---
title: "Minds ComfyUI results, outputs and troubleshooting | Minds"
canonical_url: "https://getminds.ai/guide/comfyui-results-and-troubleshooting"
last_updated: "2026-10-23T02:30:47.683Z"
meta:
  description: "Understand Minds preference charts, per-Audience rankings, JSON and native image/video outputs in ComfyUI. Troubleshoot draft approval, ties, blocked..."
  "og:description": "Understand Minds preference charts, per-Audience rankings, JSON and native image/video outputs in ComfyUI. Troubleshoot draft approval, ties, blocked..."
  "og:title": "Minds ComfyUI results, outputs and troubleshooting | Minds"
  "twitter:description": "Understand Minds preference charts, per-Audience rankings, JSON and native image/video outputs in ComfyUI. Troubleshoot draft approval, ties, blocked..."
  "twitter:title": "Minds ComfyUI results, outputs and troubleshooting | Minds"
---

Minds

October 16, 2026·Updated October 17, 2026·Minds Team # **Minds ComfyUI results, outputs and troubleshooting** Read the research widget and structured results together, route the original selected media, and diagnose the stage that stopped your graph. This reference covers ties, multiple Audiences, input limits and durable retries._Minds · Audience Preferences_ returns a research response and a set of graph outputs. The response helps you understand what the selected Audiences preferred. The outputs let you preserve that result and continue with the selected original media. A completed preference test does not automatically mean every output port has a native image or video to send. This reference assumes the pack is installed and the server has `MINDS_API_KEY`. Start with the [ComfyUI setup guide](https://getminds.ai/guide/integration-comfyui) if the node is missing or you have not run a first test. Use the [native video guide](https://getminds.ai/guide/comfyui-video-preference-testing) when the graph needs motion and audio.![A real Minds response displayed inside the ComfyUI node, including the QA Audience's image preference distribution](https://getminds.ai/_ipx/_/images/videos/integrations/comfyui/coca-cola-response.png) Read the thumbnail and label together: Ad A is the football campaign image; Ad B is Share a Coke. The completed distribution contains one selection for A and four for B, so the widget displays 20% and 80%. The result-detail capture includes the explanation and final action row. Its companion workflow shows the original B image reaching Save Image. These five QA responses establish that the display and selection agree, rather than validating a market estimate. Campaign sources: [Coca-Cola football](https://www.coca-colacompany.com/media-center/coca-cola-invites-gen-z-to-live-fully-in-the-moment), [Share a Coke](https://www.coca-colacompany.com/media-center/iconic-share-a-coke-is-back-for-a-new-generation.html). Place the pointer over the results and scroll with the mouse wheel or a two-finger trackpad gesture to reach the findings and bottom controls. The graph stays in place. This taller documentation capture shows the entire result at once. ## Read the widget as a research response The embedded widget uses the Minds response surface also used in MCP clients. Depending on the completed response, it can display the question, selected Audience, categorical distribution, findings and available individual Mind answers. Scroll inside the result region to inspect the response rather than reading only the tallest bar. First confirm the question. Context, candidate references and the authored question are assembled into the saved preference instrument. A result for a website introduction cannot automatically answer which creative will win in a silent social feed. Then confirm the Audience and the alternative labels. These define the population and choice set to which the result applies. Read the findings and available answers alongside the distribution. They can help identify issues to investigate, but a plausible explanation is not independent proof of real-world behavior. For broader qualitative exploration, questionnaire work or quantitative methods such as MaxDiff, continue in the connected [Study workflow](https://getminds.ai/guide/studies) and [feature guide](https://getminds.ai/guide/features). Minds is the end-to-end platform for commercial synthetic research, connecting qualitative and quantitative work through Minds, Audiences and Studies. PRISM provides the reasoning, inference and source-modeling foundation beneath each Mind. This adapter exposes one focused categorical preference question on that foundation. ## Understand the five output ports | Port | Type | Use |
