Image

An image request owes more than new pixels. Image records which operation was requested, which sources and controls it may use, what changed, and which exact profile produced the candidate. An image-capable chat model can interpret the picture; this Maker/editor passage produces or transforms it.

Begin with the operation and expected return, then select a simple serving or full workflow route. The study was reviewed on 2026-08-26. These are Designed contracts and candidate bakes, with delivery owned by State. Technical settlement leaves creative adoption, export, and publication with their named owners.

One job, explicit operation

The image-producing surface should accept the whole useful input family without pretending that every admitted model implements every form:

Request and provenance

Describe the intended change before resolving an engine. A candidate ImageJob@2 names its operation, prompt and negative prompt, and the authorized source, reference, mask, and control ArtifactRef values it may use. Dimensions, candidate count, seed policy, immutable preset, deadline, budget, and output policy bound the work. Supplying an optional input does not prove that an engine can use it; the selected profile must declare and prove that capability.

The job also retains the operator-facing prompt, its declared or detected source language when known, the requested model-facing language, and translation policy. A translator may derive a model-facing prompt only through an admitted transformation that preserves both texts, translator revision, and declared loss. Interface locale, model convenience, or prompt enhancement never silently replaces visual direction.

Source-grounded masks, regions, pose, depth, and other controls may come from Sight. Image retains the generative effect; the control's source coordinates, transform lineage, uncertainty, and exact artifact digest remain visible rather than becoming anonymous workflow tensors.

Technical settlement

ImageJob@2 settles whether the requested operation produced a technically valid candidate set. Atelier or another consuming CompositionAppApp then decides whether to adopt it creatively; successful custody and probes cannot make that choice for the owner. Each concrete execution uses Core's Designed ServiceJobAttempt@1 mechanics. An INDETERMINATE attempt remains contained and nonterminal, so it cannot count as either a successful or failed image result.

Qualify the interaction details separately too: numeric progress, previews, cancellation acceptance and settlement, multiple references, hard masks, deterministic seeds, and partial streaming are independent capability facts. A Connector cannot infer them from a broadly compatible endpoint.

Each result receipt retains source and output digests, original and derived prompt language and texts, translator revision and declared loss, engine and container revision, model and component digests, exact preset or graph digest, LoRA and control adapters, scheduler, steps, guidance, seed, dtype and quantization, dimensions, timing, warnings, and cancellation settlement. Provider filenames, image IDs, and embedded PNG metadata are evidence inputs, not canonical provenance.

Serving route and workflow route

The first implementation study keeps two complementary routes rather than forcing a workflow graph through a lowest-common-denominator API.

Use OpenAI Images to carry requests where its dialect fits, while keeping the inner contract with Image. Profile generation and editing separately, including masks, multiple references, partial streams, and cancellation. Because the API offers no universal durable image-job resource, LychD must retain job identity, state, authority, retry law, and artifact custody.

For a Comfy route, select only the immutable preset and parameter openings admitted under the engine contract. Arbitrary graphs, runtime checkpoint downloads, partner API nodes, and ambient custom-node installation are forbidden. The engine graph supplies no SpellweaverWorkflow EngineWorkflow Engine or effect authority.

Diffusers remains the reference implementation and model-compatibility path. It is a Python library rather than a complete multi-user job service, so LychD should not build another general Diffusers server before a supported serving engine proves insufficient. Stable Diffusion is one model family runnable through these engines, not an automation runtime; every chosen checkpoint, encoder, VAE, adapter, and dependency retains its own license and bake.

One runtime adapter and endpoint may later expose separate admitted Image and Video dialect drivers for pinned graphs. Their operations, job profiles, and editorial boundaries remain separate; a generic Connector label cannot merge them.

First permissive model profiles

The initial model bake remains deliberately small:

FLUX.2 klein 4B is the first resident-friendly default. The Qwen family is the first heavy quality and control route. Neither project logo proves every operation: FLUX generic reference editing is not a hard preservation mask, and separate Qwen checkpoints remain separate RuneService Config FileService Config File profiles.

Coven, stasis, and the proving bake

Placement must work on the operator's actual devices. One possible CovenModel GroupModel Group keeps a compact Multimodal Mind on one 24 GB GPU and FLUX.2 klein on the other. A large Qwen-Image profile may instead require affected work to release its leases through Prism's Stasis and readiness passage. Two 24 GB devices do not implicitly provide one 48 GB allocation. The exact runtime must prove sharding, offload, warm-up, cancellation, and complete release; neither Comfy nor a serving engine can choose placement or evict another AnimatorModel RuntimeModel Runtime.

The first corpus should cover natural scenes, illustration, UI and poster text, Slovak diacritics, single- and multi-reference identity, preservation edits, masks and expanded canvases, pose/depth/ edge controls, transparent output, adversarial dimensions and metadata, cancellation, OOM, restart, and deterministic-seed replay. Measure prompt alignment, text accuracy, source and region preservation, identity deviation, control adherence, diversity, latency, peak VRAM, output validity, artifact lineage, and license closure.

Promote exact profiles: the first bake compares FLUX.2 klein through vLLM-Omni with the same basic profile where SGLang-Diffusion supports it, then proves one allowlisted Comfy workflow for the advanced path. Qwen-Image enters as the heavy quality/control route only after its exact quantization or offload topology passes on operator iron.