May 2026: Open models and paid integrations need different tests
May's selected announcements show two execution routes sharing one graph interface: native model integrations and hosted Partner Nodes. Evaluate their inputs, dependencies, rights and outputs separately rather than treating the number of integrations as a measure of market adoption or production quality.
Dated announcements
On 14 May, Comfy announced native support for VOID video object removal, BiRefNet segmentation and Gemma 4. These are distinct processing tasks; their inclusion is not a common quality benchmark. Comfy: VOID, BiRefNet and Gemma 4 (2026-05-14; company announcement).
On 27 May, Comfy announced Krea 2 through Partner Nodes, with controls for style references, creativity and moodboards. This is a hosted integration, not evidence that those particular API models execute locally. Comfy: Krea 2 Partner Node (2026-05-27; company announcement).
A 6 May public interview identifies ComfyUI creator Yannik Marek as ComfyAnonymous and discusses the tension between rapid integration and production stability. fxguide interview with Yannik Marek (ComfyAnonymous) (2026-05-06; first-person founder interview).
Funding and valuation status
No additional financing announcement is established for this month by the reviewed sources. This is not proof that no financing occurred. See the dated funding register; past totals and valuations are not new monthly raises.
What this changes: editorial analysis
Our interpretation is that integration breadth increases the importance of classifying execution boundaries. A single canvas can hide meaningful differences in model ownership, service dependency and cost. The strategic question is whether creators can inspect and preserve those differences while composing a pipeline, not whether one monthly list contains more model names than another.
Start with a specific editorial task. Removing an object from video, extracting a mask and generating an image are not interchangeable outcomes. Define acceptance separately for each stage and inspect the transitions: a plausible single frame does not prove temporal consistency, and a clean mask does not establish that the subsequent generator preserves the intended subject.
For art direction, separate control from repeatability. A style reference or moodboard can help describe an intended appearance, but the stored graph must also retain the exact inputs and model selection. Record which assets can be shared and which remain private. If an upstream API changes, keep the old output as a comparison rather than silently replacing evidence with the latest result.
Proposed verification scaffold, not executed tests
- Choose one task and write a short acceptance rubric before generating. Specify the input rights, output format, defects to reject and the person who will review the result. Use the same rubric for local and hosted candidates.
- Trace each node to local weights or an external endpoint. Record version, dependency source, storage destination and any charge incurred. A local ComfyUI window alone is not proof that the whole job stays on the device.
- Save the final files and the workflow needed to reproduce them. Include failed attempts, manual corrections and post-processing. Compare cost per accepted deliverable only when the acceptance rubric and all cost components are available.
Record the environment, exact versions, permitted inputs, output location, failures and acceptance decision. No pass result is implied by this checklist. Use the workflow evidence guide and measurement method for the corresponding record formats.
Open questions and next review
- Which controls survive export and reopening in another supported environment?
- Which stages depend on mutable hosted models rather than downloadable revisions?
- The official May recap was published on 1 June; distinguish its publication date from the month it describes.
Additional dated reference
- Comfy: May retrospective, published in June (2026-06-01; retrospective company summary)