| --- | --- | --- | | `selected_images` | IMAGE list | Original native images selected for the routing Audience | | `selected_videos` | VIDEO list | Original native videos selected for the routing Audience, including their audio | | `results_json` | STRING | Serialized result with Study/run references and per-Audience rankings | | `preference_result` | MINDS_CREATIVE_RESULT | Structured result for compatible custom-node processing | | `selected_creatives` | MINDS_CREATIVES | Selected creative records, including native material or hosted/text references | The image and video ports are list outputs. A shortlist, tied result or selected set may contain several items. Ensure the downstream node accepts the native type and list behavior of your ComfyUI workflow. These custom outputs are not generic strings, so connect a node that understands the matching type. The result ports remain useful when no native media exists for a particular modality. For example, a text-concept comparison can produce a valid result and selected concept records without producing images. Likewise, hosted video references do not become native VIDEO objects simply because their MIME type is `video/mp4`. ## What is in results_json The JSON includes `studyId`, `runId`, `reviewUrl`, `question`, `audiences`, `creativeFingerprint` and `interpretation`. It associates the preference with the saved research and exact creative set. Preserve these references when you record a decision, so you can trace which run and alternatives informed it. Each Audience entry contains its `audienceId`, display name, `ranking`, `winners`, optional API `estimation` data and `confidence`. Each ranking item includes `label`, `index`, `share` and `rank`. The share is a fraction of the returned distribution total, rather than a percentage string. An 80% share is represented as `0.8`. Candidate indices are zero-based and refer to the assembled creative order. The `winners` array contains indices with rank 1, including ties. Use labels together with indices and the fingerprint when inspecting or routing a result. Input position by itself is not evidence that a candidate was preferred. The adapter normalizes the returned categorical distribution and sorts it by descending share. It leaves `confidence` as `null`; it does not calculate an unsupported confidence interval, significance score or conversion forecast. If you build your own downstream automation, preserve that distinction rather than relabeling preference as measured effectiveness. ## Compare Audiences without pooling them Each selected Audience keeps its own preference distribution and ranking. If Audience A prefers one execution and Audience B prefers another, the result records both. The node does not merge them into a population with invented weights. When exactly one Audience is selected, its ranking can drive media selection directly. With several, enter an included Audience ID in `selection_audience_id`. This identifies the population whose preference should control the selected creative outputs. It does not change the other Audiences' research results. Without that routing choice, the widget and `results_json` can still show the completed research, while media and selected-creative routing remain blocked. That is expected behavior. Review the segment differences first, then choose which Audience should govern this branch of your production workflow. An Audience ID outside the test is rejected. A friendly name is not a substitute for the ID. See the [Audience guide](https://getminds.ai/guide/audiences) and the optional _Minds · Find Audiences_ helper when you need to confirm the populations accessible to the server account. ## Interpret top_k and ties`top_k` sets the shortlist cutoff within the candidate count. The node includes candidates whose share is at least the share at that cutoff. If several candidates tie there, all of them remain selected, even when the output contains more items than `top_k`. Ranks use the number of candidates with a strictly greater share. Two alternatives tied for first both receive rank 1. With two tied leaders and a lower-scoring third alternative, that third alternative receives rank 3. This preserves the actual tie instead of implying an ordering the data did not supply. If your next step expects one file, make that assumption explicit in your own workflow. You may keep tied outputs for another iteration or review the tie in Minds. The adapter will not silently choose whichever candidate was connected first to make a tie disappear. The node also checks that the creative fingerprint matches the ranked batch before selecting originals. Changing a label or media input is a research change, not merely a cosmetic graph adjustment. Start a new test for the changed material. ## Native outputs preserve the original material Native image inputs are encoded as PNG for the uploaded research copy. Native videos are encoded as MP4, including their audio. The selected output uses the original connected tensor or VIDEO object, not the uploaded encoding or a preview extracted from the response widget.![A complete image workflow retaining the selected original image through Save Image](https://getminds.ai/_ipx/_/images/videos/integrations/comfyui/coca-cola-workflow-transparent.png) This lets your graph continue at its original working quality. The downstream node still determines how its saved file is encoded and named. Check its settings when you need a particular format or audio behavior. The preference adapter cannot restore material that was missing before it received the input. If an image wins a mixed