mirror of
https://github.com/Comfy-Org/comfy-skills.git
synced 2026-11-04 23:37:45 +00:00
[PR #44] docs(skills): fix consistent-characters run_template overrides, scope Step 5 chaining, batch Step 7 seeds #43
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
Comfy-Org/comfy-skills#43
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
📋 Pull Request Information
Original PR: https://github.com/Comfy-Org/comfy-skills/pull/44
Author: @mattmillerai
Created: 10/23/2026
Status: 🔄 Open
Base:
main← Head:matt/be-22709-consistent-characters-overrides📝 Commits (2)
7de6b6ddocs(skills): fix consistent-characters run_template overrides, Step 5 chaining scope, Step 7 batch fan-out258541cdocs(skills): split Step 7 fan-out over the 50-item submit_batch cap📊 Changes
1 file changed (+3 additions, -3 deletions)
View changed files
📝
skills/technique-consistent-characters.md(+3 -3)📄 Description
ELI5
The character-consistency recipe told agents to plug reference photos into a template in a way the server doesn't accept, so those runs either failed or quietly dropped the reference. Now it tells them to put each photo into the template's image loader. It also limits reusing the previous shot to the scene/lighting slot, and has agents state the total number of variations up front and send them in one batch (or a few, when there are more than a batch can hold), so the user approves the whole set up front.
What
skills/technique-consistent-characters.md, three sentence-level edits:run_templateoverrides now target the template's ownLoadImagenodes ({ "<loaderNodeId>": { "image": "<uploaded filename>" } }), found viaio.inputs(search_templatesdetail: "full"orget_template_schema). It also says what to do when the cast needs more references than the template has loaders:get_template, add oneLoadImageper extra reference wired to the next freemodel.images.image_N, thensubmit_workflow.use_previous_outputnow applies only to the scene/lighting slot. The character slots keep the Step 1 reference files.submit_batch. A batch holds at most 50 items, so a larger set is split across severalsubmit_batchcalls with each batch's total stated. A dependent run (the Step 4 split pass) is not batched.Nothing else changed: not the "Follow these steps exactly" framing, the headings, or the Key learnings block. The line
**One reference per auto-grow slot** - slot names come fromget_node, never an arrayis byte-identical, which matters because a downstream consumer keys on that exact sentence. The wording matches the "filename goes in the LoadImage slot" convention in the open multi-scene-comic recipe (#39).Why the old Step 2 sentence was wrong (checked against the MCP server's source on its default branch)
run_templateinput_overridesis{ nodeId: { inputName: value } }, applied after conversion withObject.assign. WithfailOnUnresolvedOverrides: true, a key that doesn't resolve to an executed node is rejected before submit. So a top-levelmodel.images.image_1key is rejected. Under a real node id, it overwrites the IMAGE link tuple with a string.GeminiNanoBanana2V2→modeloptions →imagesisCOMFY_AUTOGROW_V3, and each template slot is["IMAGE", {}](image_1…image_14). These are link slots, not filename widgets.LoadImagenodes listed inio.inputs:api_google_nano_banana2_image_edituses node 16, andtemplate_product_placementuses nodes 1, 2, 3 and 5. Both were confirmed in the server's bundled template data.submit_batch,get_template,get_template_schema,submit_workflow,use_previous_output, andsearch_templatesdetail: "full"all exist as named tools/options on that branch.Testing
claude plugin validate ./claude-code: ✔ Validation passed. This is the repo's only check; it has no test suite or lint config.search_templates/image/fullmentions in README.md and AGENTS.md, and none describes the old behaviour.Provenance
MAX_BATCH_ITEMSon its default branchrun_template/submit_batchrun: not done. These calls spend credits; see Residual.LoadImage`nodes) and an escapedmedias\[\]. I restored normal spacing and kept the file's existingmedias[]. Wording is otherwise verbatim.Residual
LoadImage-widget redirect are backed by reading the source, not by a submitted job.run_templateonapi_google_nano_banana2_image_editortemplate_product_placementand asubmit_batchfan-out both spend credits, so neither was exercised. A follow-up can confirm it with one cheaprun_templatethat setsinput_overrides: { "16": { "image": "<uploaded file>" } }onapi_google_nano_banana2_image_edit.🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.