Glossary

Adversarial distillation Training

Distillation that adds a GAN-style discriminator so the few-step student is pushed to produce sharp, realistic outputs rather than the blurry averages few-step students otherwise tend toward. Used to keep quality high at very low step counts.

• Sharper few-step output than plain distillation
• More complex training; can introduce artifacts if mistuned

Examples: Latent Adversarial Diffusion Distillation (FLUX schnell, SDXL-Turbo)

See also: Distillation , GAN (Generative Adversarial Network)

Artistic style range Capability

The breadth of styles a model can produce (illustration, painting, 3D, anime, graphic design) and how well it follows style instructions. A wide ecosystem of fine-tunes/LoRAs effectively extends this axis.

See also: Fine-tuning

Autoregressive Architecture

Generates an image as a sequence of tokens (like a language model emits text), one chunk at a time. Conceptually unifies image and text generation and can give strong prompt adherence, but sequential decoding can be slow.

• Unifies with LLM tooling; strong instruction following
• Sequential token decoding can be slow at high resolutions

Examples: Parti, token-based image models

Base (foundation) model Training

A model trained from scratch on a large image-text corpus. It defines the ceiling of what downstream fine-tunes and distillations can do. When people compare "the model," they usually mean the base checkpoint.

• Most general and capable starting point
• Largest and most expensive to train and to run at full step count

BYOK (Bring Your Own Key) Platform

An execution model where a tool calls an AI provider's API directly using a key you already hold, instead of routing the request through a third-party server first. The alternative is proxy-based: the tool operates its own server, which calls the provider on your behalf using its own credentials. BYOK removes that middle server entirely — your request goes straight from your machine to the provider you chose, with no markup beyond what the provider itself charges, no added latency from an extra hop, and no scenario where the tool's own infrastructure being down is the reason a request fails. The tradeoff is that you're responsible for holding and managing your own key for each provider you want to reach, rather than a single sign-on.

Also: bring your own key

• No markup beyond the provider's own price, no proxy latency, and no third-party outage surface — the only two parties in the request are you and the provider
• You manage your own key per provider, rather than a single sign-on across all of them
• Doesn't affect the quality of the routing decision itself — only how the resulting request reaches the model

Examples: Modelglass's VS Code router

See also: Cost-aware routing

CFG scale (classifier-free guidance) Concept

A knob controlling how strongly generation is pulled toward the prompt. The model runs both a prompt-conditioned and an unconditioned prediction each step and extrapolates between them; the CFG scale sets how far. Low values drift off-prompt but look natural; high values follow the prompt hard but can oversaturate colours and introduce artifacts. Typical values sit around 5-9 for many diffusion models.

Also: guidance scale, classifier-free guidance, cfg

• Higher scale = stronger prompt adherence
• Too high oversaturates, burns contrast, and can deep-fry the image
• Too low ignores parts of the prompt

See also: Negative prompt , Latent diffusion

Compositional accuracy Capability

Getting multi-object scenes right: correct counts, spatial relationships ("A on top of B"), and binding the right attribute to the right object (the "red cube, blue sphere" problem). Measured by GenEval and T2I-CompBench.

Examples: T2I-CompBench, GenEval counting/position

See also: Prompt adherence

ControlNet Concept

An add-on network that conditions generation on a structural input — an edge map, depth map, pose skeleton, or scribble — so the output follows a specified composition or layout while the prompt controls content and style. The standard way to get precise spatial control out of diffusion models.

• Precise control over pose, layout, and structure
• Requires preparing a control image and adds inference cost

See also: Latent diffusion

Cost-aware routing Platform

Picking which AI model handles a task by comparing real, current cost and capability data — rather than defaulting to the same model for everything, or the most expensive option out of habit. A cost-aware router filters candidate models to ones that clear a minimum capability bar for the task (a benchmark score threshold, for example), then ranks the ones that pass by price, so the cheapest model that's actually good enough gets picked — not the cheapest model overall, and not just the highest-scoring one. Because provider pricing and model releases change on their own schedules, this only works if the underlying cost and benchmark data is kept current rather than a one-time snapshot.

