Skip to content

Publication summary: local ComfyUI image and video qualification

This is derivative publishing copy. The dated finding and linked raw metadata are authoritative.

Canonical facts

  • Image workflow: FLUX.2 Klein 4B FP8; 512×512, four steps, c1; decodable PNG in 9.859 seconds; peak 12,919 MiB from a 943 MiB worker baseline.
  • Video workflow: Wan2.2 TI2V 5B FP16; 512×288, 17 frames, eight steps, 16 fps, c1; decodable 1.0625-second H.264 MP4 in 9.092 seconds; peak 18,263 MiB from a clean 943 MiB worker baseline.
  • Runtime: ComfyUI v0.33.4 at 7a131a3a, CUDA 13.0, PyTorch 2.13.0+cu130, digest-pinned base and exact custom-node revisions.
  • Measurement path: direct loopback managed worker, qualification-only, asynchronous durable jobs; no router route or agent client involved.
  • Decision: functional and capacity evidence passes; both workflows remain unavailable and no-promotion until independent perceptual quality review.
  • Restoration: final managed state is worker absent, zero reservation committed, and 448 MiB host GPU usage.
  • Canonical evidence: https://fakoli.github.io/anvil-serving/findings/2026-08-28-comfyui-media-qualification/

X / short post

RTX 5090 ComfyUI: FLUX.2 Klein 512px PNG 9.86s/12,919 MiB peak; Wan2.2 17-frame H.264 9.09s/18,263 MiB. Functional only; quality/promotion open. https://fakoli.github.io/anvil-serving/findings/2026-08-28-comfyui-media-qualification/

Reddit

Local FLUX.2 Klein and Wan2.2 qualification on an RTX 5090 via ComfyUI

I qualified two pinned ComfyUI workflows locally on one RTX 5090 through a
managed, asynchronous job path.

- FLUX.2 Klein 4B FP8: 512×512 PNG, four steps, 9.859 seconds, 12,919 MiB peak
- Wan2.2 TI2V 5B: 512×288, 17-frame H.264 MP4, eight steps, 9.092 seconds,
  18,263 MiB peak from a clean worker baseline
- Both artifacts decoded with the requested dimensions/frame metadata
- The worker was stopped afterward and GPU use returned to 448 MiB

These are single cold c1 functional/capacity runs, not a quality ranking or a
throughput distribution. The media bytes and prompt text are not published,
independent perceptual review remains open, and neither workflow was promoted
or routed.

Full identities, failures, and sanitized evidence:
https://fakoli.github.io/anvil-serving/findings/2026-08-28-comfyui-media-qualification/

Screenshot alt text

Result card for local ComfyUI qualification on one RTX 5090. FLUX.2 Klein produced a decodable 512-by-512 PNG in 9.859 seconds at 12,919 MiB peak GPU memory. Wan2.2 produced a decodable 17-frame, 512-by-288 H.264 MP4 in 9.092 seconds at 18,263 MiB peak. A caveat says perceptual quality is unreviewed and neither workflow was promoted.

Claim ledger

Public claim Conditions Evidence
FLUX.2 Klein decodable PNG in 9.859 seconds at 12,919 MiB peak one 512×512, four-step, c1 cold qualification image-qualification.json
Wan2.2 decodable 17-frame H.264 in 9.092 seconds at 18,263 MiB peak one clean-baseline 512×288, eight-step, c1 cold qualification video-qualification.json
Final worker absent and GPU baseline restored to 448 MiB managed stop/remove after all runs rollback.json
No quality approval or promotion independent perceptual review not performed; descriptors remain unavailable Finding decision boundary