comparison, `selected_images` can be usable while `selected_videos` has no selected native video. If a hosted record wins, inspect the selected reference. Do not interpret an absent native output as a failed research response without checking which candidate actually won. ## Recognize the prepare, review and run stages On the first queue, the node creates or loads a saved Study draft and opens its review dialog. It blocks downstream execution while waiting for approval. Inspect the question, alternatives, Audiences and language, approve the exact draft, then queue again to run.![The actual ComfyUI draft-review dialog before the approved run](https://getminds.ai/_ipx/_/images/videos/integrations/comfyui/draft-review.png) If you cancel review, no winner is available. If you approve but do not queue again, the research has not been triggered by that action alone. The _Review draft_ button reopens the available review. It cannot show a draft that has not yet been prepared. The adapter verifies the saved draft before execution. Changed questions, altered attachments, different Audiences or a new revision can invalidate prior approval. Prepare and review the new state. Do not copy approval values between workflows to avoid the review step. ## Retry the same request or start a new test The node maintains a request identity and a durable server-side receipt scoped to the authenticated account. An unchanged retry can reuse the saved Study, draft and run. This is particularly useful when a network interruption or local polling deadline occurs after the research has already started. For an unchanged retry, queue the same workflow. Keep the same material, labels, question, context, Audiences and study language. Replacing the server credential with another account does not grant access to the first account's saved work; the adapter checks ownership. Use _New test_ when the research inputs change. It creates a new request identity and clears the node's previous approval and response display. Prepare, review and execute the revised test. This also applies when an upstream generation node produced a new image or video, even if its prompt text stayed unchanged. A terminal failed or canceled run is different from a polling timeout. Open the saved Study to inspect it. Repeatedly queuing a terminal failure does not make its incomplete evidence suitable for selecting a winner. ## Input limits and validation | Input | Current adapter contract |
| --- | --- | | Total creative alternatives | 2–20 distinct candidates | | Creative labels | Distinct, non-empty, at most 200 characters each | | Selected Audiences | 1–100 distinct comma-separated UUIDs, subject to account access and readiness | | Assembled question | At most 2,000 characters including context, candidate references and concept text | | Local encoded creative | Non-empty, at most 50 MB per upload | | Text concept record | 1–20,000 characters, also subject to the tighter assembled-question limit | | Hosted URL | HTTPS file reference, at most 4,096 characters, no embedded login details or fragment | | `top_k` | At least 1 and within the actual candidate count | | Local polling timeout | 5–3,600 seconds; default 900 | | `study_locale` | `ar`, `de`, `en`, `es`, `fr`, `ja`, `ko`, `tr`, `zh` | These limits validate the adapter's input, not a guarantee that every file under the limit or every account can execute the same test. Account access, allowances, Audience readiness and research-service validation still apply. A separate manifest text limit does not authorize a question that exceeds the assembled-question limit. ## Troubleshoot the stage that stopped | Symptom | Check and next action |
| --- | --- | | Node is missing | Confirm the official pack is in the running server's `custom_nodes`, dependencies were installed with its Python, and the server was restarted. Inspect custom-node import errors. | | `MINDS_API_KEY` error | Inject the credential into the actual ComfyUI server process and relaunch it. A key in workflow JSON or an unrelated terminal does not configure that server. | | Label or candidate validation error | Count every native image and video in collection order. Match `labels_json` to native inputs, and give manifest records their own unique labels. | | Draft appears and saving is blocked | Review and approve the prepared draft, then queue again. This is the normal first stage. | | Review no longer matches | Start a new test for changed material or load and review the current saved draft as instructed. | | Results display but media are blocked | Check `selection_audience_id` for multi-Audience research and whether the selected records contain the native media type you expect. | | Research is still running | Queue the unchanged workflow to resume polling the same saved run. | | Incomplete evidence or unexpected distribution | Open the saved Study and investigate. The adapter deliberately does not select creatives from incompatible or incomplete results. | | Widget is blank | Check the pack's frontend assets loaded, refresh ComfyUI, and inspect browser errors. The JSON result can help distinguish a display problem