ComfyUI FLUX.2 Klein and Wan2.2 qualification on RTX 5090¶

Date: 2026-08-28

Decision: retain both as unavailable candidates; no-promotion

Evidence: functional, capacity; perceptual quality remains unreviewed

Measured hardware: one NVIDIA GeForce RTX 5090, 32,607 MiB, sm_120

Topology: isolated Windows 11 / Docker Desktop / WSL2 media lane; the router and unrelated inference services were not changed

Result card¶

The exact managed FLUX.2 Klein and Wan2.2 candidate workflows each produced a decodable artifact from a clean 943 MiB ComfyUI worker baseline, then the worker was removed and GPU use returned to 448 MiB.

Setup	Qualified value
Models	FLUX.2 Klein 4B FP8 plus Qwen3 4B encoder and FLUX.2 VAE; Wan2.2 TI2V 5B FP16 plus UMT5 FP8 encoder and Wan2.2 VAE; exact revisions and SHA-256 identities in the evidence
Hardware	one NVIDIA GeForce RTX 5090, 32,607 MiB, sm_120; isolated Windows 11 / Docker Desktop / WSL2 lane
Runtime	ComfyUI v0.33.4 at `7a131a3a`; digest-pinned base; CUDA 13.0; PyTorch 2.13.0+cu130; exact curated custom-node revisions
Managed recipes	`examples/primary-node/Dockerfile.comfyui`, `examples/primary-node/serves.comfyui.toml`, and `configs/media/workflows/registry.json`
Measurement path	direct loopback managed ComfyUI; qualification-only durable jobs; cold model load; c1; no router or agent client
Contracts	image 512×512 / four steps; video 512×288 / 17 frames / eight steps / 16 fps
Evidence	`functional`, `capacity`; exact model/runtime compatibility and format decoding pass
Decision	candidates retained unavailable, `no-promotion`; independent perceptual review remains `human_required`

Headline measurement	Local result	Conditions
FLUX.2 Klein latency	9.859 s	one cold 512×512 PNG, four steps, c1
FLUX.2 Klein peak GPU memory	12,919 MiB	943 MiB worker baseline; nine samples
Wan2.2 latency	9.092 s	one clean-baseline 512×288, 17-frame H.264 MP4, eight steps, c1
Wan2.2 peak GPU memory	18,263 MiB	943 MiB worker baseline; ten samples
Queue behavior	max running 1, pending 0	both jobs returned queued immediately and completed asynchronously
Final restoration	worker absent; 448 MiB GPU used	reservation committed 0 MiB; 28,511 MiB free

Why it matters: This closes the exact model-identity, runtime-compatibility, functional-artifact, capacity, and rollback gates for the first bounded image and video candidates without turning ComfyUI graphs or filesystem paths into a caller API.

Important caveat: These are single cold c1 runs. The transport decoded the files, but no independent reviewer assessed perceptual quality; no claim about visual quality, determinism, concurrency above one, routed behavior, or normal remote availability is supported.

Evidence manifest · Publication summary

Immutable identities¶

The image bundle contains 12,451,817,860 bytes across:

black-forest-labs/FLUX.2-klein-4b-fp8@5b4408e5, SHA-256 97ed34fe...ccb6;
Comfy-Org/flux2-klein@5f526678 Qwen3 4B encoder, SHA-256 6c671498...fc5a; and
the matching FLUX.2 VAE, SHA-256 868fe7b3...e8f3.

The video bundle contains 18,144,966,705 bytes across the three Comfy-Org/Wan_2.2_ComfyUI_Repackaged@c4f60d30 files:

Wan2.2 TI2V 5B FP16, SHA-256 456f9013...a1e;
UMT5 XXL FP8 scaled, SHA-256 c3355d30...4f68; and
Wan2.2 VAE, SHA-256 e40321bd...d156.

The image graph digest is 991b63b8...2e4f; the corrected video graph digest is bd12b2de...9572. Full values and byte counts are retained in the raw metadata and public bundle lock.

Method and measurements¶

Exact assets were staged into named Docker volumes through media bundle stage, with size and SHA-256 verification before atomic placement. The worker was built and started through serves up; compatibility queried bounded feature, per-node, and model inventories. media qualify run submitted an immutable rendered graph, recorded the durable job state sequence, sampled the ComfyUI queue and GPU memory, fetched the bounded artifact into Anvil-owned storage, and independently decoded PNG structure or ffprobe MP4 metadata.

Prompt text is intentionally omitted; only a SHA-256 digest is public. The generated media bytes are also omitted because format and capacity evidence do not require publishing model output, and perceptual review remains open.

The image produced one 258,472-byte PNG with the requested 512×512 dimensions. The final clean-baseline video produced one 111,055-byte H.264 MP4 with 17 frames at 16 fps, 512×288, and 1.0625 seconds duration. The reciprocal single-run rates in raw evidence are not throughput distributions and are not used as headline throughput claims.

Failure record¶

The qualification fixed durable product/runtime defects rather than applying one-off bypasses:

BusyBox staging checksum flags were made portable and exact complete partials became resumable.
Root-owned image builds gained an explicit Git safe-directory check against the ComfyUI release revision; bounded lifecycle build output now survives Windows decoding and redacts credentials.
Compatibility moved from the oversized global /object_info document to bounded per-required-node queries and real ComfyUI feature flags.
The derived image now explicitly contains gcc and libc6-dev, required by Triton's runtime driver compilation under --no-install-recommends.
The pinned VideoHelperSuite graph gained its required immutable save_output=true field and a new canonical digest.

Each failure occurred before a successful artifact claim, is retained in a public ticket, and was followed by the same managed qualification path.

Decision boundary¶

Keep both descriptors discoverable but available=false. The successful qualification-only compatibility override is not a promotion mechanism. A separately administered perceptual reviewer must assess the retained or newly generated artifacts before changing the quality disposition. Routed MCP/A2A acceptance, Hermes skill acceptance, normal remote enablement, and any final live deployment remain separate gates.

The final managed rollback removed the worker, released the entire 28,511 MiB reservation, and returned host GPU use to 448 MiB. No router configuration, route, deployment assignment, or unrelated model service changed.