, which are all the same U-Net architecture. Our measurement is the explanation for why that workflow works: the seed carries roughly half the compositional variance.

On Flux the seed will not do that. If you move to a Flux pipeline, the seed bank stops carrying composition and you switch to a canonical prompt prefix (encode the framing in the first tokens), IPAdapter at higher weight for character lock, and image-to-video for animation. The finding is descriptive, not prescriptive — if your current pipeline ships images that look right, it validates that choice. It becomes relevant when you migrate between model families.

MMDiT (multimodal diffusion transformer) is the architecture used by Flux and SD3/SD3.5, where text tokens and image tokens mix at every transformer block. A U-Net (SD1.5, SDXL and its fine-tunes) instead injects text at specific resolutions via cross-attention. That structural difference changes how compositional information is routed.

In our measurement, the locus of compositional control moves: on the U-Net it lives largely on the seed (the noise fiber); on MMDiT it lives on the prompt path. We hypothesize MMDiT's text-everywhere attention is what routes composition through the prompt, but we have not yet measured attention attribution — so we say “the relationship inverts between these two models,” not “MMDiT causes the inversion.” The wording matters and we hold it until SD3.5 lands.

Indirectly. Many video models inherit a text-to-image backbone — an MMDiT video model (Wan 2.2, Hunyuan Image-Video) extends the MMDiT family, and older HunyuanVideo-style models extend the U-Net family. The architectural prediction carries: a video model built on an MMDiT backbone should sit in the prompt-dominant regime, one built on a U-Net backbone closer to the seed-dominant regime.

We have not directly measured video variance decomposition. In practice the recommended pattern is to set composition once in a still keyframe — where the seed gives you that control on SDXL — then animate forward with an image-to-video model that takes the composition as given.

Li et al. documented initial-seed effects on object placement within a single architecture family (Stable Diffusion and PixArt-α). It establishes that the seed influences where things go, but does not decompose that effect into a variance fraction or compare U-Net against MMDiT.

Our contribution is the cross-architecture contrast: we measure the variance fraction explicitly and show that the locus of compositional control moves — from the seed (U-Net, ~50%) to the prompt (MMDiT, ~5%). Single-architecture findings show the seed matters in SDXL; ours is the first to show that whether it matters is itself an architectural property.

Yes. The headline inversion on centroid_y is backed by a permutation test with B = 5000 shuffles of the model label. Observed Δ = 0.472 (NoobAI 0.498 − Flux 0.026); the 99th percentile of the null was 0.040, and the observed Δ was exceeded by 1 of 5000 shuffles — p ≤ 0.0002.

Bootstrap CIs alone are necessary but not sufficient for a difference claim. The permutation test is the formal test, it is distributional-assumption-free, and it survives Bonferroni correction for ~50 comparable tests.

No, because we replicated it with an independent feature extractor. DINOv2-large semantic features from the same NoobAI sweep give the top principal components as 75–91% prompt-driven, while centroid_y is 54% seed-driven. That dissociation — “where the subject is” is seed-driven, “what the subject is” is prompt-driven — comes from two extractors trained on different objectives.

On Flux the same DINOv2 measure gives 1.4% seed contribution, agreeing with the 5% centroid number. If centroid were an Otsu artifact, DINOv2 would not have to agree — and it does.

Yes — and we say so explicitly. NoobAI and Flux differ on six axes (architecture, training objective, parameter count, training data, text encoder, distillation). Until the disentanglement cells land we report “the relationship inverts between these two models,” not “the architecture causes it.”

The critical control is SD3.5-Large (MMDiT like Flux, but DDPM-style training like SDXL), which separates architecture from training objective. It is pre-registered with three outcome probabilities: architecture-wins 0.55, intermediate 0.30, training-wins 0.15. An intermediate result is one we committed to interpreting in advance.

Yes. All raw artifacts are already public at github.com/quivent/lambda — image grids, feature tensors, h-space activations, manifests, variance decompositions, and bootstrap CIs, at the specific commit hashes that produced each reported number.

The lambda topology CLI provides one-command reproduction; total wall time from scratch on a single 96 GB GH200 is about three hours. The companion documents are public at github.com/quivent/anime.productions and updated as data lands. There is no gatekept release.

Five experiments are queued in priority order:

1. SD3.5-Large — MMDiT architecture with DDPM training, the disentanglement cell separating architecture from training objective.
2. PixArt-Sigma — a DiT that is not MMDiT and not distilled; the cleanest control for “is it the transformer, or MMDiT specifically?” (cell currently running).
3. Larger NoobAI scale — 256 × 32 = 8192 images, tightening every reported CI by about 3.5×.
4. DINOv2 per-patch features — richer spatial measurement than the centroid we currently use.
5. Sampler ablation (deterministic vs ancestral), plus the missing U-Net rectified-flow cell — the absent corner of the 2×2 architecture-by-training grid — and the mechanism work on attention attribution.