Image

Prism's Maker/editor office covers image-producing effects. It is distinct from a Multimodal Mind that accepts an image and returns text: ordinary vLLM can serve that Mind, while a diffusion or omni-generation engine must produce the new pixels. Every result is a new derivative artifact, never an observation that can replace its source.

This candidate study was reviewed on 2026-08-26. It records a proposed contract and bake, not delivery, automatic fallback, or permission to generate, edit, publish, or export an image.

One job, explicit operation

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

A candidate ImageJob@1 therefore carries an explicit operation, prompt and negative prompt, authorized source, reference, mask, and control ArtifactRef values, requested dimensions and candidate count, seed policy, immutable preset, deadline, budget, and output policy. Optional inputs do not grant an engine capabilities it has not declared and proved.

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.

ImageJob@1 owns the requested operation, candidate set, technical validation, and result settlement. Creative adoption belongs to Voidlight or another consuming Composition; successful custody and probes never accept the image on that owner's behalf. Each concrete execution uses Core's Designed ServiceJobAttempt@1 mechanics; INDETERMINATE remains contained and nonterminal rather than a successful or failed image result. Numeric progress, previews, cancellation acceptance and settlement, multiple references, hard masks, deterministic seeds, and partial streaming are independent capability facts. A Connector must not synthesize support 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.

OpenAI Images is a useful transport dialect, not the inner contract. Generation and edit support must be profiled separately, along with mask, multi-reference, partial-stream, and cancellation behavior. The API does not provide a universal durable image-job resource; LychD retains job identity, state, authority, retry law, and artifact custody.

A Comfy workflow is likewise an engine dialect, not a Spellweaver Pattern or another workflow jurisdiction. A Rune selects an immutable LychD preset whose compiled graph, node set, model set, and parameter openings passed Assimilation. Arbitrary imported graphs, runtime checkpoint downloads, partner API nodes, and ambient custom-node installation are forbidden. The ComfyUI server routes provide execution primitives; they do not provide LychD authority or provenance.

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 Rune profiles.

Coven, stasis, and the proving bake

One possible Coven 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 active work to settle and release its leases. The requesting Run enters Graph Stasis while Orchestrator drains and transitions the affected Animators before both cards are reassigned. Two 24 GB devices are not implicitly one 48 GB device; the exact runtime must prove sharding, offload, warm-up, cancellation, and complete release. A Comfy graph or serving engine never chooses placement or evicts another Animator.

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 drift, 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.