MiniMax H3 Evening Report 2026-11-10 | First Recommendation: H3 Continuity Studio - ComfyUI Timeline, Story and Visual Memory for MiniMax H3

This issue reviews 11 new versions of H3, with 1 S-tier, 1 A-tier, 4 B-tier, and 5 C-tier entries. The top recommendation is H3 Continuity Studio: a ComfyUI timeline node package that strings sequential clips into a timeline, allowing it to "remember" selected frames from previous clips and carry them into subsequent shots; it also supports integration with Codex / Claude Code. The A-tier entry is a realistic video LoRA featuring views of Earth from low Earth orbit, including 2 visible videos and example prompts. Two CROWN hand/casual smartphone snapshot LoRAs have complete parameters but lack visible examples, and their training data sources are questionable; the frost flow speed slider is paid; two character voice LoRAs involve game character voices, placing them in B/C tier; a combat action LoRA with only external links, an int8 integrated base model with only two sentences of description, and two original

Zhenzhen's AI Radar · MiniMax H3 · 2026-11-10 Evening Report

This issue features 10 new versions: Strongly Recommended 1 · Recommended 1 · Worth Watching 4 · Listed Only 4

Editor's Note This issue reviews 11 new versions of H3, with 1 S-tier, 1 A-tier, 4 B-tier, and 5 C-tier entries. The top recommendation is H3 Continuity Studio: a ComfyUI timeline node package that strings sequential clips into a timeline, allowing it to "remember" selected frames from previous clips and carry them into subsequent shots; it also supports integration with Codex / Claude Code. The A-tier entry is a realistic video LoRA featuring views of Earth from low Earth orbit, including 2 visible videos and example prompts. Two CROWN hand/casual smartphone snapshot LoRAs have complete parameters but lack visible examples, and their training data sources are questionable; the frost flow speed slider is paid; two character voice LoRAs involve game character voices, placing them in B/C tier; a combat action LoRA with only external links, an int8 integrated base model with only two sentences of description, and two original character LoRAs are listed as C-tier.

Strongly Recommended (1)

[S] H3 Continuity Studio - ComfyUI Timeline, Story and Visual Memory for MiniMax H3 · v0.20.0

Workflow · New Model · Author KadoBOT · Published 11-10 19:25 · Downloads 30

A ComfyUI node package for H3: uses a timeline to arrange clips and carry context into sequential clips; "Visual Memory" can bring selected frames from previous clips into subsequent shots; also includes story writing, reference images, and a timeline editing interface.

It is suitable to first test continuity with short clips, confirming the transitions before extending the timeline.

Images

Example 1920x1080

Reasons for Recommendation

  • New model (new Civitai page, version v0.20.0): serves text-to-video needs for multi-shot short films and continuous narratives where scenes and characters must remain consistent across shots.
  • The author states that Visual Memory can retrieve previously selected frames even after inserting an outdoor shot or changing the camera angle; H3 Story can plan and write to the timeline via any OpenAI-compatible interface (local or cloud).
  • This version adds MCP tools and H3 skills, allowing Codex / Claude Code to drive Studio through a separate stdio adapter.
  • Installation steps are fully documented, source code is MIT licensed, and GitHub includes a user guide and demo videos; 3 visible example images.

How to Use

  • Extract to ComfyUI/custom_nodes/ (or clone the GitHub repository), use ComfyUI's built-in Python to run pip install -r requirements.txt, restart and load examples/h3_story_studio.json, then select your H3 model, text encoder, and VAE in the loader.
  • Add clips and write prompts in Timeline Studio; render the first segment first, then add a segment set to "Continuation" to render and check the transition; the Story feature requires configuring a separate language model interface.
  • Generation parameters such as sampling steps and resolution are not provided by the author and depend on the selected H3 model and workflow.