from a research failure. | Authentication errors are separate from draft errors. HTTP 401 points to the server key. HTTP 402 or 403 can indicate access or account allowances. HTTP 409 can indicate a changed request or Minds that are not ready. HTTP 413 indicates an asset-size problem. For HTTP 429, retry the same request later rather than creating repeated new Studies. The [official issue tracker](https://github.com/minds-ai-co/ComfyUI-Minds/issues) is the support location for the pack. Provide the pack and ComfyUI versions, the stage that failed, and a sanitized error. Keep the server key, private media, signed URLs and unredacted account data out of public reports. ## What has been verified The published integration's evidence includes real image and native-video workflows through draft approval, live Minds API execution, embedded response rendering and native saving. The video fixtures were two seconds long and included audio. Repeating the same request resumed the existing durable run in the tested workflow. Those checks used ComfyUI 0.41.0 and frontend 1.55.17, with a five-Mind QA Audience. They do not establish long-video acceptance, Comfy Cloud execution, a native Partner Node listing, research calibration or campaign uplift. Keep this distinction when presenting your own creative decisions to colleagues. For research design and interpretation, read [getting started](https://getminds.ai/guide/overview), [Minds](https://getminds.ai/guide/minds), [Audiences](https://getminds.ai/guide/audiences) and [Studies](https://getminds.ai/guide/studies). The [integration overview](https://getminds.ai/guide/integrations), [ChatGPT setup](https://getminds.ai/guide/integration-chatgpt), [Claude setup](https://getminds.ai/guide/integration-claude) and [MCP overview](https://getminds.ai/mcp/overview) explain other host workflows. The [ComfyUI creative article](https://getminds.ai/blog/comfyui-audience-preference-testing) shows how to turn a scoped preference question into a useful next creative iteration. ## **Frequently asked questions**### **What does the preference percentage mean?** It describes the returned distribution among the exact alternatives for one selected Audience. It is directional synthetic research, not measured campaign performance, a representative population estimate or a predicted conversion rate. ### **Does top_k always return exactly that many creatives?** No. The node includes all ties at the cutoff, so a shortlist can contain more candidates than top_k. It preserves tied first place rather than choosing by input order. ### **Why do results appear but selected media remain blocked?** When several Audiences are included, choose one included selection_audience_id to drive media routing. Also confirm that the selected candidate has the native modality you connected; hosted URL and text records do not create native IMAGE or VIDEO outputs. ### **Can I retry without starting another Study run?** Queue the same workflow with the same request identity and unchanged inputs to resume its durable saved request. Use New test and review again when material, labels, question, context, Audiences or study language change. ### **Is the response widget sent the API key?** No. Authentication belongs to the ComfyUI Python server. The result iframe receives the response data for display; it does not need the server's Minds key. Treat the research content itself as information you choose to share. ### **Does the node provide a statistical confidence interval?** The structured result leaves confidence null and does not invent an interval or a significance claim. Interpret the distribution with the Audience, stimulus, question and evidence limitations of the run. [Minds](https://getminds.ai/)© 2026 Minds. Your target audience. AI-driven and grounded in transparent evidence. Build within minutes. [Minds on X (Twitter)](https://x.com/mindsai_co) [Minds on LinkedIn](https://www.linkedin.com/company/mindsaicompany/) [Minds on Instagram](https://www.instagram.com/getminds.ai/)Minds is part of [![ESOMAR](https://getminds.ai/images/newsroom/logos/esomar-logo.svg)ESOMAR](https://esomar.org/) [![GreenBook](https://getminds.ai/images/newsroom/logos/greenbook.svg)GreenBook Directory](https://greenbook.org/company/Minds) [![Insight Platforms](https://getminds.ai/images/newsroom/logos/insight-platforms.png)Insight Platforms](https://www.insightplatforms.com/platforms/minds/) [![Capterra](https://getminds.ai/images/newsroom/logos/capterra.svg)Capterra](https://www.capterra.com/p/10046203/Minds/) [![G2](https://getminds.ai/images/newsroom/logos/g2.svg)G2](https://www.g2.com/products/minds/reviews) [![CSSDA Best UX Design Award](https://getminds.ai/images/newsroom/logos/cssda-best-ux-award.png)CSSDA Best UX Design Award](https://www.cssdesignawards.com/) [![CSSDA Best Innovation Award](https://getminds.ai/images/newsroom/logos/cssda-best-innovation-award.png)CSSDA Best Innovation Award](https://www.cssdesignawards.com/) [![CSSDA Best UI Design Award](https://getminds.ai/images/newsroom/logos/cssda-best-ui-award.png)CSSDA Best UI Design Award](https://www.cssdesignawards.com/)