Zhenzhen's AI Radar · MiniMax H3 · 2026-11-01 Evening Report
This issue features 5 new versions: Strongly Recommended 0 · Recommended 1 · Worth Watching 2 · Listed Only 2
Editor's Note This issue reviews 5 new H3 versions. The top recommendation is LongForge 7.4: a ComfyUI workflow that stitches H3 long videos into finished pieces with audio, scene by scene. In this version, the prompt assistant allows you to choose the large language model backend for writing prompts (local, Ollama, or OpenAI-compatible interfaces). Saved scenes can also be individually upscaled and refined. Maintains an A rating. The weapon attack speed slider parameters and examples are complete, but it is paid; maintains a B rating. The Proteus one-click studio only released a link file, with the main body on Patreon; maintains a B rating. Two character LoRAs lack H3-specific instructions or parameters, downgraded from B to C.
Recommended (1)
[A] MiniMax H3 LongForge FL2VA & REF2VA (LONG VIDEO) · LongForge v7.4
Workflow · Version Update · Author AIBOB1199 · Published 11-01 09:30 · Downloads 251
A ComfyUI workflow suite that generates MiniMax H3 long videos segment by segment per scene and then stitches them into finished pieces with audio (including a bundled node package, FL2VA / REF2VA main workflows, and RefMod Studio); the 7.4 version primarily upgrades the prompt assistant.
Maintains A: The scene-by-scene long video workflow addresses a real need for H3, and the 7.4 prompt assistant changes are specifically documented; points deducted for the lack of visible finished video examples and heavy dependencies.
Images
Reasons for Recommendation
- Version Update: The update notes section is empty, but the model card details the 7.4 changes—the prompt assistant (16 · Prompt Assistant) backend can be selected from NATIVE (text models compatible with ComfyUI TextGenerate, such as Gemma), LOCAL, GGUF / TRANSFORMERS, OLLAMA, and OpenAI-compatible SERVER interfaces; writing rules can use built-in, custom, or both combined; you can separately toggle whether to pass the previous segment's prompt, reference image, and last frame to the model
- Serves the need for multi-scene long films using first/last frame to video (FL2VA) and reference image to video (REF2VA): from a single idea, it plans scenes, writes prompts, selects shots, edits, and exports; the author states that saved scenes can undergo only UPSCALE or UPSCALE + REFINE without re-running the base generation
- Free, license allows image output and site rental; this version has approximately 12 hours 251 downloads since release, with the entire model accumulating 9266 downloads (representing only attention level)
How to Use
- Import minimaxH3LongforgeFL2VA_longforgeV74.zip (518 KB); the node package, main workflows, and RefMod Studio inside must all be installed; controls are in English, and both workflows include built-in Russian-English bilingual instructions
- The prompt assistant requires a separate text-generation model (H3's own text encoder cannot be used): place local models in ComfyUI/models/llm/, or connect to Ollama / OpenAI-compatible interfaces, with interface keys in environment variables; the author's scene transition parameters are CROSSFADE / BEST MATCH, aligned to the H3 grid at 24 FPS
- Basic generation parameters such as steps, CFG, and sampler follow the author's previous version settings and are not separately specified on the card
Risks and Limitations
- The author did not write update notes; 7.4 changes can only be found in the lengthy model card
- Only 1 visible example images (without prompts); 6 sensitive-level examples are hidden, so the finished video effect cannot be seen
- The suite is large with many dependencies: a custom node package, two workflows, plus an external large model, resulting in a high learning curve
- The beginning of the model card contains cryptocurrency donation addresses; this entry only read the model card and was not run
Worth Watching (2)
[B] Proteus: one-click MiniMax H3 + Krea 2 video & picture studio for ComfyUI · 1.1
Workflow · New Model · Author MythicalCreator · Published 11-01 19:56 · Downloads 23
A browser application called "Proteus" that runs locally on NVIDIA GPUs via ComfyUI: it turns H3 text-to-video/first-last frame/reference image-to-video, actor swapping, video upscaling/frame interpolation, and Krea 2 image editing into form-based buttons; the Civitai page contains only a link file, with the main body available to Patreon supporters.
