[PR #36] [MERGED] docs(generate-video): rank the H3 speed-up options in the MiniMax H3 notes #34

New issue

Closed

opened 2026-10-26 22:34:13 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:34:13 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-skills/pull/36
Author: @mattmillerai
Created: 10/5/2026
Status: ✅ Merged
Merged: 10/22/2026
Merged by: @mattmillerai

Base: main ← Head: matt/be-18911-h3-speed-up

📝 Commits (2)

727f2a8 docs(generate-video): rank the H3 speed-up options in the MiniMax H3 notes
1ca6b6b docs(generate-video): tie the H3 speed-up values to the fetched template graph

📊 Changes

2 files changed (+2 additions, -0 deletions)

View changed files

📝 claude-code/commands/generate-video.md (+1 -0)
📝 skills/comfy-generate-video.md (+1 -0)

📄 Description

ELI-5

When someone asks an agent to make their MiniMax H3 video generation faster, the skill gave it no default, so it just repeated whatever search_templates returned. This adds one bullet to the H3 model notes. It makes the default a single switch in the shipped template (the Lightning LoRA), ranks the next options in order, and says the paid Max Turbo templates are a route change to offer, not something to pick quietly.

What changed

One bullet appended to ## Model notes: MiniMax H3 (as of 2026-08), with the values marked as 2026-08 values to confirm against the fetched template graph, in both skills/comfy-generate-video.md and claude-code/commands/generate-video.md. The added line is byte-identical in both files. The heading and its "as of" date are left alone.
Nothing else is touched. Step 0 and the area #32 edits are untouched, so this should rebase cleanly whichever merges first.

Judgment calls

One clause reworded to pass the MCP negative-claim lint. The suggested wording said the Max Turbo templates go "via run_template, not a partner_generate slug". comfy-cloud-mcp-server's guide-negative-claims lint flagged that sentence as a non-support claim with no inventory row (2 failures). I changed it to "launch them with run_template", which says the same thing affirmatively. The skill copy's existing Hailuo bullet already says minimax/hailuo-03-t2v is the only registered MiniMax slug.
The wording is neutral about what's "above" it. The claude-code/ copy's H3 notes are an older version that never mentions run_template for the other paid H3 templates. So the bullet doesn't point back to that text. Its one back-reference, "the OSS-vs-partner rule above", is the Step 0 both-routes rule, which exists in both files.

Verification

Cited names checked against comfy-cloud-mcp-server origin/main data/templates.json and the bundled catalog. All found: video_minimax_h3_t2v / _i2v / _r2v, video_fastvideo_fasth3_t2v / _i2v (and no fasth3 R2V entry, which matches "no R2V variant"), api_minimax_h3_max_turbo_t2v / _i2v, and MiniMaxH3SigmaShift. The Max Turbo T2V catalog description itself says "768p clips in roughly 1–3 seconds".
MCP guards: in a temporary worktree of comfy-cloud-mcp-server at origin/main, I replaced the bundled data/skills.json comfy-generate-video body with the new file (the bundled body matched current main byte-for-byte, so this is the same as a refresh) and ran the guard tests.

Residual

Template internals checked against the live graphs (review round): in Comfy-Org/workflow_templates main, Boolean (Enable Lightning LoRA) is off by default in video_minimax_h3_t2v / _i2v / _r2v, and the step switch picks 20 or 6 (R2V: 20 or 4). FastH3 t2v/i2v run 8 steps with MiniMaxH3SigmaShift 10/3, and there is no FastH3 R2V. The H3 graphs have no shift node, so the earlier "12/3" default-shift figure was dropped. The bullet now tells the agent to read these values from the fetched graph before acting on them. The 768p 4-step LoRA is listed in the template notes, but its 6/3 shift was not checked against a graph.
The two copies of the H3 section already differ (before this PR) in the routing/prefix bullet, the evidence-precedence bullet and the Hailuo bullet. The claude-code/ copy is the older wording. So the H3 sections as a whole don't match, though the added bullet does in both files. Syncing the claude-code/ copy is a separate change and wasn't done here.
Not run: npm run refresh:skills. It fetches from GitHub, so I injected the body locally instead. The real nightly refresh will run the same guards after merge.