Risks and Limitations

  • Does not include model weights; requires a new version of ComfyUI capable of running H3, and Context Windows nodes require a ComfyUI version with the H3 Context Windows core node.
  • If Story connects to an external interface, it will send story text and (if enabled) reference images to that service; pay attention to privacy and costs.
  • "Consistency maintained after changing camera angles" is claimed by the author; visible examples are screenshots, with no generated videos attached on this page; newly listed, likes are 0.

Recommended (1)

[A] Earth from ISS | Low Earth Orbit | Planetary view | H3 Minimax LoRA · v1

LoRA · New Model · Author NOLIM_AI · Published 11-10 15:06 · Downloads 59

A realistic video LoRA featuring views of Earth from low Earth orbit: curved horizon, a thin layer of blue atmosphere, deep blue oceans, extensive cloud cover and desert coasts, with black space in the background; spacecraft structures may appear in the corner of the frame.

Suitable for opening establishing shots or transitions: Earth's curvature with spacecraft foreground.

Images

Video 864x480

Reasons for Recommendation

  • New model; serves Earth shots from the International Space Station perspective and space documentary-style planetary establishing shots. The author's goal is a realistic NASA footage feel rather than 3D rendered planets.
  • The author provides a complete example prompt (shot from a satellite or spacecraft, curved horizon, spacecraft robotic arm visible in the top right corner, black space background).
  • 2 visible example videos.

How to Use

  • Follow the author's example prompt style, using long sentences to describe the curvature of the horizon, distribution of oceans and clouds, and spacecraft structures in the frame; trigger words are not provided by the author.
  • Strength, steps, and CFG are not provided by the author.

Risks and Limitations

  • The description mainly covers creative intent and training costs, with no parameters provided; additionally, 2 examples are hidden.
  • Covers only the specific subject of viewing Earth from low Earth orbit, making the scope narrow.

Worth Watching (4)

[B] CROWN Perfect 5-Finger Hands, Eye Catchlights & Human Anatomy Booster | MiniMax H3 · v1.0 - True768P VideoOnly r64

LoRA · New Model · Author NineCrowns · Published 11-10 10:21 · Downloads 167

A human detail enhancement LoRA: the author claims it keeps each hand with five distinct fingers, clear knuckles and nails during fast gestures, and enhances eye highlights and natural blinking.

Images

Diagram 1086x1448

Reasons for Recommendation

  • New model; targets issues in H3 such as hand blurring, extra/merged fingers, and dull eyes during close-up actions like gesture dances, heart signs, waving, or holding faces.
  • The author provides a complete test configuration: 768×1344 (or 1344×768), 25 steps, euler + beta, CFG 1.0, MiniMaxH3SigmaShift (shift_video 12.0, shift_audio 3.0), recommended strength 0.70–0.80.
  • The author claims it works with Ref2VA / I2VA / T2VA, and can be used alone or stacked with character/style LoRAs.

How to Use

  • Add the trigger word crown_perfect_hands_eyes to the prompt, strength 0.70–0.80.
  • The author's tested base model is minimax_h3_ref2va_bf16 + qwen3vl_32b text encoder + fp32 video VAE; 25 steps, euler + beta, CFG 1.0.

Risks and Limitations

  • No visible examples (1 are hidden), so hand correction effects cannot be verified from the page; all claims are by the author.
  • The author states training materials are real gesture dance videos from Douyin / TikTok, but licensing is not specified.
  • Card images are AI-generated diagrams, not actual model outputs; the author has no publicly available standard-level examples.

Trigger Word: crown_perfect_hands_eyes

[B] CROWN InstaGirl & TikTok Raw iPhone Candid Deep-Focus Realism | MiniMax H3 · v1.0 - True768P VideoOnly r64

LoRA · New Model · Author NineCrowns · Published 11-10 10:20 · Downloads 131

A realistic smartphone snapshot LoRA: the author claims it removes H3's default studio lighting, oily skin, and large aperture background blur, replacing them with deep focus, matte skin, and natural window/daylight for a vertical fashion snapshot feel.