• Avoids paying premium prices for tasks a cheaper model handles just as well
• Only as good as the underlying pricing/benchmark data — stale or invented numbers create false confidence, not savings
• A fixed quality bar can exclude a genuinely capable model that's just measured on a different benchmark than the one used to set the bar

Examples: Modelglass's VS Code router (cost-aware-vscode-router)

See also: BYOK (Bring Your Own Key) , Router ranking eval

DDIM (Denoising Diffusion Implicit Models) Architecture

A deterministic, non-Markovian sampler that produces good images in far fewer steps than DDPM (e.g. 20-50). It is a sampling method layered on a diffusion model, not a different model — most schedulers you see in tools are DDIM-style.

• Much faster sampling than DDPM at similar quality
• Deterministic given a seed, which aids reproducibility

See also: DDPM (Denoising Diffusion Probabilistic Models) , LCM (Latent Consistency Models)

DDPM (Denoising Diffusion Probabilistic Models) Architecture

The original diffusion formulation: a fixed forward process gradually adds Gaussian noise, and the model learns to reverse it one small step at a time. High quality but needs many steps (often 50-1000), so it is slow.

• Stable training, high sample quality
• Slow: many sequential denoising steps

See also: DDIM (Denoising Diffusion Implicit Models) , Latent diffusion

Diffusion transformer (DiT) Architecture

A diffusion/flow model whose backbone is a transformer operating on image latent patches, replacing the convolutional U-Net. Transformers scale more predictably with data and parameters, which is why most frontier image models have moved to DiT-style backbones.

• Predictable scaling, strong text-image alignment at scale
• Compute-hungry; benefits most at large parameter counts

Examples: FLUX.1, Stable Diffusion 3, PixArt

See also: Flow matching / rectified flow , Latent diffusion

Distillation Training

Training a smaller or fewer-step "student" to reproduce the behaviour of a slower "teacher" model. The headline payoff is dramatically cheaper, faster inference (e.g. 1-4 steps instead of 30+), usually with a small quality cost.

• Large inference speed/cost reduction
• Some loss of fine detail, prompt nuance, or diversity vs the teacher

Examples: FLUX.1 [schnell] distilled from FLUX.1 [dev], LCM

See also: Adversarial distillation , LCM (Latent Consistency Models) , Base (foundation) model

DPO (Direct Preference Optimization) Training

A simpler alternative to RLHF that optimises directly on preference pairs without training a separate reward model. Increasingly applied to image models to nudge them toward preferred aesthetics and adherence.

• Lighter-weight than full RLHF pipelines
• Still inherits the biases of the preference data

See also: RLHF / RLAIF

Elo rating Concept

A relative-skill score computed from head-to-head (pairwise) comparisons rather than a fixed pass/fail answer key — the same system used to rank chess players, adapted to rank model outputs. Two outputs are shown blind and a judge (human or automated) picks the stronger one; ratings shift based on the result and the opponent's own rating, so a win against a higher-rated opponent moves the number more than a win against a lower-rated one. Because it's relative, an Elo score only means something against the pool it was computed in. To make the numbers stable, a leaderboard pins a fixed reference point into that pool: a human-expert baseline (GDPval-AA v2 used 1,000) or a named reference model (GDPval-AA v2.1 pins DeepSeek V4.1 Flash at 1,600).

Also: Elo score

• Captures relative quality via pairwise judgment, useful when there's no single correct answer to grade against
• Only meaningful relative to the pool it was computed in — not comparable across different Elo systems
• A model's rating can shift as new models enter the comparison pool, even with no change to the model itself

Examples: GDPval-AA v2.1

Evaluation harness Concept

The specific scaffold and tooling setup used to run a model against a benchmark — what tools it's given, how many retries it gets, and how its output is graded. The harness itself, not just the model, affects the result: a more permissive harness with more retries or better tooling can inflate a weaker model's score, and a stricter grader can undercount a genuinely correct answer. This is why scores are only safely compared within the same harness — the same model can score very differently under two harnesses testing the exact same underlying capability.

