What Midjourney does, described accurately
Midjourney is an independent research lab that operates a subscription text-to-image generation service. Users originally interacted with it through a Discord bot and can now also use a web application, submitting natural-language prompts and receiving generated images across successive model versions and the anime-oriented Niji model line. Its outputs are widely regarded for composition, lighting, color, and stylistic coherence, and the product is a capable, mature generative-art system. Nothing here disputes that.
Like essentially every production text-to-image model, Midjourney is trained by gradient-based optimization over a large image corpus. That is a standard, effective methodology. It carries one architectural property worth naming precisely and neutrally: in ordinary training, every admitted example contributes to every layer's parameters with equal structural authority, and the pipeline keeps no record that later lets an operator say which content was permitted to influence which layers, or how deeply. This is a property of conventional gradient descent, not a Midjourney-specific defect. The comparison below is scoped strictly to that axis.
Midjourney is also a named defendant, alongside Stability AI and DeviantArt, in Andersen v. Stability AI (N.D. Cal., No. 3:23-cv-00201), a putative class action by visual artists alleging their works were used in training image generators. Related publicly reported matters include Getty Images v. Stability AI and The New York Times v. OpenAI. These cases are cited only to establish that courts and content owners are asking evidentiary questions about training. Nothing in this article characterizes any lab's training-data sources or the merits or outcome of any case.
The governance axis: depth-selective training provenance
The invention reconceives the training loop itself as a governed execution environment. As disclosed in Chapter 11 of United States Patent Application 19/647,395, the semantic execution substrate is positioned at the boundary between the forward-pass loss computation and the backward-pass gradient application. It does not alter gradient math or the optimizer; it governs which gradient signals reach which layers, and with what magnitude, based on the semantic properties of the content that produced them. Each training example is treated as a proposed semantic mutation to the model's knowledge state that must be evaluated for admissibility before commitment, so declining to learn from an example is a valid, recorded result rather than an error.
Four primitives follow from that positioning, each grounded in the specification:
Entropy-band-indexed depth profiles. Each example carries an entropy-band classification derived from the semantic divergence of its embedding relative to the model's current representational state (computed by a KL-divergence or Jensen-Shannon metric). Each band maps to a training depth profile: a per-layer or per-block contribution-weight vector where a weight of one passes the full gradient to a layer, zero blocks it, a fraction attenuates it, and a value above one amplifies it. Low-entropy content receives shallow profiles; high-entropy content receives deep profiles. The result is a model whose representations are organized by semantic complexity rather than uniformly aggregated across depth.
Append-only training-provenance log. For each batch or example the substrate records the entropy band, slope position, depth-aggregation profile, the per-layer gradient magnitude that actually reached each block after modulation, the governing policy objects, the content-provenance record (source, acquisition pathway, chain of custody), and the admissibility determination. The log is chronologically ordered and append-only, so entries cannot be silently altered or reordered, and it may be periodically sealed using the cryptographic sealing infrastructure of the cross-referenced Governance filing, producing tamper-evident checkpoints for third-party verification.
Reverse provenance and memorization detection. A forward query traces a piece of content to the layer blocks it was permitted to influence; a reverse query starts from a behavior observed at inference and identifies the bounded set of training content whose depth profiles encompassed the active layer blocks. The specification is explicit that this does not exactly attribute behavior to specific content, because non-linear optimization dynamics preclude exact attribution, but it yields an attribution set substantially narrower than the full corpus. On top of this, a memorization detector classifies similarity between output and a known training artifact as shallow (confined to shallow layers by a suppressed profile), deep (encoded in specific layers, possibly indicating content that should have been depth-restricted), or absent (coincidental).
Differential privacy by confinement, and fine-tuning provenance. Rather than injecting worst-case global noise, privacy-sensitive content is routed primarily to shallow, generic layers and suppressed at deep, memorizable layers; the guarantee is structural rather than statistical, because the model cannot memorize what it was not permitted to encode. Separately, a fine-tuning corpus is recorded as a provenance record structurally distinct from pre-training, so behaviors can later be attributed to the pre-training versus the fine-tuning region of the parameter space.
Where this differs from a conventional art pipeline
A conventional text-to-image pipeline, Midjourney's included, produces an excellent artifact but an opaque process: after training, the relationship between a specific image in the corpus and a specific learned capability is not recoverable, cannot be selectively attenuated, and cannot be re-derived for an audit. The governance layer does not claim to make Midjourney's outputs better. It changes what an operator can answer afterward. When a content owner asks whether their work trained the model, the log answers definitively either way. When a regulator asks whether restricted content was kept out of deep, memorizable layers, the depth-profile records show the contribution weights that were applied. When an output resembles a known artifact, the memorization classifier says whether the influence was shallow, deep, or coincidental, and that assessment feeds inference-time governance. Content whose origin cannot be structurally verified can be confined to shallow layers by policy rather than admitted at full depth by default.
Enablement and embodiment scope
A skilled implementer can build this. The gradient-modulation hook sits between loss computation and optimizer application: the substrate multiplies each example's per-layer gradient contribution by the weight vector of its entropy band's depth profile, writes the resulting magnitudes to the provenance log, and hands the modulated buffer to the optimizer, which operates on it as an ordinary gradient buffer. Because the optimizer is unchanged, the mechanism drops into existing pipelines with minimal disruption.
The approach is intended to be read broadly. Embodiments include per-example and per-batch admissibility; depth profiles expressed per layer or per block; gating that blocks, attenuates, passes, or amplifies gradient flow; entropy bands computed by KL-divergence, Jensen-Shannon divergence, or comparable measures; provenance logs sealed continuously or at checkpoints; anchored content identity derived from structural entropy so provenance survives format conversion; forward and reverse queries; shallow, deep, and absent memorization classes; differential privacy by confinement as an alternative or complement to noise injection; and application to pre-training, fine-tuning, reinforcement-style updates, and real-time interactive updates. Modalities include images, text, audio, and multimodal training; the same substrate governs both what a model may say at inference and what it may learn at training time.
Disclosure Scope
The mechanisms attributed to the invention in this article, depth-selective gradient routing, entropy-band-indexed depth profiles, the append-only and cryptographically sealable training-provenance log, forward and reverse provenance queries, shallow/deep/absent memorization classification, differential privacy by architectural confinement, and governed fine-tuning provenance, are those of United States Patent Application 19/647,395. This article is a dated public description of that disclosure. All references to Midjourney and to any other company, product, dataset, or litigation are external context describing the surrounding market and legal landscape; they are not claims of the filing, and no statement here characterizes any third party's proprietary training data, model internals, or the merits or outcome of any pending case. Midjourney and Niji are used descriptively to identify a real product and are the marks of their respective owner.
