ComfyUI 0.39: New Media Nodes, Credit Budgets and Offline Controls

ComfyUI released v0.39.0 on October 5, 2026, at 22:49 UTC. For teams combining local generation with hosted media models, it adds image-editing and reference-guided video options, alongside separate controls for disabling paid Partner Nodes and restricting frontend internet access.

Stage an upgrade if a new integration solves a specific production task, or if the documented Qwen Image 2.1 cache crash or asset-catalog problems affect your installation. An unaffected local-only workflow does not need an urgent upgrade just to acquire more paid model choices.

Choose the media task first

Partner Nodes are an existing paid service, not a new billing system introduced by this release. The useful change is which hosted capabilities can now sit inside a Comfy graph. These options are defined in the v0.39.0 source:

The release notes also list ElevenLabs Eleven v4/v4 Turbo and Anthropic Sonnet 5.5 integrations. Inclusion in the release is not proof of access for every account or immediate availability in every Desktop or Cloud installation; those rollout details were not independently checked.

For direct-API context, see Grok Imagine Video 1.5's three media modes. That article concerns the underlying 1.5 API, not this Lite integration.

Choose the right launch flag

Version 0.39.0 separates two behaviors previously bundled together. The merged flag change keeps the existing frontend restriction but gives it a distinct name:

When replacing --disable-api-nodes, choose --offline to preserve its frontend restriction. Switching to --disable-partner-nodes deliberately removes that restriction; it is not just a rename.

For an existing manual installation, append the chosen flag to its usual launch command, retaining other necessary options. For example, disable Partner Nodes while keeping frontend internet access:

python main.py --disable-partner-nodes

Or disable those nodes and restrict frontend requests:

python main.py --offline

These examples follow the tagged argument definitions. The server implementation applies a browser Content Security Policy, not an operating-system firewall. Do not treat it as proof that backend code or custom nodes cannot make network requests. For a genuinely disconnected installation, provision dependencies and models beforehand and enforce network restrictions outside ComfyUI too.

Budget paid executions before batching

Comfy's pricing guidance says Partner Nodes consume credits when executed, not when you merely build or edit a graph. It states 211 credits per US dollar; model and settings affect the charge. Partner Nodes and Comfy Cloud GPU runtime draw from the same balance.

Check the selected model, duration, resolution, connected references and current price display before running. The published pricing table checked on October 6 did not list Grok 1.5 Lite explicitly. Do not substitute the adjacent 1.5 model's rate.

Planning example, not a measured bill: the v0.39.0 Grok Lite price-badge expression uses $0.03 per output second at 720p or $0.14 at 1080p, then multiplies by 1.43. For 100 separately executed, ten-second, text-only clips:

Estimated USD = 100 × 10 × resolution rate × 1.43
Nominal credits = estimated USD × 211

Under those assumptions, 1080p costs about 4.67 times as much. This excludes input images, retries, other paid nodes, Cloud GPU runtime, taxes and later rate changes. It is arithmetic from a tagged estimate, not a live backend quote, quality comparison or spending cap.

Budget the paid executions across every stage, including separately run alternatives, rather than assuming one graph equals one charge. Comfy documents temporary negative balances caused by usage-reporting delays, so the remaining balance is not a guaranteed ceiling. Its account-wide Partner Node concurrency limit is also separate from Cloud workflow concurrency; excess calls can return HTTP 429.

Keep hosted inputs and durable outputs explicit

A local canvas does not keep every input local. Comfy says Partner Nodes send prompts and other inputs to third-party providers. Review the actual reference media and text crossing that boundary before enabling the hosted stage.

Comfy's media-retention policy deletes both input and output media from its managed storage after 24 hours. Save outputs to your own approved storage before expiry, and refresh uploads just before delayed calls. That policy does not establish every provider's prompt, log or training-data retention terms.

For a reference-editing pipeline, a useful design is to save the hosted result before local post-processing. This also matters for FLUX 3 Image: its documented nondeterministic generation means a seed value alone cannot recover the exact image needed by a later edit.

Upgrade, then check the paths you actually use

Keep a recoverable copy of the current environment, workflow JSON and asset database before changing versions. Record custom nodes and launch flags. Follow Comfy's installation-specific update guide: Portable has a stable updater; Desktop's Manage panel updates the engine separately from the app; manual installations need the selected release's dependencies updated inside the Comfy environment, not just a code pull.

After restarting, confirm the loaded engine version and reopen a representative workflow to check for missing nodes, dependency errors or stale frontend/templates. Then use these targeted acceptance checks; they are proposed checks, not results obtained for this article.

1. Launch controls: verify which nodes load and whether the browser response carries the expected Content Security Policy. Check those separately; neither an absent Partner Node catalog nor a CSP header proves whole-host isolation.

2. Paid workflow: use one approved input within a small test budget. Record the exact model/settings and displayed estimate, save the output, and compare the charge with account credit history before expanding the batch.

3. Qwen Image 2.1: if your workflow uses int8 or int4 cache, repeat those configurations and a previously working setting. The specific crash fix supports checking errors and output validity, not promising a speedup.

4. Enabled asset scanning: check uploads and output saves during a representative scan. Verify catalog records and job associations as well as files. The scan-transaction change addresses registration contention, but its author documents longer initial scans and a durability trade-off: recent catalog commits may roll back after power loss or an OS crash. The queue-resumption fix lets scanning resume when a finished prompt leaves the queue empty; a continuous backlog still keeps it paused.

For hybrid media teams, the upgrade decision is concrete: identify the new editing or video capability you need, choose the launch mode that matches the installation, and bound its paid executions before scaling. Local-only teams can make the decision on affected fixes and their normal maintenance schedule.

Methodology: AI-assisted reporting and analysis based on Comfy's release, tagged source and official documentation, rechecked October 6, 2026. No upgrade, generation, billing, quality or network-isolation tests were performed. The cost example is a transparent calculation, not an observed charge.