Validation — 0.2.0 preview

Updated on 2026-09-29. This release has working ComfyUI registration, CPU contract tests and real local API checks. GPU image generation is still pending. The 0.2.0 interface checks below were made on 2026-09-29. Earlier Gallery and sampling observations retain their original scope.

0.2.0 interface and Manual defaults

New and starter Engines use Manual with enhancement none. CPU schema and engine regressions confirm the defaults and the actionable empty-Enhanced error. Existing selected legacy presets still resolve.
In the real Canvas frontend with the older DMAI Suite also installed, Manual displayed no preset selector, upload or export. Switching to Enhanced exposed Upload JSON. A local QA JSON imported successfully through the actual file chooser and validation route, showing 28 steps. Returning to Manual restored the edited 24-step draft.
DPM++ 2M and Karras selections, the imported settings and the manual draft survived workflow save/reload. A native Prompter-to-Engine drag recreated the link; the saved graph contained the new link ID with unchanged port types.
Inline LoRA CLIP strength changed from 0.4 to 0.3, independently of Model strength 0.65. Moving that LoRA to position 2 retained both values. Regression tests also cover invalid CLIP drafts blocking reorder/queue serialization.
Canvas captions identify native sockets at 14 px. Connector tests cover hit-target clearance, minimum dimensions after loading/resizing, caption geometry, connected metadata and collapsed nodes.
The rendered interface was inspected at 70% canvas zoom. Prompter Generate and LoRA Add remain visible at the new default sizes. Gallery uses a fixed selection footer with an independently scrolling image grid; its panel has no outer overflow in the measured fixture.
The compact Enhanced card measured 616 px client height and 616 px scroll height, with Upload and Seed visible. Returning to Manual grew it to 716 px without overflow, restored 24 steps and removed the upload control.
Preset mode and adaptive-height tests preserve larger saved/user dimensions, including Comfy's extra native widget padding. Package-specific LoRA classes avoid the matching global selectors in the older Suite and standalone LoRA extension.
Versioned frontend imports and stylesheet URLs refresh the changed assets. Existing frontend initialization/deprecation warnings from the reference host remain; this release does not claim a warning-free third-party installation.

The images in the QA Gallery are synthetic solid-color fixtures. These checks do not add GPU inference or image-quality coverage.

0.1.3 coexistence and migration

Real ComfyUI loader tests load the new extension alongside legacy V1 and Suite-style V3 Prompter fixtures in both orders. Each ID retains its own schema and execution contract. The new extension does not claim DMAIPrompter.
A running CPU ComfyUI instance loaded the actual older ComfyUI-DMAI-Suite and the new package together. /object_info contained all seven new nodes and all four Suite nodes. DMAINodesPrompter returned DMAI_PROMPT; legacy DMAIPrompter retained CONDITIONING, INT, INT, STRING.
In the real Canvas frontend, both Prompters appeared in their separate package categories and rendered their own controls. A saved 0.1.2 DMAI NODES workflow upgraded on load with its five links, 1152x768 dimensions, image count 2, model choices and LoRA strengths intact. A legacy Prompter was added, edited, saved and reloaded alongside it; its original ID and widget format were preserved.
JavaScript regressions cover nested serialized subgraphs, named/positional widget formats, invalid or ambiguous signatures, unchanged settings/links and canonical progress tracking. The new frontend mounts only DMAINodesPrompter.
Eleven Python migration tests cover mixed UI/API graphs, conservative signature checks, invalid links, UTF-8/configuration boundaries, source/output preservation and rejection of legacy nodes before API submission.
The new LoRA row/name classes are isolated from the matching selectors found in older DMAI frontend styles.

These checks establish registration, UI coexistence and workflow migration. They do not add GPU inference coverage or certify unrelated third-party extension combinations.

Reference environment

Windows, Python 3.12.11, CPU PyTorch 2.14.0.
ComfyUI v0.37.0, commit 73c9bad4d21e7addbe1d13bc92eee0f1431b017d.
The tag's frontend package: 1.52.7.
A separate loopback-only installation, tiny explicitly named QA inventory fixtures, no usable image models and isolated output storage.
No paid compute, Runpod start or model-weight download.