Also: harness, scaffold

• Necessary to actually run an agentic benchmark — the model needs tools, retries, and a grading procedure to be evaluated at all
• Makes cross-harness score comparisons unreliable even on the same benchmark and same model
• A stricter or more permissive harness can move a score independently of the model's real capability

Examples: SWE-bench Verified (agentless vs. SWE-agent), Terminal-Bench 2.1 (terminus-2 vs. native)

Fine-tuning Training

Continued training of a base model on a narrower dataset to specialise it (a style, a domain, a refiner stage). LoRA is the common lightweight form. Output characteristics shift toward the fine-tuning data.

• Cheap way to specialise; rich ecosystem (LoRAs, refiners)
• Can narrow diversity or overfit a look if overdone

Examples: SDXL refiner, community LoRAs

See also: Base (foundation) model

Flow matching / rectified flow Architecture

A newer generative formulation that learns a straight-line "velocity field" transporting noise to data, instead of a stepwise denoising chain. The straighter paths mean high quality in fewer steps and cleaner scaling. FLUX and Stable Diffusion 3 use rectified-flow transformers.

• Strong quality with fewer steps; scales well
• Newer, with a smaller tooling/fine-tuning ecosystem than classic diffusion

Examples: FLUX.1, Stable Diffusion 3

See also: Diffusion transformer (DiT) , Latent diffusion

GAN (Generative Adversarial Network) Architecture

A generator and a discriminator trained against each other; the generator produces an image in a single forward pass. Extremely fast at inference, but historically harder to train and weaker at open-ended prompt following than diffusion. Now mostly used for speed-critical or distillation roles.

• Single-pass, very fast generation
• Unstable training; weaker prompt diversity/coverage than diffusion

See also: Adversarial distillation

Hero / marketing asset Quality axis

The image is the focal point at large size — landing pages, ads, key art. Quality means high fidelity, correct details, on-brief composition, and usually legible text. Worth paying for a stronger model and more steps.

See also: Photorealism , Text rendering , Resolution ceiling

Inference Concept

Running a trained model to produce an output, as distinct from training (teaching the model in the first place). When you pay per image at an API provider you are paying for inference compute — the model's weights are already fixed; each request just runs them forward over your prompt.

• The cost you actually pay per generation at an API provider
• Latency and price scale with steps, resolution, and model size

See also: Steps , Sampler

Inference speed Capability

How quickly the model produces an image, driven mainly by step count and backbone size. Directly tied to per-image cost on compute-billed hosts and to user-facing latency.

• Few-step/distilled models are fast and cheap but trade some fidelity

See also: Distillation , LCM (Latent Consistency Models)

Inpainting Concept

Regenerating only a masked region of an existing image while keeping the rest fixed — used to remove objects, fix hands, or edit a detail without redrawing the whole picture. Outpainting is the same idea extended beyond the original canvas to expand the scene.

Also: outpainting

• Targeted edits without regenerating the entire image
• Seams and lighting/style mismatches at the mask boundary if guidance is weak

Latent diffusion Architecture

A diffusion model that denoises in a compressed latent space (via a VAE) rather than at full pixel resolution. Generation starts from noise and is iteratively denoised over many steps. Working in latent space is what made high-resolution generation affordable. Most "Stable Diffusion"-family models are latent diffusion with a U-Net backbone.

• Strong quality and a huge fine-tuning/LoRA ecosystem
• Many denoising steps means slower, more expensive inference than few-step methods
• U-Net backbones scale less cleanly than transformers

Examples: Stable Diffusion 1.5/2/XL, most open-weight image models

See also: DDPM (Denoising Diffusion Probabilistic Models) , DDIM (Denoising Diffusion Implicit Models) , Diffusion transformer (DiT) , Flow matching / rectified flow

LCM (Latent Consistency Models) Architecture

A distillation technique that trains a model to jump most of the way to the final image in a handful of steps (often 1-8) by learning a consistency mapping. Trades a little fidelity for a large speed/cost win.