Images

Diagram 1254x1254

Reasons for Recommendation

  • New model; serves vertical OOTD outfit displays, street/indoor snapshots, and other character videos that need to look "shot with a raw phone camera."
  • The author claims background walls, furniture, and street scenes remain clear throughout, and handles perspective for 45° side profiles.
  • Parameters are identical to the same author's hand LoRA: 768×1344, 25 steps, euler + beta, CFG 1.0, recommended strength 0.70–0.80.

How to Use

  • Add the trigger word crown_instagirl_real to the prompt, strength 0.70–0.80; works with Ref2VA / I2VA / T2VA.
  • Set MiniMaxH3SigmaShift shift_video to 12.0 and shift_audio to 3.0.

Risks and Limitations

  • No visible examples (1 are hidden), all effects are claimed by the author.
  • The author states training used real person snapshot videos from Douyin / TikTok, but licensing is not specified; visuals lean towards influencer-style snapshots, so pay attention to portrait usage rights.
  • Card images are AI-generated diagrams, not actual model outputs; the author has no publicly available standard-level examples.

Trigger Word: crown_instagirl_real

[B] Naoto Shirogane Voice - Persona 4 - Audio Only [Minimax H3] · Minimax H3 v1.0

LoRA · New Model · Author OneMoreLurker · Published 11-10 11:13 · Downloads 49

Pure audio character voice LoRA: Trained on Japanese voice recordings of Naoto Shirogane from Persona 4 Golden, enabling H3 to generate dialogue in this character's voice.

Images

Video 544x800

Reasons for Recommendation

  • New model; serves the need to add Japanese dialogue with corresponding character voices to video content. The author claims the data covers various emotions such as shyness, nervousness, confidence, and happiness.
  • The author provides a training subtitle format: Naoto says "...", with dialogue in Japanese; English dialogue can also be used, but the author believes it will differ from the official English dub.
  • 3 visible example videos.

How to Use

  • Write prompts in the format Naoto says "<Japanese dialogue>"; trigger words, strength, and steps are not provided by the author.

Risks and Limitations

  • The author admits this version has not fully learned the character's somewhat neutral tone and plans to retrain later.
  • Trained on game character voiceovers; the voice belongs to the original voice actor, so confirm rights before public or commercial use.

[B] KinetiX Elemental VFX Suite · KinetiX Freeze Speed

LoRA · Version Update · Author make676767 · Published 11-10 18:21 · Downloads 3 · Paid

A frost spread speed slider LoRA: positive values cause the frost front to advance rapidly across ground, water, and buildings; negative values result in slow creeping ice. The author claims it does not change ice crystal size, nor affect character faces, clothing, or poses.

Images

Diagram 1086x1448

Reasons for Recommendation

  • Version update (same KinetiX element effects suite page; the previous version featured a flame spread slider): This version adds frost flow speed control
  • Intensity tiers are clearly defined: +1.0 to +2.0 for rapid spread, +2.0 to +5.0 for instant freezing, 0 is the base model default, -1.0 to -2.0 for slow crawling, and -2.0 to -5.0 for extremely slow freezing
  • Provides three ready-to-use prompts for ground ice waves, slow castle freezing, and combat freeze-frame effects, along with a workflow resource archive

How to Use

  • No trigger words required; set positive values in strength_model to accelerate and negative values to decelerate. Start with the author's three scenario prompts

Risks and Limitations

  • Paid download; visible examples are 0 (4 hidden)
  • The filename stated in the description does not match the actual filename
  • The card image is an AI-generated illustration, not the model's actual output; the author has no publicly available standard-tier examples

Included Only (4)


Data: Civitai public API (base model MiniMax H3), inclusion window 11-10 08:31 → 11-10 21:01 (Beijing Time). Recommendations have been reviewed by editors; download counts reflect attention only.

Civitai

Explore the latest AI Radar updates