Passed

Check	Evidence
Python suite	102 tests passed with optional real-Comfy CPU tests enabled, including registration coexistence, workflow migration, scoped native progress registry updates and multi-image progress.
JavaScript suite	80 tests passed: dimensions, exact seeds, ordered manifests, strength edits, mode persistence, sampler overrides, ZIP selection, keyboard containment, Comfy change events, progress presentation, job lifecycle handling, native socket geometry, model-file selection and safe workflow migration.
Node registration	All seven native V3 nodes appear in real `/object_info`.
Runtime choices	Four model profiles, 45 native sampler names and 9 schedulers in the test installation. Extension installations can change these counts.
LoRA loading	Four real CPU comparisons use tiny safetensors, Comfy's file loader, ModelPatcher and LoRA weight calculation. Results match the standard loader and expected low-rank math exactly.
Model boundaries	Real CPU latent conversions for Wan, FLUX and SDXL; native BasicScheduler/V3 outputs; sampler registry and Krea family checks.
Enhancer dependency	Both pinned upstream implementations, 1.1.0 and 1.4.2, pass source/signature checks and register the expected patch on a small test object. No model inference.
Gallery	Real Comfy queue jobs save synthetic 64×64 PNGs. Two selected ZIP entries pass CRC and match original download bytes exactly. Unit tests cover history snapshots, cross-gallery rejection and missing originals.
UI transport	Native Comfy merges `dmai_gallery` correctly; no second native image-preview payload is emitted.
Progress transport	A real `/prompt` job and WebSocket emitted `preparing` then `error` with the correct prompt and Engine IDs when the optional enhancer was absent. No completion event was fabricated.
Progress lifecycle	Automated tests cover cached jobs, subgraph identity, multiple Engine branches, queued-job removal, failure, cancellation, disconnect and cleanup. Only native workflow success reaches 100%.
Progress browser check	Generate appears only in Prompter. A real Comfy execution failure re-enabled it and showed the actionable error. The production component harness displayed a 42% perimeter, disabled Generate while active, and restored it at 100% completion using explicitly labeled fixture states.
Native connections	Reproduced the old failure: dragging the visible edge socket panned the canvas. After repair, mouse drags created Prompter→Engine, LoRA→Engine, Engine IMAGE→Gallery and Engine REPORT→Gallery links. Dragging from the advanced SAMPLER input opened Comfy's compatible-node search and connected native KSamplerSelect. All five links survived save and reload.
Model picker	Real Comfy inventories supplied one checkpoint, two diffusion files (including a nested path), one encoder and one VAE fixture. The top selector exposed both groups. Switching checkpoint/diffusion source, explicitly choosing Qwen architecture, and saving/reloading preserved the exact file paths and supporting selections. Removing/reselecting a VAE updated the missing-file message immediately without closing Model files. No fixture is a usable generation model.
Browser	Production panels render in real Comfy Canvas. Custom dimensions, image count, Manual/Enhanced round trips and native sampler/scheduler menus were exercised. Saving and reloading retained 1152×768, count 2, 22 steps, DPM++ 2M and Karras. Reordering retained LoRA model strength 0.65 and independent CLIP strength 0.4. Comfy's unsaved marker and Undo respond to edits. Two non-adjacent Gallery images could be selected.

The visual Gallery fixtures are solid-color test images, not generated model results.

Not yet verified

Full Krea, SDXL, Qwen or FLUX image inference and visual quality on GPU.
Original enhancer 1.1.0 versus 1.4.2 image parity.
Installed RES4LYF sampling and its stochastic or custom scheduling behavior. Integration is checked against upstream interfaces and guarded at the Engine boundary.
Runpod templates, Linux GPU installations and memory requirements for each model/precision.
Complete Nodes 2.0 renderer behavior. Native sockets remain available through the renderer's own layout; Canvas is the visual reference.
Browser download completion: the in-app test browser ran the ZIP button request successfully but did not expose a completed download event. ZIP contents and original files were verified directly over HTTP.
Every third-party model variant, LoRA or custom-node combination.

Do not interpret a registered patch or a passing tensor fixture as a model-quality guarantee. Start with one image and the recommended model files when validating a new machine.

Reproduce

Run from a Git checkout of DMAI NODES; the install ZIP does not include tests. Lightweight Python tests need only requirements.txt; real-Comfy checks skip by default. JavaScript tests require Node.js 22 or newer:

python -m pip install -r requirements.txt
python -m unittest discover -s tests -v
node --test tests-js/*.test.mjs
python tools/build_release.py

For real CPU checks, use an environment with the reference ComfyUI and CPU torch available. Set DMAI_RUN_COMFY_TESTS=1 and DMAI_COMFY_SOURCE to your ComfyUI source directory before running unittest. No weights are required. These checks use explicit CPU configuration.

The release builder includes allowlisted package files and verifies ZIP CRC. The SHA-256 file identifies the exact download. CI is configured to check the lightweight suite on Windows and Ubuntu; optional real-Comfy tests are local evidence, not part of the lightweight CI job.