• Very fast, few-step inference
• Slightly lower peak fidelity and detail than the full multi-step model

Examples: LCM-LoRA variants of SD/SDXL

See also: Distillation , DDIM (Denoising Diffusion Implicit Models)

LoRA (Low-Rank Adaptation) Concept

A lightweight fine-tuning method that freezes the base model and trains small low-rank weight matrices alongside it. The resulting adapter is a few megabytes instead of gigabytes, can be trained on modest hardware, and is loaded on top of the base model at inference to add a style, character, or concept. LoRAs can be stacked and weighted, which is why open-weight ecosystems have thousands of them.

Also: low-rank adaptation

• Cheap to train and tiny to share versus a full fine-tune
• Composable: stack/blend multiple adapters at inference
• Less expressive than a full fine-tune; quality depends on the base model

See also: Distillation , Latent diffusion

MCP (Model Context Protocol) Platform

An open protocol, developed by Anthropic, for connecting an AI assistant to external tools and data sources through a standard interface, rather than a bespoke integration per assistant. A service exposes a set of callable "tools" over MCP; any MCP-compatible client (Claude Code, Claude Desktop, and others) can then call those tools directly inside a conversation, without custom glue code for each one. Modelglass exposes its pricing feed this way — a paid subscriber can ask an MCP-connected assistant a natural-language pricing question and get a live, structured answer instead of manually querying an API.

Also: Model Context Protocol

• One integration works with every MCP-compatible client, instead of building a bespoke connector per assistant
• The client needs to actually support MCP — it isn't a fallback for assistants that don't
• Tools are typically read-only and scoped by the server; an MCP connection can only ever do what the server explicitly exposes

Examples: Modelglass's pricing feed MCP server, Claude Code, Claude Desktop

Model type Platform

The top-level split of the catalogue — image, language, video, audio, and 3D — that the selector on the Models page switches between. Models are grouped by what they produce and how they are bought and compared, not by how similar the underlying technology is. An image model is priced per image or per megapixel; a 3D model is priced per generation, per credit, or per operation (mesh, texture, rig); a language model is priced per million tokens. Those units do not convert into one another, and what makes one model better than another differs by type — prompt adherence and style control for image, geometry and topology and rigging for 3D, reasoning and context length for language. Keeping the types separate means every price and capability comparison on the site is between models that actually do the same job.

Also: category, model category

• Every comparison stays apples-to-apples — you are only ever weighing models that are real substitutes for each other
• A model with priced offerings in more than one type appears under each of them, rather than being forced into a single bucket
• The split follows the buying decision, not the architecture — image, video, and some 3D models are all diffusion-based and still sit in separate types because you do not cross-shop them

Examples: Image, Language, Video, Audio, and 3D on the Models page

Negative prompt Concept

Text describing what you do not want in the image (e.g. "blurry, extra fingers, watermark"). Implemented via classifier-free guidance: instead of guiding away from an empty unconditioned prediction, the model guides away from the negative prompt's prediction, steering the result clear of those concepts. Not all models or APIs expose it.

• A direct lever for suppressing common failure modes and unwanted content
• Overuse can flatten variety or fight the positive prompt

See also: CFG scale (classifier-free guidance)

Photorealism Capability

How convincingly the model renders real-world scenes, lighting, skin, and materials. Distinct from "looks nice" — a model can be highly aesthetic but stylised rather than photoreal.

Print / high-resolution asset Quality axis

The image must hold up at print resolution and physical size. Demands the highest native resolution (or clean upscaling), no artifacts under scrutiny, and accurate detail. The most demanding axis; favours top-tier models.

See also: Resolution ceiling , Photorealism

Prompt adherence Capability

How faithfully the image reflects everything the prompt asked for — objects, attributes, relationships, and intent. The single most important axis for most production use, and what alignment benchmarks (GenEval, DPG-Bench, CLIPScore) try to measure.

Examples: GenEval, DPG-Bench

See also: Compositional accuracy

Rapid iteration / draft Quality axis

Quality good enough to explore ideas, storyboard, or generate many options fast and cheap. The operator cares about throughput and cost per image far more than peak fidelity — a fast, distilled model is usually the right call.

Examples: concepting, A/B option generation, moodboards

See also: Inference speed

Resolution ceiling Capability

The largest / highest-quality native output the model produces before needing upscaling, and the set of aspect ratios it supports well. Matters for hero and print assets.