Provenance

Authored by: agent-work loop
Verified:
- The comfy-cloud-mcp-server guide guards passed with the new bullet injected into data/skills.json (273 passed, same as baseline). They ran on the original wording; the review-round edit adds no negative claims and was not re-run through them.
- The template names, the Lightning switch default, the 20→6 / R2V 20→4 steps and the FastH3 8-step 10/3 shift were checked against the live workflow_templates graphs. The Max Turbo 1–3 s figure is in the official catalog description.
- The added bullet is byte-identical in both skill files.
Deviations:
- One clause was reworded from the suggested text to pass the negative-claim lint.
- The unverifiable "12/3" default-shift figure was dropped after review.
- The H3 sections as a whole already differed between the two files before this PR. Only the added bullet matches.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/comfy-skills/pull/36 **Author:** [@mattmillerai](https://github.com/mattmillerai) **Created:** 10/5/2026 **Status:** ✅ Merged **Merged:** 10/22/2026 **Merged by:** [@mattmillerai](https://github.com/mattmillerai) **Base:** `main` ← **Head:** `matt/be-18911-h3-speed-up` --- ### 📝 Commits (2) - [`727f2a8`](https://github.com/Comfy-Org/comfy-skills/commit/727f2a87e213339f611dd98e544ec6ea5c163493) docs(generate-video): rank the H3 speed-up options in the MiniMax H3 notes - [`1ca6b6b`](https://github.com/Comfy-Org/comfy-skills/commit/1ca6b6be565784ab7d9d27549a552d2b57415125) docs(generate-video): tie the H3 speed-up values to the fetched template graph ### 📊 Changes **2 files changed** (+2 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `claude-code/commands/generate-video.md` (+1 -0) 📝 `skills/comfy-generate-video.md` (+1 -0) </details> ### 📄 Description ## ELI-5 When someone asks an agent to make their MiniMax H3 video generation faster, the skill gave it no default, so it just repeated whatever `search_templates` returned. This adds one bullet to the H3 model notes. It makes the default a single switch in the shipped template (the Lightning LoRA), ranks the next options in order, and says the paid Max Turbo templates are a route change to offer, not something to pick quietly. ## What changed - One bullet appended to `## Model notes: MiniMax H3 (as of 2026-08)`, with the values marked as 2026-08 values to confirm against the fetched template graph, in both `skills/comfy-generate-video.md` and `claude-code/commands/generate-video.md`. The added line is byte-identical in both files. The heading and its "as of" date are left alone. - Nothing else is touched. Step 0 and the area #32 edits are untouched, so this should rebase cleanly whichever merges first. ## Judgment calls - **One clause reworded to pass the MCP negative-claim lint.** The suggested wording said the Max Turbo templates go "via `run_template`, not a `partner_generate` slug". `comfy-cloud-mcp-server`'s `guide-negative-claims` lint flagged that sentence as a non-support claim with no inventory row (2 failures). I changed it to "launch them with `run_template`", which says the same thing affirmatively. The skill copy's existing Hailuo bullet already says `minimax/hailuo-03-t2v` is the only registered MiniMax slug. - **The wording is neutral about what's "above" it.** The `claude-code/` copy's H3 notes are an older version that never mentions `run_template` for the other paid H3 templates. So the bullet doesn't point back to that text. Its one back-reference, "the OSS-vs-partner rule above", is the Step 0 both-routes rule, which exists in both files. ## Verification - Cited names checked against `comfy-cloud-mcp-server` `origin/main` `data/templates.json` and the bundled catalog. All found: `video_minimax_h3_t2v` / `_i2v` / `_r2v`, `video_fastvideo_fasth3_t2v` / `_i2v` (and no fasth3 R2V entry, which matches "no R2V variant"), `api_minimax_h3_max_turbo_t2v` / `_i2v`, and `MiniMaxH3SigmaShift`. The Max Turbo T2V catalog description itself says "768p clips in roughly 1–3 seconds". - MCP guards: in a temporary worktree of `comfy-cloud-mcp-server` at `origin/main`, I replaced the bundled `data/skills.json` `comfy-generate-video` body with the new file (the bundled body matched current `main` byte-for-byte, so this is the same as a refresh) and ran the guard tests. ## Residual - **Template internals checked against the live graphs (review round):** in `Comfy-Org/workflow_templates` `main`, `Boolean (Enable Lightning LoRA)` is off by default in `video_minimax_h3_t2v` / `_i2v` / `_r2v`, and the step switch picks 20 or 6 (R2V: 20 or 4). FastH3 t2v/i2v run 8 steps with `MiniMaxH3SigmaShift` 10/3, and there is no FastH3 R2V. The H3 graphs have no shift node, so the earlier "12/3" default-shift figure was dropped. The bullet now tells the agent to read these values from the fetched graph before acting on them. The 768p 4-step LoRA is listed in the template notes, but its 6/3 shift was not checked against a graph. - **The two copies of the H3 section already differ** (before this PR) in the routing/prefix bullet, the evidence-precedence bullet and the Hailuo bullet. The `claude-code/` copy is the older wording. So the H3 sections as a whole don't match, though the added bullet does in both files. Syncing the `claude-code/` copy is a separate change and wasn't done here. - **Not run:** `npm run refresh:skills`. It fetches from GitHub, so I injected the body locally instead. The real nightly refresh will run the same guards after merge. ## Provenance - **Authored by:** agent-work loop - **Verified:** - The comfy-cloud-mcp-server guide guards passed with the new bullet injected into `data/skills.json` (273 passed, same as baseline). They ran on the original wording; the review-round edit adds no negative claims and was not re-run through them. - The template names, the Lightning switch default, the 20→6 / R2V 20→4 steps and the FastH3 8-step 10/3 shift were checked against the live `workflow_templates` graphs. The Max Turbo 1–3 s figure is in the official catalog description. - The added bullet is byte-identical in both skill files. - **Deviations:** - One clause was reworded from the suggested text to pass the negative-claim lint. - The unverifiable "12/3" default-shift figure was dropped after review. - The H3 sections as a whole already differed between the two files before this PR. Only the added bullet matches. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:34:13 +00:00