Publication summary: Hermes image-quality production enablement

Canonical facts

• Local result: Real Hermes generated eight FLUX.2 Klein images through the production Anvil MCP gateway and a managed on-demand ComfyUI worker on one RTX 5090.
• Profiles: draft 512×512, standard 768×768, and high 1024×1024; four steps each; c1; prompt and optional seed are the only generation inputs.
• Quality: six of eight bounded samples over six prompts passed strict independent visual review. Two later draft samples are retained failures for origami/material fidelity and exact bird count. This is not a broad quality comparison or tier ranking.
• Warm gateway E2E: 1.352, 1.242, and 1.650 seconds for one draft, standard, and high request. Complete Hermes turns took 38.354, 24.420, and 41.962 seconds and include agent reasoning/tool overhead.
• Cold path: explicit approval, managed build/start, identical idempotent replay, generation, artifact return, and teardown passed. Its 2,045.626-second E2E includes multiple fix-forward iterations and is not normal latency.
• Final regression: the b46f6ce server returned an exact resume bundle and skill 1.0.4 copied it unchanged, reattached with created=false to the same job, polled to terminal in one English turn, returned a native PNG that matched the authenticated resource byte for byte, and measured 908.936 seconds E2E with 0.087 seconds generation.
• Final client/gateway contract: revision 5ea1edc and skill 1.0.5 preserve an explicit empty profile for profileless workflows, distinguish the five legal replay inputs from two correlation fields, and make malformed- bundle cancellation an explicit fail-closed exception. The deployed router and eight-tool Hermes smoke passed, including exact cold bundle return and explicit cancellation; this does not relabel the 1.0.4 image measurement.
• Safety: incorrect, retired, and cross-boundary credentials returned 401; the router-to-lifecycle and lifecycle-to-resource boundaries use distinct credentials; arbitrary dimensions failed before execution; video remains quality_failed with no fallback; the unrelated Qwen service remained in place.
• Decision: the exact image workflow is available=true at the three fixed profiles and remains promoted=false. Video is unavailable.
• Evidence: canonical finding and machine-readable summary.

X / short post

RTX 5090: Hermes made 8 FLUX.2 images via Anvil MCP at 512/768/1024px; 6 passed strict review and 2 draft prompts exposed fidelity/count limits. Cold resume/teardown passed; video stays blocked. https://fakoli.github.io/anvil-serving/findings/2026-08-28-hermes-image-quality-production/

Reddit

Title: Local Hermes → Anvil MCP → ComfyUI image generation is live on one RTX 5090

Body:

I enabled a bounded production image path from real Hermes through the Anvil MCP gateway to an on-demand ComfyUI worker on one RTX 5090.

• FLUX.2 Klein 4B FP8, ComfyUI v0.33.4, c1
• fixed profiles: 512×512 draft, 768×768 standard, 1024×1024 high
• six of eight images over six prompts passed strict bounded review; two draft failures expose origami/material and exact-count limits
• exact cold approval/build/resume/artifact/teardown path passed
• caller-controlled graphs, models, node paths, and arbitrary dimensions are rejected
• Wan2.2 video remains quality-failed and unavailable; there is no fallback

Warm gateway E2E was 1.242–1.650 seconds for one request per profile, while full Hermes turns were 24.420–41.962 seconds because they include agent reasoning and tool use. The one 2,045.626-second cold run includes every defect found and fixed forward during deployment, so it is incident evidence—not a normal cold-start benchmark. A final server-issued exact-resume regression completed in 908.936 seconds E2E with 0.087 seconds generation.

Full configuration, caveats, latency layers, hashes, and fix-forward record: https://fakoli.github.io/anvil-serving/findings/2026-08-28-hermes-image-quality-production/

Accessible alt text

Result card for a local RTX 5090 image-generation deployment. Real Hermes requested FLUX.2 images through an authenticated Anvil MCP gateway and managed ComfyUI worker. Eight bounded samples covered all three profiles and five cold paths; six passed strict visual review, while two draft samples missed origami/material fidelity and an exact three-bird count. Warm gateway E2E ranged from 1.242 to 1.650 seconds, and the final exact-resume cold regression took 908.936 seconds E2E. Video remains blocked on quality.

Claim ledger

Public claim	Conditions	Evidence
Real Hermes completed all three fixed image profiles	one c1 warm request per profile; one shared prompt	Warm profile results; summary
Six of eight bounded images passed strict review	six prompts; two retained draft failures; independent visual inspection	Warm profile results; final regressions
Warm gateway E2E was 1.242–1.650 seconds	one request per profile; phase telemetry, not caller wall	Result card; summary
Cold approval/build/resume/teardown passed	one cold draft; duration includes fix-forward iterations	Cold lifecycle; summary
Final skill reattached exactly and completed in one English turn	one cold draft; server-issued exact bundle copied unchanged by skill 1.0.4; created=false; same job; 908.936-second E2E	Final regressions; summary
Final generic resume contract is replayable and fail-closed	profileless real-operations cold replay/cancel regression; deployed 5ea1edc router; skill 1.0.5; five submission fields plus two correlation fields	Defects fixed forward; summary
Arbitrary dimensions and video fail closed	one width-override control and one video request; no execution/fallback	Safety controls; summary
Exact image workflow is available; video is not	three fixed c1 image profiles; Wan2.2 exact workflow remains quality-failed	Product contract; evidence boundary