See also: Print / high-resolution asset

RLHF / RLAIF Training

Post-training that optimises the model against a reward signal derived from human (RLHF) or AI (RLAIF) preferences. Pulls outputs toward what people tend to rate highly — aesthetics, prompt adherence, safety — sometimes at the cost of diversity.

• Better-aligned, more 'preferred' outputs
• Can reduce diversity or over-optimise to the reward model's biases

See also: DPO (Direct Preference Optimization)

Router ranking eval Platform

A regression test that checks a router's model-ranking logic against real production data and a small set of manually-verified reference values, rather than a live LLM judge or synthetic fixtures. The goal is to catch a specific, dangerous failure mode: the ranking logic silently substituting the wrong score for a model that has more than one named-benchmark result, which changes its rank without any error being raised. Deterministic and reference-comparable — a run either matches the expected reference values or it doesn't, with no judgment call involved.

Also: ranking eval

• Deterministic and reference-comparable — no LLM-judge variance, a failure is unambiguous
• Runs against real production data, catching bugs synthetic fixtures wouldn't surface
• Reference values need periodic manual review as the underlying data changes, rather than re-verifying themselves automatically

Examples: Modelglass's routing.eval.test.ts (rankModelsForCategory)

See also: Evaluation harness , Cost-aware routing

Sampler Concept

The algorithm that performs each denoising step — deciding how far to move and in what direction at every iteration. Different samplers (DDIM, DPM++, Euler, UniPC) trade speed against quality and reproducibility. Some providers expose sampler choice; many hosted APIs fix it.

Also: scheduler

• Sampler choice affects quality, convergence speed, and determinism
• Not all hosted APIs let you pick one

See also: Steps , DDIM (Denoising Diffusion Implicit Models) , DDPM (Denoising Diffusion Probabilistic Models)

Seed Concept

The number that initializes the random noise a generation starts from. Fixing the seed (with the same prompt and settings) makes a result reproducible; changing it explores variations. Seeds are how you A/B a single prompt change or recover an image you liked.

• Reproducibility: same seed + same inputs = same image
• Seed behaviour is implementation-specific and rarely portable across models

Steps Concept

The number of denoising iterations a diffusion model runs to turn noise into an image. More steps generally improve quality up to a point of diminishing returns, at the cost of higher latency and — for per-step or per-second pricing — higher cost. Few-step and distilled models (LCM, Turbo, schnell) deliberately cut this to a handful.

• More steps = more detail/coherence, up to a plateau
• Each step adds latency and, on metered pricing, cost

See also: Inference , Sampler , Latent diffusion

Text rendering Capability

The ability to render legible, correctly-spelled text inside the image (signs, logos, labels). A long-standing weak spot for diffusion models; newer transformer-backbone models are markedly better.

Thumbnail / small surface Quality axis

The image is consumed small (feed thumbnails, avatars, icons, list rows). Fine detail and perfect text are largely invisible at display size, so "quality" means clean composition and strong silhouette, not pixel-level fidelity. Cheap models punch above their weight here.

Upscaling Concept

Increasing an image's resolution after generation. Diffusion-based upscalers add plausible detail rather than just interpolating pixels, which is how models that natively target ~1-2 megapixels reach print sizes. Often paired with a light second diffusion pass ("hi-res fix") to sharpen detail.

Also: super-resolution, hi-res fix

• Reaches resolutions beyond a model's native ceiling
• Can hallucinate detail that was not in the original

See also: Resolution ceiling

VAE (Variational Autoencoder) Concept

The encoder/decoder that compresses an image into a small latent representation and reconstructs it back into pixels. Latent diffusion models denoise in this compressed space — far cheaper than working at full resolution — and the VAE decoder turns the final latent into the visible image. A weak or mismatched VAE shows up as washed-out colour, mushy detail, or artifacts even when the diffusion model itself is good.

Also: variational autoencoder, autoencoder

• Makes high-resolution generation affordable by shrinking what the model operates on
• Lossy: fine detail and exact colour can be degraded in the encode/decode round-trip

See also: Latent diffusion

No terms match the selected category.