Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- The original maintenance window ended at 22:00 UTC on September 22, but that was a planned date. Kayra and Erato actually recovered at 22:08 UTC on September 23.
- At 21:47 UTC on September 24, the official update said V3/V4/V4.5/V5 were moving back to the intended GPU class and explicitly warned Expect seed changes. At 22:44 UTC, it confirmed migration and maintenance were complete.
- The official status page currently lists image and text generation as Operational. That does not establish pixel-for-pixel restoration for every account, historical task, or old seed output.
- Preserve model, prompt, negative prompt, seed, sampler, steps, dimensions, enhancement/Director Tools parameters, old output, and generation time. Compare representative samples first.
- Returning to the intended GPU class does not imply old images can be reproduced byte for byte. Do not overwrite your only baseline assets before examining differences.
- PRODUCT_FIT=NONE: this is a compute-migration and output-reproducibility issue. Changing IP cannot eliminate seed differences or replace account and model availability checks.
Direct Answer: Maintenance and Migration Are Complete, but Check Your Own Results
NovelAI's official status page records that a compute shortage starting September 18 temporarily moved image models including V5, V3, V4, and V4.5 to another GPU class. Kayra and Erato text models were disabled during maintenance. GLM-4.6 and Xialong were outside this shutdown's scope.
A direct check on September 27, 2026 found Emergency Maintenance marked Maintenance is done, with image- and text-generation components Operational. This is the official service-level final state, not evidence that this site tested a particular account, old session, or historical seed output.
Planned Versus Actual Dates: Text Recovered First, Images Returned Later
The original planned window ran from 2026-09-18 13:00 UTC to 2026-09-22 22:00 UTC, ending at 06:00 Beijing time on September 23. At 22:18 UTC on September 22, the official update said CoreWeave had supplied only initial compute capacity and Phase 2 GPU migration back was postponed to the next day. The planned end must therefore not be treated as the actual recovery time.
At 22:08 UTC on September 23, NovelAI explicitly announced full functionality restored for Kayra and Erato. At 21:47 UTC on September 24, it announced V3/V4/V4.5/V5 were moving back to the intended GPU class and warned that seeds would change. At 22:44 UTC, it confirmed migration completion and the end of maintenance. These are actual status-page updates and must not be backdated to the original planned recovery day.
Why the Same Seed Can Change: A Random Seed Is Not a Cross-Hardware Output File
A seed fixes the random initialization path, not a contract for final pixels. Different GPU classes, numerical precision, kernels, and execution order can produce computational differences. NovelAI's original announcement warned that temporary migration would affect seed reproducibility for V4.5, V4, and V3, and very likely V5. The September 24 return migration for V3/V4/V4.5/V5 again explicitly stated Expect seed changes.
Returning to the intended GPU class therefore does not mean old images automatically match pixel for pixel, nor does it prove all seeds have changed permanently. The official updates give no model-by-model difference magnitude, affected proportion, or per-project reproducibility guarantee. Keep old outputs as baselines and generate comparisons with the complete same configuration to assess your own workflow.
Verify Image Projects: Preserve Metadata Before Replacing Assets
For projects using V5, V4.5, V4, or V3, first preserve the original baseline image without modifying it, along with model, prompt, negative prompt, seed, sampler, steps, guidance, dimensions, enhancement options, Director Tools, reference images, generation time, and project purpose. If original metadata can be exported, retain it together with the file hash.
Choose a representative image with a known seed and generate a comparison in the restored environment using the same configuration. Examine the subject, composition, text, dimensions, and differences that matter to the project. Record the comparison before deciding whether to replace character references, product hero images, storyboard baselines, or auditable deliverables. This site has not performed that pixel-level testing.
Text Workflows: Kayra and Erato Are Restored, but Account Checks Still Matter
NovelAI announced Kayra and Erato restored at 22:08 UTC on September 23. If a project switched models for continuity, export the story, preset, module, and lorebook and retain the temporary version first. Use a non-critical passage to check original-model selection, sampling configuration, style, and context behavior before deciding to move back.
GLM-4.6 and Xialong were not disabled in this incident, but that does not make their writing style or parameters equivalent to Kayra or Erato. If your account still lacks a restored model, save the plan, time, and error and contact official support. Overall Operational status does not prove recovery for each individual account.
Recovery Checks and Stop Conditions: Service Status Does Not Mean Identical Old Output
The status page has announced the final maintenance state. Record the official dates, then run a minimal retest using one preserved image configuration and one non-sensitive Kayra or Erato text sample. Separately verify model selection, request success, preset loading, and generated output. Do not start by rerunning an entire batch.
If requests still fail, save the model, plan, time, original error, and status-page snapshot before contacting support. If image generation works but an old seed produces different output, first confirm that model, parameters, tools, and inputs match the baseline. Then decide whether to accept the difference or pause final asset replacement. Model-specific numerical differences, compensation, and individual account outcomes have not been published and remain UNKNOWN.
Network Boundaries: Do Not Confuse Compute Maintenance With a Localized Routing Failure
This maintenance announcement explicitly concerned a compute shortage, GPU migration, and temporary model shutdowns. Maintenance has now officially ended. If individual requests still show DNS, TLS, 407, or connection-timeout errors, use the proxy connection troubleshooting checklist to inspect the exit, keeping the target domain, time, and error. Do not disable TLS verification.
Changing IP cannot remove post-migration seed differences or replace account eligibility and model availability checks. After connectivity returns, still validate model status and representative samples. Network success cannot establish that historical outputs now match.
Sources
Frequently Asked Questions
Has NovelAI's emergency GPU maintenance ended?
Yes. At 22:44 UTC on September 24, 2026, NovelAI confirmed model migration and maintenance were complete. The original September 22 endpoint was a plan, not the actual final state.
When did Kayra and Erato recover?
The official status page announced full functionality restored for both at 22:08 UTC on September 23, 2026. Individual accounts should still verify actual availability.
Will old seeds definitely reproduce after V3, V4, V4.5, and V5 return to the original GPUs?
No guarantee is given. NovelAI explicitly warned Expect seed changes in its September 24 return-migration update, without publishing model-specific difference ranges.
Is saving the seed alone enough to compare old and new images?
No. Also preserve model version, prompt, negative prompt, sampler, steps, dimensions, tool parameters, reference images, and the old output, then compare representative samples.
Does an all-green status page mean my account and old outputs are restored?
No. The status page provides the official service-level final state. Account availability and historical seed outputs require checks against your own project.
Can a different proxy or a fixed IP repair seed differences?
No. Investigate networking only with evidence such as DNS, TLS, 407, or connection timeouts. Seed differences require preserved configuration and old-output comparisons.