Maintains B: The feature list is comprehensive, but the core is behind a paywall, and there are no visible examples.
Images
Reasons for Recommendation
- New Model: Serves users who do not want to connect nodes themselves and prefer generating H3 videos directly via forms; the author claims long shots can be generated continuously in multiple segments, with a Fast preset (4 steps), and displays the execution plan before output
- In version 1.1, the author claims new features including segmented long shots, continuation, character sheets, storyboards, and low-res video restoration; frame interpolation can optionally use RIFE or RTX Frame Generation up to 48 or 60 fps
- Free page, but the downloaded file is only proteusOneClickMinimaxH3Krea_11.txt (a link), with no visible examples
How to Use
- Obtain the installer from Patreon; the installer will find or install the portable version of ComfyUI and install the required node packages; models are not included in the package, and the Setup button in the app will indicate where each file should be placed
- Requires Windows 10/11 and an NVIDIA RTX GPU; parameters such as steps and CFG are not provided by the author except for the Fast preset
Risks and Limitations
- The main body is on Patreon, requiring supporters to access it; the Civitai file is just a link
- No visible examples (1 sensitive-level items are hidden); all features are as claimed by the author
- The license does not allow merging; the "actor swap" feature involves other people's video materials, and the author requires using only clips you have the right to use; this entry was not run
- The card image is an AI-generated schematic, not the actual model effect; the author has no publicly available general-level examples
[B] Motion-Speed-MMH3-SLIDER · KinetiX Weapon Speed
LoRA · Version Update · Author make676767 · Published 11-01 10:32 · Downloads 7 · Paid
An H3 motion slider LoRA that only accelerates weapon and projectile actions (drawing swords, stabbing with spears, rapid fire, cycling bolts); the author claims character bodies, faces, and environments remain static; negative weights result in heavy slow motion.
Maintains B: Usage instructions are complete and actionable, but it is paid, this version was just released, and there are only two examples.
Images
Reasons for Recommendation
- Version Update: This is the 4th slider on the same model page; previous versions adjusted limb action speed and camera movement speed, while this version isolates "weapon attack speed"
- Complete parameters: No trigger word; strength +1.3 ~ +1.6 is the common range, +0.8 ~ +1.2 is realistic martial arts rhythm, +1.8 ~ +2.0 is the upper limit, -1.0 ~ -1.5 and -2.0 ~ -3.0 are varying degrees of slow motion; the author also provides copy-paste storyboard-style prompts, requiring the first sentence to clearly define the initial contact between hands and weapons
- Both 2 visible video examples include prompts; this version is paid, with the model accumulating 4750 downloads (representing only attention level)
How to Use
- Load KinetiX_act_weapon_speed_h3_v1.safetensors (approx. 3 MB · BF16), no trigger word, start adjusting strength from +1.3 ~ +1.6; you can add the author's enhancement phrases, such as blistering weapon swing velocity
- The first sentence of the prompt should clearly describe the initial state of the weapon (e.g., thumb pressing on the sword guard), use continuous action descriptions throughout, and avoid static words like motionless or frozen; when stacking with the limb action speed slider, prioritize the weapon slider
Risks and Limitations
- Paid download, with 7 downloads for this version and very little feedback
- The automatically captured ±4.0 / ±8.0 strengths belong to other sliders on the same page, not this version
- "Characters and environment remain completely unchanged" is the author's claim, with only 2 videos; 4 sensitive-level examples are hidden; this entry was not run
Listed Only (2)
- Sister Vita Nox (Minimax-H3) · v1 — LoRA: An H3 character LoRA for the fictional character "Tattooed Battle Sister Sister Vita Nox": black nun habit, gothic tattoos covering face, neck, and arms, silver cross necklace.
- Claudia WhiteMoon · MiniMaxH3 — LoRA: The H3 version of the character LoRA for the fictional steampunk sports model character "Claudia WhiteMoon": corseted light armor, brass gauntlets, pocket watch pinned in hair.
Data: Civitai Public API (base model MiniMax H3), inclusion window 11-01 08:31 → 11-01 21:01 (Beijing Time). Recommendation reasons have been reviewed by editors; download counts represent only attention level.